betaFeedback

WorksMathematics

Weak and Sharp Stretched-Exponential Flocking in the Kinetic Cucker-Smale Model with Noncompactly Supported Initial Data

Abstract

We study the long-time behavior of the kinetic Cucker-Smale model under initial data with unbounded support in both position and velocity. Such data invalidate classical diameter-based flocking arguments and make sustained spatial cohesion difficult to establish. We prove an estimate for spatial characteristics under a finite spatial second moment and a sufficiently high velocity moment, without any requirement on coupling strength. We also prove weak flocking, i.e. control of spatial variance and integrability of velocity fluctuations in time. Under stronger assumptions on the initial tails, we establish uniform propagation of spatial exponential moments of explicitly determined order, a substantially stronger conclusion than bounded spatial variance. These estimates lead to explicit stretched-exponential rates of velocity alignment and convergence of the centered spatial distribution toward a limiting profile. Another contribution is the construction of countably atomic solutions that show optimality of the decay exponent in the spatial-tail-limited class. Together, the results clarify how initial phase-space tails govern spatial confinement and quantitative flocking for kinetic distributions without compact support.

Introduction

From classical flocking to noncompact kinetic distributions

Collective alignment is a basic mechanism in the mathematical description of self-organized motion. In the Cucker–Smale (CS) model, individuals adjust their velocities by averaging the velocities of the surrounding population, with an influence that decreases as the spatial separation increases. The original particle system of Cucker and Smale [7] prompted a large literature on unconditional and conditional flocking, mean-field limits, and the associated kinetic and hydrodynamic equations. The passage from particle dynamics to kinetic descriptions was developed, in particular, by Ha and Tadmor [10], while the Lyapunov approach of Ha and Liu [9] established a useful connection between alignment mechanisms and mean-field stability. Kinetic flocking and measure-valued well-posedness were further investigated by Carrillo, Fornasier, Rosado, and Toscani [3] and by Cañizo, Carrillo, and Rosado [2]. An exhaustive presentation of the literature on this subject is far beyond the scope of this work, so we give the following surveys [5, 8, 21], and also the work of the author on the subject [18, 19, 20].

In this article, we work with the kinetic CS equation

∂tft+v⋅∇xft+∇v⋅(F[ft]ft)=0,F[ft](t,x,v)=κ∫ψ(∣x−y∣)(w−v)ft(dy,dw),(1)\partial_t f_t+v\cdot\nabla_x f_t+\nabla_v\cdot\left(F[f_t]f_t\right)=0,\qquad F[f_t](t,x,v)=\kappa\int\psi\left(\lvert x-y\rvert\right)(w-v)f_t(\mathrm{d}y,\mathrm{d}w), \tag*{(1)}

with coupling strength κ>0\kappa> 0, and communication weight

ψ(r)=(1+r2)−β/2,0<β<1.(2)\psi(r) = (1+r^2)^{-\beta/2}, \qquad0 < \beta< 1. \tag*{(2)}

The unknown, ft(x,v)f_t(x,v) is a probability measure that describes the probability of particles located at the phase-space point (x,v)∈Rxd×Rvd(x,v) \in\mathbb{R}^d_x \times\mathbb{R}^d_v at time tt. The measure ftf_t is normalized to have unit total mass. Conservation of momentum fixes the mean velocity vcv_c, and the center of mass of the system moves as xc(t)=xc(0)+tvcx_c(t) = x_c(0) + tv_c. The two natural macroscopic fluctuations are

X(t)=∫∣x−xc(t)∣2 dft(x,v),V(t)=∫∣v−vc∣2 dft(x,v).(3)\mathcal{X}(t) = \int|x-x_c(t)|^2 \,df_t(x,v), \qquad\mathcal{V}(t) = \int|v-v_c|^2 \,df_t(x,v). \tag*{(3)}

For compactly supported initial data, the traditional flocking argument controls the diameter of the spatial support and thereby recovers a strictly positive lower bound on the communication weight along interacting pairs. This ultimately yields exponential decay of the velocity diameter and a uniformly bounded spatial diameter in the unconditional regime. This mechanism of control is not available for fully noncompact data: even when nearly all of the mass stays within a small region, particles of arbitrarily small mass can occur at arbitrarily large distances. In this setting, the infimum of the pairwise communication coefficients may be zero at every time.

This obstruction motivates a distinction between several notions of flocking. Velocity alignment in the mean-square sense means V(t)→0\mathcal{V}(t) \to0. The usual weak flocking conclusion supplements this by sup⁡t≥0X(t)<∞\sup_{t\geq0}\mathcal{X}(t) < \infty. The latter is an averaged form of spatial cohesion but does not, by itself, propagate a prescribed class of spatial tails. In particular, boundedness of X(t)\mathcal{X}(t) yields only a Chebyshev estimate of the type ft{∣x−xc(t)∣>R}≤CR−2f_t\{|x-x_c(t)|>R\} \leq CR^{-2}, for any R>0R>0; it does not imply a uniform exponential or stretched-exponential spatial bound. A central objective here is the stronger property

sup⁡t≥0∫R2dexp⁡(a∗∣x−xc(t)∣σ) dft(x,v)<∞for some a∗,σ>0.(4)\sup_{t\geq0} \int_{\mathbb{R}^{2d}} \exp\left(a_*|x-x_c(t)|^\sigma\right) \,df_t(x,v) < \infty\qquad\text{for some } a_*,\sigma>0. \tag*{(4)}

Besides quantifying spatial cohesion, such a bound provides an exponentially small error when the velocity dissipation is localized to an expanding ball. This connection is essential to the convergence rates proved below.

Previous results for noncompactly supported data

The study of the kinetic model with noncompact support is not new. From the point of view of well-posedness, Chen and Yin [4] proved global weak existence under finite first phase-space moments. They also showed stronger classical and measure well-posedness results for compactly supported data. Their results illustrate why one must separate the existence of a noncompact weak solution from the propagation of strong flocking estimates. Wang and Xue [22] studied formation and velocity consensus for kinetic CS systems with noncompact spatial distributions, advancing the asymptotic theory beyond the usual bounded spatial support setting.

A particularly important sequence of advances is due to S.-Y. Ha, X. Wang et al. In particular, Ha, Wang, and Xue [15] considered kinetic distributions with noncompact spatial support and bounded initial velocity support. Under sufficiently rapid spatial decay and an appropriate coupling hypothesis, they obtained weak flocking, namely decay of the centered velocity variance together with boundedness of the centered spatial second moment. Their analysis uses an infinite-particle approximations and tail estimates in place of the classical support-diameter functional.

Ha and Wang [12] subsequently treated the fully noncompact setting, allowing both position and velocity support to be unbounded for any d≥1d \ge1. They distinguished polynomially and exponentially decaying initial distributions and developed a time-varying effective-region method. Their polynomial-moment theorem gives algebraic decay of the velocity variance and uniform spatial second-moment control; the moment exponents are linked by explicit inequalities. For data with an exponential moment of the form ∫ea(∣x∣+∣v∣) dfin<∞\int e^{a(|x|+|v|)}\,\mathrm{d}f_{\mathrm{in}} < \infty, they established weak flocking with convergence faster than any prescribed algebraic rate. If the velocity-tail hypothesis is strengthened to ∫ea(∣x∣+∣v∣δ) dfin<∞\int e^{a(|x|+|v|^\delta)}\,\mathrm{d}f_{\mathrm{in}} < \infty, with δ>1\delta> 1 and β<1−1/δ\beta< 1 - 1/\delta, their Theorem 2.2 gives the explicit estimate

V(t)≤Cexp⁡[−ct1−βδδ−1].(5)\mathcal{V}(t) \le C\exp\left[-ct^{1-\frac{\beta\delta}{\delta-1}}\right]. \tag*{(5)}

The same paper proves weak spatial cohesion. Its stated theorems do not, however, assert a uniform-in-time spatial exponential moment of the form (4). This distinction matters: an exponential bound on the initial distribution and a uniform second spatial moment at later times do not automatically furnish such a uniform exponential-moment estimate.

The question of constructing the corresponding solutions is equally important. The noncompact weak-existence theory of Chen–Yin [4] gives one point of reference, while Ha and Wang [12] establish uniqueness under additional bounded-density and exponential velocity-tail assumptions. More recently, Ha and Wang [13] obtained finite-time Osgood-type stability and random mean-field approximation for a class with exponential velocity moments and finite spatial second moments. An essential technical point is that boundedness and smoothness of ψ\psi do not give a globally uniform phase-space Lipschitz constant for F[f]F[f] when ∣v∣|v| is unbounded: the spatial derivative of the alignment force involves the factor w−vw-v. We therefore prove existence of Lagrangian probability solutions in the moment class used below and state our estimates for every such solution, without asserting a general nonlinear uniqueness theorem.

Other recent contributions clarify adjacent, but distinct, mechanisms. Ha and Wang’s relativistic model [11] treats noncompact phase-space data with a modified velocity structure. Ha, Wang, and Yoon [17] investigate difficulties in uniform-in-time stability and propose a modified noncompact interaction model. Ha and Wang [14] establish quantitative flocking in the presence of a confining force; genuine exponential fluctuation-energy relaxation there is assisted by confinement and should not be confused with the free alignment dynamics in (1). Finally, Ha, Wang, and Xue [16] analyze clustering and spectral properties of a countably infinite particle system. That perspective is relevant to the construction of infinitely many distant “satellite” particles, but their infinite-graph spectral conclusions address a different question from the propagation of kinetic spatial exponential moments.

Motivation of this work and main results

The developments we mentioned leave room for a more detailed description of the spatial distribution itself. Rather than merely asking whether X(t)\mathcal{X}(t) remains finite, we ask whether the initial position and velocity tails control the location of each characteristic for all time, and whether that confinement yields sharp quantitative velocity decay. The answers presented in this paper rest on three related results.

Our first result is a finite-moment confinement principle, valid for every d≥1d \ge1. Suppose the centered initial spatial second moment is finite and the initial centered velocity has a moment of order

m>21−β.(6)m > \frac{2}{1-\beta}. \tag*{(6)}

We prove weak spatial cohesion, quantitative algebraic velocity alignment, and the integrability of the velocity standard deviation,

sup⁡t≥0X(t)<∞,∫0∞V(t) dt<∞.(7)\sup_{t\ge0} \mathcal{X}(t)<\infty,\qquad\int_{0}^{\infty}\sqrt{\mathcal{V}(t)}\,\mathrm{d}t<\infty. \tag*{(7)}

The key additional consequence is the support-independent characteristic estimate

sup⁡t≥0∣Xt(z)∣≤C(1+∣X0(z)∣+∣V0(z)∣)1/(1−β)for fin-a.e. z,(8)\sup_{t\ge0}|X_t(z)|\le C\left(1+|X_0(z)|+|V_0(z)|\right)^{1/(1-\beta)} \qquad\text{for }f_{\mathrm{in}}\text{-a.e. }z, \tag*{(8)}

where XtX_t, VtV_t are centered characteristic variables. The constant is controlled by averaged initial moments and fixed model parameters; it is independent of the support diameter, particle number, and the individual label zz. While related characteristic damping ideas are classical, the combination with the finite-moment Lyapunov estimate allows their use for fully noncompact measures in any dimension.

The second result is uniform propagation of strong spatial tails. If for some p,q,a,b>0p,q,a,b>0,

∫exp⁡(a∣x−xc(0)∣p+b∣v−vc∣q)fin(x,v)<∞,(9)\int\exp\left(a|x-x_c(0)|^p+b|v-v_c|^q\right)f_{\mathrm{in}}(x,v)<\infty, \tag*{(9)}

then (4) holds with the explicitly determined order

σ=min⁡{p,(1−β)q}.(10)\sigma=\min\{p,(1-\beta)q\}. \tag*{(10)}

Thus the spatial tail is limited by the initial position exponent pp and by the velocity exponent qq after multiplication by the coercivity exponent 1−β1-\beta. The statement includes fully noncompact phase-space distributions, and neither compact velocity support nor a large-coupling condition is assumed. The associated velocity-variance estimate is

V(t)≤Cexp⁡[−c(1+t)σ/(σ+β)],t≥0.(11)\mathcal{V}(t)\le C\exp\left[-c(1+t)^{\sigma/(\sigma+\beta)}\right],\qquad t\ge0. \tag*{(11)}

In addition, the centered spatial marginal converges to a limiting probability measure in W2W_2, with a quantitative rate. This conclusion identifies an asymptotic spatial profile even though the original spatial support need never become bounded.

Our third result concerns optimality. For every p>0p>0, we construct a symmetric countably atomic kinetic solution, supported on a fixed line in Rd\mathbb{R}^d, with bounded initial velocities and uniformly bounded spatial exponential moments of order pp, for which

V(t)≥cexp⁡[−C(1+t)p/(p+β)].(12)\mathcal{V}(t)\ge c\exp\left[-C(1+t)^{p/(p+\beta)}\right]. \tag*{(12)}

The interaction of an arbitrarily distant satellite with the central mass may be extremely weak, even when the satellite has exponentially small weight. Choosing a geometric sequence of radii and suitable satellite masses realizes this slow mode simultaneously with strong spatial-tail bounds. The lower bound shows that the exponent p/(p+β)p/(p+\beta) cannot be improved uniformly in the spatial-tail-limited measure class with bounded velocities. The construction does not claim sharpness for smooth strictly positive kinetic densities or for the regime in which the velocity-tail exponent determines σ\sigma.

For comparison with the exponential part of Ha–Wang [12], Theorem 2.2, take β=1/4\beta=1/4 and assume initial exponential position tails of order one and Gaussian-type velocity tails of order two. Their stated estimate yields exponent 1/21/2, whereas (10)–(11) yield σ=1\sigma=1 and exponent 4/54/5 in arbitrary dimension. Under merely exponential velocity tails of order one, the present result instead gives σ=3/4\sigma=3/4 and exponent 3/43/4; the corresponding statement in [12] asserts super-polynomial alignment but does not specify this exponent. These are comparisons between stated sufficient estimates, not claims that the earlier methods could not be sharpened. Our finite-moment theorem also improves the sufficient moment conditions for weak flocking. It requires only a finite initial spatial second moment and a finite velocity moment of order m>2/(1−β)m > 2/(1-\beta), for every 0<β<10 < \beta< 1. The polynomial-moment criterion in [12 Theorem 2.2] also admits a spatial moment of order two in part of the parameter range. More precisely, setting their spatial and velocity moment orders to D1=2D_1 = 2 and D2=mD_2 = m, respectively, and optimizing over their auxiliary parameters gives the sufficient conditions

0<β<12,m>2(1−β)1−2β.0 < \beta< \frac{1}{2}, \qquad m > \frac{2(1-\beta)}{1-2\beta}.

For 0<β<1/20 < \beta< 1/2, our velocity-moment threshold is strictly smaller; for 1/2≤β<11/2 \le\beta< 1, their stated polynomial-moment criterion cannot be satisfied with D1=2D_1 = 2, whereas our theorem still applies under the stated velocity-moment assumption. Thus the improvement concerns the admissible combinations of spatial and velocity moments, rather than the admission of finite-second-moment spatial data alone.

The results of the present article consequently contribute in three complementary directions: a finite-moment, dimension-independent criterion for spatial cohesion and characteristic confinement; uniform propagation of quantitative spatial exponential tails with a corresponding explicit velocity decay rate; and a countably atomic example matching that rate in the spatial-tail-limited class. The first result identifies a tractable threshold within our method, but we do not claim that the strict moment inequality (6) is necessary in every dimension. The atomic construction proves sharpness in a specified measure class, without settling smooth-density sharpness. These distinctions are important for separating proven results from possible extensions.

The rest of the paper is arranged as follows. Section 2 states the main results and notation. Section 3 establishes existence of global Lagrangian solutions under the required finite moments. Section 4 records the velocity dissipation identities and localization tools. Section 5 proves the coupled Lyapunov estimate and the characteristic confinement theorem. Section 6 establishes strong tail propagation, stretched-exponential alignment, and convergence of the centered spatial marginal. Section 7 develops the countably atomic sharpness construction. The concluding section identifies remaining mathematical questions.

Setting and main theorems

Let d≥1d \ge1, 0<β<10 < \beta< 1, and κ>0\kappa> 0. The kinetic CS model (for (x,v)∈Rd×Rd(x,v) \in\mathbb{R}^d \times\mathbb{R}^d) is

{∂tft+v⋅∇xft+∇v⋅(F[ft]ft)=0,F[ft](t,x,v)=κ∫ψ(∣x−y∣)(w−v)ft(dy,dw),ψ(r)=(1+r2)−β/2.(13)\left\{ \begin{aligned} \partial_t f_t + v \cdot\nabla_x f_t + \nabla_v \cdot\left(F[f_t]f_t\right) &= 0,\\ F[f_t](t,x,v) &= \kappa\int\psi(|x-y|)(w-v)f_t(dy,dw), \qquad\psi(r) = (1+r^2)^{-\beta/2}. \end{aligned} \right. \tag*{(13)}

Let fin(x,v)f_{\mathrm{in}}(x,v) be a Borel probability measure of total mass one, i.e. ∫dfin(x,v)=1\int\mathrm{d}f_{\mathrm{in}}(x,v) = 1. Conservation of mass implies that any solution remains a probability measure ft∈P(R2d)f_t \in\mathcal{P}(\mathbb{R}^{2d}). The equation admits solutions that can be represented as the push-forward of the initial measure finf_{\mathrm{in}} by the transport flow. A global Lagrangian solution is a family ft=(Xt,Vt)#finf_t = (X_t,V_t)_\# f_{\mathrm{in}} with almost-everywhere locally absolutely continuous trajectories satisfying the kinetic characteristic equations and locally continuous in L2(fin;R2d)L^2(f_{\mathrm{in}};\mathbb{R}^{2d}). The push-forward identity gives, for every bounded Borel function ϕ:R2d→R\phi:\mathbb{R}^{2d} \to\mathbb{R}, or more generally every ϕ∈L1(ft)\phi\in L^1(f_t) that

∫ϕ(x,v) dft(x,v)=∫ϕ(Xt(x,v),Vt(x,v)) dfin(x,v).\int\phi(x,v)\,\mathrm{d}f_t(x,v) = \int\phi\left(X_t(x,v),V_t(x,v)\right)\,\mathrm{d}f_{\mathrm{in}}(x,v).

Theorem 3.1 below establishes the existence of at least one such solution for every initial probability measure with the stated finite moments; all our a priori estimates hold for every Lagrangian solution in this class. We do not impose or assert general nonlinear uniqueness. Lemma 4.1 supplies the finite-time identities.

Lemma 2.1 (Spatial and velocity averages). We define the following averages for positions and velocities

xc(t):=∫x dft,vc(t):=∫v dft.x_c(t) := \int x\,\mathrm{d}f_t,\qquad v_c(t) := \int v\,\mathrm{d}f_t.

Then we have that vc(t)=vc(0)=:vcv_c(t)=v_c(0)=:v_c, and xc(t)=xc(0)+vctx_c(t)=x_c(0)+v_ct.

Proof. The proof follows by simply differentiating xc(t)x_c(t) and vc(t)v_c(t) along solutions of (13). □\square

Before we go to the main theorems we give some definitions for later use, starting from the centered characteristics and position-velocity fluctuations

Definition 2.2 (Centered characteristics and variances). For any point z=(x,v)∈Rd×Rdz=(x,v)\in\mathbb{R}^d\times\mathbb{R}^d we define the centered characteristics from point zz by

Xt(z):=Xt(z)−xc(t),Vt(z):=Vt(z)−vc.(14)\mathbf{X}_t(z):=\mathbf{X}_t(z)-x_c(t),\qquad\mathbf{V}_t(z):=\mathbf{V}_t(z)-v_c. \tag*{(14)}

We also define the variances around mean position and velocity by

X(t):=∫∣Xt∣2 dfin,V(t):=∫∣Vt∣2 dfin,Z(t):=1+X(t).(15)\mathcal{X}(t):=\int|\mathbf{X}_t|^2\,\mathrm{d}f_{\mathrm{in}},\qquad\mathcal{V}(t):=\int|\mathbf{V}_t|^2\,\mathrm{d}f_{\mathrm{in}},\qquad\mathcal{Z}(t):=\sqrt{1+\mathcal{X}(t)}. \tag*{(15)}

The quantity Z(t)\mathcal{Z}(t) is a renormalized spatial variance that comes handy later.

It is not difficult to verify that

∫Xt dfin=∫Vt dfin=0\int\mathbf{X}_t\,\mathrm{d}f_{\mathrm{in}}=\int\mathbf{V}_t\,\mathrm{d}f_{\mathrm{in}}=0

We also have the characteristic system around the average

X˙t=Vt,V˙t(z)=−κ∫ψ(∣Xt(z)−Xt(z′)∣)(Vt(z)−Vt(z′)) dfin(z′).(16)\dot{\mathbf{X}}_t=\mathbf{V}_t,\qquad\dot{\mathbf{V}}_t(z)=-\kappa\int\psi\bigl(|\mathbf{X}_t(z)-\mathbf{X}_t(z')|\bigr)\bigl(\mathbf{V}_t(z)-\mathbf{V}_t(z')\bigr)\,\mathrm{d}f_{\mathrm{in}}(z'). \tag*{(16)}

We also give the standard definition of the W2W_2 Wasserstein distance which is a metric distance for distributions. We denote by P2(Rd)\mathcal{P}_2(\mathbb{R}^d) the space of Borel probability measures with finite second moments. For ρ,ν∈P2(Rd)\rho,\nu\in\mathcal{P}_2(\mathbb{R}^d), the quadratic Wasserstein distance is defined by

W2(ρ,ν)=(inf⁡π∈Π(ρ,ν)∫Rd×Rd∣x−y∣2 dπ(x,y))1/2,W_2(\rho,\nu)=\left(\inf_{\pi\in\Pi(\rho,\nu)}\int_{\mathbb{R}^d\times\mathbb{R}^d}|x-y|^2\,\mathrm{d}\pi(x,y)\right)^{1/2},

where Π(ρ,ν)\Pi(\rho,\nu) is the set of all couplings of ρ\rho and ν\nu.

In our first theorem we prove weak-flocking under the assumptions of quadratic spatial initial moments, and velocity moments of order m>2/(1−β)m>2/(1-\beta). We also show an important global (in time) estimate of the spatial characteristics.

Theorem 2.3 (Weak flocking and control of spatial characteristic). Assume X(0)<∞\mathcal{X}(0)<\infty and for some m>2/(1−β)m>2/(1-\beta), we assume finite m’thm\text{’th} velocity moment of the initial data, i.e.

Mm:=∫∣v−vc∣m dfin<∞.(17)M_m:=\int|v-v_c|^m\,\mathrm{d}f_{\mathrm{in}}<\infty. \tag*{(17)}

Then the following estimates hold:

sup⁡t≥0X(t)<∞,(18)\sup_{t\ge0}\mathcal{X}(t)<\infty, \tag*{(18)}
∫0∞V(t) dt<∞,(19)\int_{0}^{\infty}\sqrt{\mathcal{V}(t)}\,\mathrm{d}t<\infty, \tag*{(19)}
sup⁡t≥0∣Xt(z)∣≤C(1+∣X0(z)∣+∣V0(z)∣11−β)for fin-a.e. z.(20)\sup_{t\ge0}|X_t(z)|\le C\left(1+|X_0(z)|+|V_0(z)|^{\frac{1}{1-\beta}}\right)\quad\text{for }f_{\mathrm{in}}\text{-a.e. }z. \tag*{(20)}

Moreover, if V(0)>0\mathcal{V}(0)>0, then for η=βm2(m−2)∈(0,1/2)\eta=\frac{\beta m}{2(m-2)}\in(0,1/2),

V(t)≤C(1+t)−1/η.(21)\mathcal{V}(t)\le C(1+t)^{-1/\eta}. \tag*{(21)}

The constant in (20) depends only on d,β,κ,X(0),V(0),m,Mmd,\beta,\kappa,\mathcal{X}(0),\mathcal{V}(0),m,M_m. If V(0)=0\mathcal{V}(0)=0, all centered velocities vanish and the conclusions are immediate.

A by-product of Theorem 2.3 is that we can actually get L2L^2 and W2W_2 control of the centered spatial tails

Corollary 2.4 (Polynomial spatial convergence under finite moments). Under the assumptions of Theorem 2.3, there exists X∞∈L2(fin;Rd)X_\infty\in L^2(f_{\mathrm{in}};\mathbb{R}^d) such that

∥Xt−X∞∥L2(fin)+W2((Xt)#fin,(X∞)#fin)≤C(1+t)−(12η−1),η=βm2(m−2)<12.\|X_t-X_\infty\|_{L^2(f_{\mathrm{in}})}+W_2((X_t)_{\#}f_{\mathrm{in}},(X_\infty)_{\#}f_{\mathrm{in}})\le C(1+t)^{-\left(\frac{1}{2\eta}-1\right)},\qquad\eta=\frac{\beta m}{2(m-2)}<\frac{1}{2}.

For V(0)=0\mathcal{V}(0)=0, the centered positions are constant.

Theorem 2.5 (Exponential spatial tails and stretched-exponential flocking). Assume, for p,q>0p,q>0 and a,b>0a,b>0,

Ma,b,p,q:=∫exp⁡(a∣x−xc(0)∣p+b∣v−vc∣q) dfin<∞.(22)M_{a,b,p,q}:=\int\exp\left(a|x-x_c(0)|^p+b|v-v_c|^q\right)\,\mathrm{d}f_{\mathrm{in}}<\infty. \tag*{(22)}

Set

σ=min⁡{p,(1−β)q},ϑ=σσ+β.(23)\sigma=\min\{p,(1-\beta)q\},\qquad\vartheta=\frac{\sigma}{\sigma+\beta}. \tag*{(23)}

There exist a∗,c,C>0a_*,c,C>0 depending on the initial moments and parameters such that

sup⁡t≥0∫ea∗∣x−xc(t)∣σ dft≤C,(24)\sup_{t\ge0}\int e^{a_*|x-x_c(t)|^\sigma}\,\mathrm{d}f_t\le C, \tag*{(24)}
V(t)≤Ce−c(1+t)ϑ.(25)\mathcal{V}(t)\le Ce^{-c(1+t)^\vartheta}. \tag*{(25)}

All assertions hold for arbitrary d≥1d\ge1 and κ>0\kappa>0.

Corollary 2.6 (Spatial tails, strong convergence and W2W_2). Under Theorem 2.5, for every finite s≥0s\ge0,

sup⁡t≥0∫∣Xt∣s dfin<∞,lim⁡R→∞sup⁡t≥0∫{∣Xt∣>R}∣Xt∣s dfin=0.(26)\sup_{t\ge0}\int|X_t|^s\,\mathrm{d}f_{\mathrm{in}}<\infty,\qquad\lim_{R\to\infty}\sup_{t\ge0}\int_{\{|X_t|>R\}}|X_t|^s\,\mathrm{d}f_{\mathrm{in}}=0. \tag*{(26)}

There is X∞∈L2(fin;Rd)X_\infty\in L^2(f_{\mathrm{in}};\mathbb{R}^d) satisfying

∥Xt−X∞∥L2(fin)+W2((Xt)#fin,(X∞)#fin)≤Ce−c(1+t)ϑ,(27)\|X_t-X_\infty\|_{L^2(f_{\mathrm{in}})}+W_2((X_t)_{\#}f_{\mathrm{in}},(X_\infty)_{\#}f_{\mathrm{in}})\le Ce^{-c(1+t)^\vartheta}, \tag*{(27)}

and the limiting spatial measure has an exponential moment of order σ\sigma (for the same a∗a_* after weakening it if necessary).

Remark 2.7. We mention the fact that in both corollaries above it suffice to prove the inequalities for ∥Xt−X∞∥L2(fin)\|X_t-X_\infty\|_{L^2(f_{\mathrm{in}})} due to the inclusion

W2((Xt)#fin,(X∞)#fin)2≤∫∣Xt(z)−X∞(z)∣2 dfin(z).W_2\left((X_t)_{\#}f_{\mathrm{in}},(X_\infty)_{\#}f_{\mathrm{in}}\right)^2 \le\int|X_t(z)-X_\infty(z)|^2\,df_{\mathrm{in}}(z).

The inequality follows from the definition of W2W_2, and the fact that πt=(Xt,X∞)#fin\pi_t=(X_t,X_\infty)_{\#}f_{\mathrm{in}} is only one of the possible couplings.

Proposition 2.8 (Conditional quantitative flocking for d≥1d\ge1). Suppose d≥1d\ge1 and β>0\beta>0. If a global centered Lagrangian solution satisfies

sup⁡t≥0∫ea∣Xt∣σ dfin≤Mx<∞,∫∣V0∣2+ε dfin<∞(28)\sup_{t\ge0}\int e^{a|X_t|^\sigma}\,df_{\mathrm{in}}\le M_x<\infty,\qquad\int|V_0|^{2+\varepsilon}\,df_{\mathrm{in}}<\infty \tag*{(28)}

for some a,σ,ε>0a,\sigma,\varepsilon>0, then

V(t)≤Cexp⁡[−c(1+t)σ/(σ+β)].(29)\mathcal V(t)\le C\exp\left[-c(1+t)^{\sigma/(\sigma+\beta)}\right]. \tag*{(29)}

This proposition does not use the restriction 0<β<10<\beta<1; that restriction enters through Theorem 2.3.

Existence theory and Lagrangian compatibility

Theorem 3.1 (Existence for the noncompact initial data). Let d≥1d\ge1, κ>0\kappa>0, and let ψ\psi be bounded, nonnegative, globally Lipschitz and s.t. ∥ψ∥L∞≤1\|\psi\|_{L^\infty}\le1. For every Borel probability measure fin∈P(R2d)f_{\mathrm{in}}\in\mathcal P(\mathbb R^{2d}) with

∫(∣x∣2+∣v∣m) dfin<∞,m≥2,\int(|x|^2+|v|^m)\,df_{\mathrm{in}}<\infty,\qquad m\ge2,

there exists a global weak kinetic solution f∈C([0,∞);P1(R2d))f\in C([0,\infty);\mathcal P_1(\mathbb R^{2d})) which has a deterministic Lagrangian representation ft=(Xt,Vt)#finf_t=(X_t,V_t)_{\#}f_{\mathrm{in}}. The characteristic map is locally continuous in L2(fin)L^2(f_{\mathrm{in}}), and

sup⁡t≤T∫(∣x∣2+∣v∣m) dft<∞(T<∞).\sup_{t\le T}\int(|x|^2+|v|^m)\,df_t<\infty\qquad(T<\infty).

The centered velocity mm-moment is nonincreasing and the exact velocity-variance identity holds. These conclusions hold for every such Lagrangian solution, not only for the one constructed below.

Proof. Step 1: finite atomic approximation. Approximate finf_{\mathrm{in}} by finite atomic probabilities finNf_{\mathrm{in}}^N obtained by truncation and quantization so that finN→finf_{\mathrm{in}}^N\to f_{\mathrm{in}} in W2W_2, their initial spatial second moments converge, and their velocity mm-moments converge (truncation can be chosen separately in the two variables). Every finite particle system is global. Its velocities remain in their initial convex hull, and symmetric pairing of the interaction terms gives conservation of mass and momentum, exact second-moment dissipation, and nonincrease of each convex centered velocity moment. In particular, with constants independent of NN,

sup⁡N,t≤T∫∣v∣m dftN<∞,sup⁡N,t≤T∫∣x∣2 dftN<∞(T<∞).(30)\sup_{N,t\le T}\int|v|^m\,df_t^N<\infty,\qquad\sup_{N,t\le T}\int|x|^2\,df_t^N<\infty\qquad(T<\infty). \tag*{(30)}

The second estimate follows from

∫∣x∣2 dftN≤∫∣x∣2 dfinN+t∫∣v∣2 dfinN.\sqrt{\int|x|^2\,df_t^N}\le\sqrt{\int|x|^2\,df_{\mathrm{in}}^N}+t\sqrt{\int|v|^2\,df_{\mathrm{in}}^N}.

Step 2: a global weak limit. The phase-space velocities of the particle curves are (v,F[fN])(v,F[f^N]), and ∣F[fN](x,v)∣≤κ(∣v∣+∫∣w∣ dfN)|F[f^N](x,v)|\le\kappa(|v|+\int|w|\,df^N). Thus t↦ftNt\mapsto f_t^N is equi-Lipschitz in W1W_1 on each finite time interval. Estimate (30) gives tightness and uniform integrability of first moments; metric Arzelà–Ascoli and a diagonal argument yield a subsequence converging in C([0,T];P1)C([0,T];\mathcal{P}_1) for every TT. In the weak kinetic formulation, the nonlinear term is

κ∬∇vφ(x,v)⋅ψ(∣x−y∣)(w−v) dftN(x,v) dftN(y,w).\kappa\iint\nabla_v\varphi(x,v)\cdot\psi(|x-y|)(w-v)\,\mathrm{d}f_t^N(x,v)\,\mathrm{d}f_t^N(y,w).

The bounded continuity of the kernel and the uniform mm-moment bound (m≥2m\ge2) give convergence after a velocity truncation; the remaining tails are uniformly small because the integrand grows at most linearly in ∣v∣+∣w∣|v|+|w|. Hence the limiting curve solves the kinetic equation globally, has the initial measure finf_{\mathrm{in}}, and satisfies the finite-time moment bounds by lower semicontinuity.

Step 3: deterministic characteristics. For this fixed solution, write F[f](t,x,v)=−κat(x)v+κbt(x)F[f](t,x,v)=-\kappa a_t(x)v+\kappa b_t(x) with

at(x)=∫ψ(∣x−y∣) dft(y,w),bt(x)=∫ψ(∣x−y∣)w dft(y,w).a_t(x)=\int\psi(|x-y|)\,\mathrm{d}f_t(y,w),\qquad b_t(x)=\int\psi(|x-y|)w\,\mathrm{d}f_t(y,w).

If L=Lip⁡ψL=\operatorname{Lip}\psi and M1(t)=∫∣w∣ dftM_1(t)=\int|w|\,\mathrm{d}f_t, then

∣at(x)−at(y)∣≤L∣x−y∣,∣bt(x)−bt(y)∣≤LM1(t)∣x−y∣.|a_t(x)-a_t(y)|\le L|x-y|,\qquad|b_t(x)-b_t(y)|\le LM_1(t)|x-y|.

Consequently, on a region ∣v∣≤K|v|\le K, the full field (v,F[f](t,x,v))(v,F[f](t,x,v)) is locally Lipschitz in (x,v)(x,v), with a constant depending on KK and the finite-time first velocity moment. It is not globally Lipschitz in phase space when vv is unbounded. The field has at most linear growth and is integrable against ftf_t. The superposition principle for the continuity equation (Theorem 8.2.1 in [1]) represents ftf_t as a probability on its absolutely continuous characteristic curves. Local Lipschitz continuity and linear growth give uniqueness and global continuation of the ordinary differential equation for this fixed ff, so disintegration by the initial label is a Dirac mass. Thus ft=(Xt,Vt)#finf_t=(X_t,V_t)_{\#}f_{\mathrm{in}} for a deterministic flow; this does not imply uniqueness among distinct nonlinear solutions.

Step 4: regularity and identities. For any such characteristic solution the damping variation-of-constants formula first gives, on [0,T][0,T], ∣Vt(z)∣≤∣V0(z)∣+CT|V_t(z)|\le|V_0(z)|+C_T. The finite-time moment bound then justifies differentiation of the centered velocity variance, recovering its exact dissipation identity, and improves this to ∣Vt(z)∣≤∣V0(z)∣+κTV(0)|V_t(z)|\le|V_0(z)|+\kappa T\sqrt{\mathcal{V}(0)}. Integration of X˙t=Vt\dot X_t=V_t yields an L2(fin)L^2(f_{\mathrm{in}})-dominating function for positions and velocities, so the flow is locally continuous in L2(fin)L^2(f_{\mathrm{in}}). The same domination and convex truncations establish nonincrease of centered velocity mm-moments. The argument applies to every Lagrangian solution with the stated finite-time moments. □

Remark 3.2 (Relation to earlier solution theory). Chen–Yin [4] supply global noncompact weak existence under a first-moment hypothesis. Ha–Wang [12] and [13] give uniqueness or quantitative weak stability under additional density and exponential-velocity conditions. Neither result should be read as a blanket uniqueness theorem for all the atomic data admitted here. The argument above provides existence in the exact Lagrangian class needed for the a priori estimates, without such a uniqueness claim.

Finite time control of characteristics and main velocity localization estimate

We begin with a simple technical lemma that is helpful when we differentiate in time across integrals

Lemma 4.1 (Finite-time control of characteristics and velocity variance identity). Assume X(0)<∞\mathcal{X}(0)<\infty and V(0)<∞\mathcal{V}(0)<\infty. Then on each finite interval [0,T][0,T] the characteristic solution satisfies

sup⁡0≤t≤T∣Vt(z)∣≤∣V0(z)∣+κTV(0),(31)\sup_{0\leq t\leq T}|V_t(z)|\leq|V_0(z)|+\kappa T\sqrt{\mathcal{V}(0)}, \tag*{(31)}
sup⁡0≤t≤T∣Xt(z)∣≤∣X0(z)∣+T∣V0(z)∣+κT22V(0).(32)\sup_{0\leq t\leq T}|X_t(z)|\leq|X_0(z)|+T|V_0(z)|+\frac{\kappa T^2}{2}\sqrt{\mathcal{V}(0)}. \tag*{(32)}

The velocity variance is locally absolutely continuous, and

V˙(t)=−κ∬ψ(∣Xt−Xt′∣)∣Vt−Vt′∣2 dfin dfin′≤0for a.e. t.(33)\dot{\mathcal{V}}(t)=-\kappa\iint\psi(|X_t-X'_t|)|V_t-V'_t|^2\,\mathrm{d}f_{\mathrm{in}}\,\mathrm{d}f'_{\mathrm{in}}\leq0\qquad\text{for a.e. }t. \tag*{(33)}

When an initial mm’th velocity moment is finite, the same is true on every compact time interval.

Proof. Write the characteristic equation in damping form V˙t=−κat(Xt)Vt+κbt(Xt)\dot{V}_t=-\kappa a_t(X_t)V_t+\kappa b_t(X_t), where 0≤at≤10\leq a_t\leq1 and ∣bt(x)∣≤m1(t):=∫∣Vt∣ dfin|b_t(x)|\leq m_1(t):=\int|V_t|\,\mathrm{d}f_{\mathrm{in}}. The scalar integrating-factor formula, followed by integration in finf_{\mathrm{in}}, yields

∣Vt(z)∣≤∣V0(z)∣+κ∫0tm1(s) ds,m1(t)≤m1(0)eκt.|V_t(z)|\leq|V_0(z)|+\kappa\int_0^t m_1(s)\,\mathrm{d}s,\qquad m_1(t)\leq m_1(0)e^{\kappa t}.

The stipulated local L2L^2 continuity ensures that the quantities in this initial argument are locally finite. For t≤Tt\leq T these inequalities give a pointwise bound ∣Vt(z)∣≤∣V0(z)∣+CT|V_t(z)|\leq|V_0(z)|+C_T; hence the second moment, and any initially finite higher polynomial velocity moment, admit a uniform integrable dominator on [0,T][0,T]. We may consequently differentiate ∫∣Vt∣2 dfin\int|V_t|^2\,\mathrm{d}f_{\mathrm{in}} and symmetrize, obtaining (33). Thus V(t)≤V(0)\mathcal{V}(t)\leq\mathcal{V}(0), whence ∣bt(x)∣≤m1(t)≤V(0)|b_t(x)|\leq m_1(t)\leq\sqrt{\mathcal{V}(0)}. Returning to the integrating-factor formula gives (31); integrating X˙t=Vt\dot{X}_t=V_t gives (32). For a higher convex polynomial test function, its differentiated integrand is dominated on [0,T][0,T] by a constant times 1+∣V0(z)∣m+∣V0(z′)∣m1+|V_0(z)|^m+|V_0(z')|^m, which permits direct dominated convergence, or equivalently a truncation argument.

With the help of the above lemma we get

Lemma 4.2 (Convex velocity moments). For a convex differentiable G:Rd→RG:\mathbb{R}^d\to\mathbb{R} of suitable polynomial growth, the map t↦∫G(Vt) dfint\mapsto\int G(V_t)\,\mathrm{d}f_{\mathrm{in}} is nonincreasing. In particular, for m≥2m\geq2,

∫∣Vt∣m dfin≤∫∣V0∣m dfin.(34)\int|V_t|^m\,\mathrm{d}f_{\mathrm{in}}\leq\int|V_0|^m\,\mathrm{d}f_{\mathrm{in}}. \tag*{(34)}

Proof. The identity

ddt∫G(Vt) dfin=−κ2∬ψ(∣Xt−Xt′∣)(∇G(Vt)−∇G(Vt′))⋅(Vt−Vt′) dfin dfin′\frac{\mathrm{d}}{\mathrm{d}t}\int G(V_t)\,\mathrm{d}f_{\mathrm{in}}=-\frac{\kappa}{2}\iint\psi(|X_t-X'_t|)(\nabla G(V_t)-\nabla G(V'_t))\cdot(V_t-V'_t)\,\mathrm{d}f_{\mathrm{in}}\,\mathrm{d}f'_{\mathrm{in}}

is nonpositive by monotonicity of the gradient of a convex function. The previous finite-time bounds justify the identity for G(V)=∣V∣mG(V)=|V|^m by truncation.

Lemma 4.3 (Localized velocity coercivity). For every R>0R>0, we define

BR(t):={∣Xt∣≤R},QR(t):=fin(BR(t)),TR(t):=∫BR(t)c∣Vt∣2 dfin.B_R(t):=\{|X_t|\leq R\},\qquad Q_R(t):=f_{\mathrm{in}}(B_R(t)),\qquad T_R(t):=\int_{B_R(t)^{\mathrm{c}}}|V_t|^2\,\mathrm{d}f_{\mathrm{in}}.

Then for almost every tt,

V˙(t)≤−2κψ(2R)[(1−QR(t))V(t)−TR(t)].(35)\dot{\mathcal{V}}(t)\leq-2\kappa\psi(2R)\left[(1-Q_R(t))\mathcal{V}(t)-T_R(t)\right]. \tag*{(35)}

Proof. On BR×BRB_R \times B_R, the communication weight is at least ψ(2R)\psi(2R). The exact identity

∬BR×BR∣Vt−Vt′∣2 dfin dfin′=2(1−QR)∫BR∣Vt∣2 dfin−2∣∫BRVt dfin∣2\iint_{B_R \times B_R} \lvert V_t-V_t'\rvert^2\,\mathrm{d}f_{\mathrm{in}}\,\mathrm{d}f_{\mathrm{in}}'=2(1-Q_R)\int_{B_R}\lvert V_t\rvert^2\,\mathrm{d}f_{\mathrm{in}}-2\left\lvert\int_{B_R}V_t\,\mathrm{d}f_{\mathrm{in}}\right\rvert^2

and the zero-mean condition give ∣∫BRVt dfin∣2=∣∫BRVt dfin∣2≤QRTR\left\lvert\int_{B_R}V_t\,\mathrm{d}f_{\mathrm{in}}\right\rvert^2=\left\lvert\int_{B_R}V_t\,\mathrm{d}f_{\mathrm{in}}\right\rvert^2\leq Q_RT_R. Consequently the restricted double integral is bounded below by 2[(1−QR)V−TR]2[(1-Q_R)\mathcal{V}-T_R]. Combining this with (33) proves (35).

Weak flocking and spatial characteristic confinement

In the result that follows, we put to use the coercivity estimate we gave in Lemma 4.3 to prove an inequality for the dissipation of the velocity variance in terms of the spatial and velocity variances. Then we introduce a Lyapunov functional to prove control of spatial variance, and show weak flocking.

Lemma 5.1 (The Lyapunov functional). Assume the hypotheses of Theorem 2.3, and V(0)>0\mathcal{V}(0)>0. Set

γ=m2(m−2),η=βγ=βm2(m−2),δ=12−η>0.(36)\gamma=\frac{m}{2(m-2)},\qquad\eta=\beta\gamma=\frac{\beta m}{2(m-2)},\qquad\delta=\frac{1}{2}-\eta>0. \tag*{(36)}

There exists k>0k>0 depending only on the initial moments and κ,β,m\kappa,\beta,m such that

V˙(t)≤−kZ(t)−βV(t)1+η(37)\dot{\mathcal{V}}(t)\leq-k\mathcal{Z}(t)^{-\beta}\mathcal{V}(t)^{1+\eta} \tag*{(37)}

for almost every tt with V(t)>0\mathcal{V}(t)>0. Moreover, the functional

L(t):=Z(t)1−β+1−βkδV(t)δ(38)\mathcal{L}(t):=\mathcal{Z}(t)^{1-\beta}+\frac{1-\beta}{k\delta}\mathcal{V}(t)^\delta \tag*{(38)}

is dissipative, i.e. $\dot{\mathcal{L}}(t)\leq 0.

Proof. The spatial moment gives

Z˙(t)=X˙(t)2Z(t)=∫Xt⋅Vt dfinZ(t)≤V(t).(39)\dot{\mathcal{Z}}(t)=\frac{\dot{\mathcal{X}}(t)}{2\mathcal{Z}(t)}=\frac{\int X_t\cdot V_t\,\mathrm{d}f_{\mathrm{in}}}{\mathcal{Z}(t)}\leq\sqrt{\mathcal{V}(t)}. \tag*{(39)}

At a time where V(t)>0\mathcal{V}(t)>0, choose

R(t)=AZ(t)(V(0)V(t))γ,(40)R(t)=A\mathcal{Z}(t)\left(\frac{\mathcal{V}(0)}{\mathcal{V}(t)}\right)^\gamma, \tag*{(40)}

with A≥2A\geq2 fixed below. Since V(t)≤V(0)\mathcal{V}(t)\leq\mathcal{V}(0), Chebyshev’s inequality and (34) give

QR(t)(t)≤A−2(V(t)V(0))2γ,(41)Q_{R(t)}(t)\leq A^{-2}\left(\frac{\mathcal{V}(t)}{\mathcal{V}(0)}\right)^{2\gamma}, \tag*{(41)}
TR(t)(t)≤Mm2/mQR(t)(t)1−2/m≤Mm2/mA−2(m−2)/mV(t)V(0).(42)T_{R(t)}(t)\leq M_m^{2/m}Q_{R(t)}(t)^{1-2/m}\leq M_m^{2/m}A^{-2(m-2)/m}\frac{\mathcal{V}(t)}{\mathcal{V}(0)}. \tag*{(42)}

Choose AA so large that A−2≤1/4A^{-2}\leq1/4 and Mm2/mA−2(m−2)/m/V(0)≤1/4M_m^{2/m}A^{-2(m-2)/m}/\mathcal{V}(0)\leq1/4. Then QR(t)≤1/4Q_{R(t)}\leq1/4 and TR(t)≤V(t)/4T_{R(t)}\leq\mathcal{V}(t)/4. Since (35) follows pointwise from the variance identity at each of its differentiability times and is valid for every R>0R > 0 at such a time, substituting the time-dependent radius causes no moving-boundary differentiation issue. Thus (35) becomes

V˙(t)≤−κψ(2R(t))V(t).(43)\dot{\mathcal{V}}(t) \le-\kappa\psi(2R(t))\mathcal{V}(t). \tag*{(43)}

Writing Λ=(V(0)/V(t))γ≥1\Lambda= (\mathcal{V}(0)/\mathcal{V}(t))^\gamma\ge1 and using Z≥1\mathcal{Z} \ge1,

ψ(2R(t))=(1+4A2Z2Λ2)−β/2≥(1+4A2)−β/2Z−β(V(t)V(0))η.\begin{aligned} \psi(2R(t)) &= (1+4A^2\mathcal{Z}^2\Lambda^2)^{-\beta/2}\\ &\ge(1+4A^2)^{-\beta/2}\mathcal{Z}^{-\beta}\left(\frac{\mathcal{V}(t)}{\mathcal{V}(0)}\right)^\eta. \end{aligned}

Thus (37) holds with

k=κ(1+4A2)−β/2V(0)−η>0.(44)k=\kappa(1+4A^2)^{-\beta/2}\mathcal{V}(0)^{-\eta}>0. \tag*{(44)}

Differentiating the two terms in (38), and using (39) and δ+η=1/2\delta+\eta=1/2, gives

L˙(t)≤(1−β)Z(t)−βV(t)−1−βkδkZ(t)−βV(t)δ+η=0.\dot{\mathcal{L}}(t)\le(1-\beta)\mathcal{Z}(t)^{-\beta}\sqrt{\mathcal{V}(t)}-\frac{1-\beta}{k\delta}k\mathcal{Z}(t)^{-\beta}\mathcal{V}(t)^{\delta+\eta}=0.

The estimates extend through possible zero-variance times by the absorbing character of perfect alignment, or by limiting arguments.

Corollary 5.2 (Spatial cohesion and integrable velocity variance). Under the hypotheses of Theorem 2.3, Z(t)≤Z∗\mathcal{Z}(t)\le\mathcal{Z}_* uniformly in time. In particular,

V˙(t)≤−k∗V(t)1+η,V(t)≤(V(0)−η+ηk∗t)−1/η,(45)\dot{\mathcal{V}}(t)\le-k_*\mathcal{V}(t)^{1+\eta},\qquad\mathcal{V}(t)\le\left(\mathcal{V}(0)^{-\eta}+\eta k_*t\right)^{-1/\eta}, \tag*{(45)}

where k∗=kZ∗−β>0k_* = k\mathcal{Z}_*^{-\beta}>0. Moreover,

∫0∞V(t) dt<∞.(46)\int_0^\infty\sqrt{\mathcal{V}(t)}\,\mathrm{d}t<\infty. \tag*{(46)}

Proof. Monotonicity of (38) gives the explicit bound

Z(t)≤Z∗:=(Z(0)1−β+1−βkδV(0)δ)1/(1−β).\mathcal{Z}(t)\le\mathcal{Z}_*:=\left(\mathcal{Z}(0)^{1-\beta}+\frac{1-\beta}{k\delta}\mathcal{V}(0)^\delta\right)^{1/(1-\beta)}.

Integrating (37) with Z(t)≤Z∗\mathcal{Z}(t)\le\mathcal{Z}_* proves (45). Since η<1/2\eta<1/2, V(t)1/2≲(1+t)−1/(2η)\mathcal{V}(t)^{1/2}\lesssim(1+t)^{-1/(2\eta)} is integrable; indeed the polynomial estimate gives the explicit bound

∫0∞V(t) dt≤V(0)δk∗δ<∞,δ=12−η>0.\int_0^\infty\sqrt{\mathcal{V}(t)}\,\mathrm{d}t\le\frac{\mathcal{V}(0)^\delta}{k_*\delta}<\infty,\qquad\delta=\frac{1}{2}-\eta>0.

Proof of Corollary 2.4. For t≤st\le s, the fundamental theorem of calculus and Minkowski’s inequality yield ∥Xs−Xt∥L2(fin)≤∫tsV(τ) dτ\|X_s-X_t\|_{L^2(f_{\mathrm{in}})}\le\int_t^s\sqrt{\mathcal{V}(\tau)}\,\mathrm{d}\tau. The bound (45) implies

∫t∞V(τ) dτ≤C(1+t)1−1/(2η).\int_t^\infty\sqrt{\mathcal{V}(\tau)}\,\mathrm{d}\tau\le C(1+t)^{1-1/(2\eta)}.

The Cauchy criterion gives Xt→X∞X_t\to X_\infty strongly in L2(fin)L^2(f_{\mathrm{in}}) with that rate, and the canonical coupling gives the same bound in W2W_2.

We now proceed to the proof of the spatial characteristic estimate

Lemma 5.3 (Uniform control of spatial characteristic). Under the hypotheses of Theorem 2.3, we get

sup⁡t≥0∣Xt(z)∣≤C(1+∣X0(z)∣+∣V0(z)∣)1/(1−β)(47)\sup_{t\ge0}|X_t(z)| \le C\left(1+|X_0(z)|+|V_0(z)|\right)^{1/(1-\beta)} \tag*{(47)}

for fin-a.e. zz, with CC controlled by the initial averaged moments and κ,β\kappa,\beta.

Proof. By (18), choose R0<∞R_0<\infty such that ∫{∣Xt∣≤R0}dfin≥1/2\int_{\{|X_t|\le R_0\}} \mathrm{d}f_{\mathrm{in}}\ge1/2 for almost every t>0t>0. The characteristic velocity equation is

V˙t=−κat(Xt)Vt+κbt(Xt),(48)\dot{V}_t=-\kappa a_t(X_t)V_t+\kappa b_t(X_t), \tag*{(48)}

for

at(x)=∫ψ(∣x−Xt′∣) dfin′,bt(x)=∫ψ(∣x−Xt′∣)Vt′ dfin′.a_t(x)=\int\psi(|x-X'_t|)\,\mathrm{d}f'_{\mathrm{in}},\qquad b_t(x)=\int\psi(|x-X'_t|)V'_t\,\mathrm{d}f'_{\mathrm{in}}.

For every x∈Rdx\in\mathbb{R}^d and $t\ge 0,

at(x)≥12ψ(∣x∣+R0)≥c0(1+∣x∣)−β,c0=12(1+R0)−β,(49)a_t(x)\ge\frac{1}{2}\psi(|x|+R_0)\ge c_0(1+|x|)^{-\beta},\qquad c_0=\frac{1}{2}(1+R_0)^{-\beta}, \tag*{(49)}

and

∣bt(x)∣≤∫∣Vt′∣ dfin′≤V(t).(50)|b_t(x)|\le\int|V'_t|\,\mathrm{d}f'_{\mathrm{in}}\le\sqrt{\mathcal{V}(t)}. \tag*{(50)}

Fix a label zz and put r(t)=∣Xt(z)∣r(t)=|X_t(z)|, u(t)=∣Vt(z)∣u(t)=|V_t(z)|. The upper Dini derivatives satisfy

D+u(t)≤−κc0(1+r)−βu+κV(t),D+r(t)≤u(t).(51)D^+u(t)\le-\kappa c_0(1+r)^{-\beta}u+\kappa\sqrt{\mathcal{V}(t)},\qquad D^+r(t)\le u(t). \tag*{(51)}

The use of upper Dini derivatives defined by

D+u(t):=lim sup⁡h↓0u(t+h)−u(t)hD^+u(t):=\limsup_{h\downarrow0}\frac{u(t+h)-u(t)}{h}

accounts for the possible vanishing of r(t)r(t) or u(t)u(t). At nonzero velocity, the first estimate follows by taking the scalar product of the velocity equation with the unit velocity vector. At zero velocity, its upper right Dini derivative is bounded by just the magnitude of the forcing term. Similarly, the position equation implies that the upper right Dini derivative of the spatial radius does not exceed the velocity magnitude.

Now, for α=1−β\alpha=1-\beta,

D+(1+r)α≤α(1+r)−βu.D^+(1+r)^\alpha\le\alpha(1+r)^{-\beta}u.

Therefore the damping terms cancel in the combination

D+[u(t)+κc0α(1+r(t))α]≤κV(t).(52)D^+\left[u(t)+\frac{\kappa c_0}{\alpha}(1+r(t))^\alpha\right]\le\kappa\sqrt{\mathcal{V}(t)}. \tag*{(52)}

Integrating, by (46), gives

(1+r(t))α≤(1+r(0))α+ακc0u(0)+αc0∫0∞V(s) ds.(53)(1+r(t))^\alpha\le(1+r(0))^\alpha+\frac{\alpha}{\kappa c_0}u(0)+\frac{\alpha}{c_0}\int_0^\infty\sqrt{\mathcal{V}(s)}\,\mathrm{d}s. \tag*{(53)}

The additivity and elementary power inequalities yield (47). The constant c0c_0 is independent of the characteristic label; thus the final coefficient is independent of any support diameter.

Proof of Theorem 2.3. If V(0)=0\mathcal{V}(0)=0, then Vt=0V_t=0 and Xt=X0X_t=X_0 almost everywhere; the conclusions follow. Otherwise Lemma 5.1 and Corollary 5.2 give the spatial second-moment and velocity-integrability conclusions, and Lemma 5.3 gives characteristic confinement. Equation (21) follows from (45).

Exponential tails and stretched-exponential velocity variance

Having established the uniform spatial characteristic control inequality, we can use it to prove propagation of exponential tails

Proof of the spatial-tail assertion of Theorem 2.5. The initial exponential velocity moment provides a finite moment of order m>2/(1−β)m > 2/(1-\beta), while the initial exponential spatial moment provides X(0)<∞\mathcal{X}(0) < \infty. Thus Theorem 2.3 applies. It gives

sup⁡t≥0∣Xt(z)∣≤C(1+∣X0(z)∣+∣V0(z)∣1/(1−β)).\sup_{t\geq0}|X_t(z)| \leq C\left(1+|X_0(z)|+|V_0(z)|^{1/(1-\beta)}\right).

For σ=min⁡{p,(1−β)q}\sigma=\min\{p,(1-\beta)q\}, there is C1C_1 such that

∣Xt(z)∣σ≤C1(1+∣X0(z)∣p+∣V0(z)∣q).(54)|X_t(z)|^\sigma\leq C_1\left(1+|X_0(z)|^p+|V_0(z)|^q\right). \tag*{(54)}

Choosing a∗>0a_* > 0 with a∗C1≤min⁡{a,b}a_*C_1 \leq\min\{a,b\}, taking the exponent and integrating proves (24).

Next comes the proof of stretched-exponential velocity variance. First we prove it in Proposition 2.8, and then in Theorem 2.5

Proof of Proposition 2.8. Assume that for some σ\sigma, a>0a > 0 and Mx<∞M_x < \infty,

sup⁡t≥0∫ea∣Xt(z)∣σ dfin(z)≤Mx,(55)\sup_{t\geq0}\int e^{a|X_t(z)|^\sigma}\,\mathrm{d}f_{\mathrm{in}}(z) \leq M_x, \tag*{(55)}

and that for some r=2+ε>2r=2+\varepsilon>2,

Mr:=∫∣V0(z)∣r dfin(z)<∞.(56)M_r:=\int|V_0(z)|^r\,\mathrm{d}f_{\mathrm{in}}(z)<\infty. \tag*{(56)}

Then there exist constants c,C>0c,C>0 such that

V(t)≤Cexp⁡[−c(1+t)σ/(σ+β)],t≥0.(57)\mathcal{V}(t)\leq C\exp\left[-c(1+t)^{\sigma/(\sigma+\beta)}\right],\qquad t\geq0. \tag*{(57)}

Step 1: The spatial mass and velocity energy outside a ball. For any R>0R>0 and t≥0t\geq0, remember that

BR(t)={z:∣Xt(z)∣≤R},QR(t)=fin(BR(t)c),TR(t)=∫BR(t)c∣Vt(z)∣2 dfin(z).B_R(t)=\{z:|X_t(z)|\leq R\},\qquad Q_R(t)=f_{\mathrm{in}}(B_R(t)^c),\qquad T_R(t)=\int_{B_R(t)^c}|V_t(z)|^2\,\mathrm{d}f_{\mathrm{in}}(z).

Since 1{∣Xt∣>R}≤e−aRσea∣Xt∣σ\mathbf{1}_{\{|X_t|>R\}}\leq e^{-aR^\sigma}e^{a|X_t|^\sigma}, the spatial exponential moment (55) gives, by Markov’s inequality,

QR(t)≤Mxe−aRσ(R>0, t≥0).(58)Q_R(t)\leq M_xe^{-aR^\sigma}\qquad(R>0,\ t\geq0). \tag*{(58)}

Convexity of u↦∣u∣ru\mapsto|u|^r and the alignment equation imply the nonincrease of centered velocity moments:

∫∣Vt∣r dfin≤Mr(t≥0).(59)\int|V_t|^r\,\mathrm{d}f_{\mathrm{in}}\leq M_r\qquad(t\geq0). \tag*{(59)}

By Hölder’s inequality with exponents r/2r/2 and r/(r−2)r/(r-2),

TR(t)≤(∫BR(t)c∣Vt∣r dfin)2/rQR(t)(r−2)/r≤Mr2/rMx(r−2)/rexp⁡(−a(r−2)rRσ).(60)\begin{aligned} T_R(t)\leq\left(\int_{B_R(t)^c}|V_t|^r\,\mathrm{d}f_{\mathrm{in}}\right)^{2/r}Q_R(t)^{(r-2)/r} \\ &\le M_r^{2/r}M_x^{(r-2)/r}\exp\left(-\frac{a(r-2)}{r}R^\sigma\right). \tag*{(60)} \end{aligned}

Thus, with

D:=Mr2/rMx(r−2)/r,b:=a(r−2)r>0,(61)D:=M_r^{2/r}M_x^{(r-2)/r},\qquad b:=\frac{a(r-2)}{r}>0, \tag*{(61)}

we have TR(t)≤De−bRσT_R(t)\le De^{-bR^\sigma}, uniformly in tt. Step 2: Localized variance dissipation. The exact variance identity is

V˙(t)=−κ∬ψ(∣Xt−Xt′∣)∣Vt−Vt′∣2 dfin dfin′≤0for a.e. t.(62)\dot{\mathcal V}(t)=-\kappa\iint\psi(|X_t-X'_t|)|V_t-V'_t|^2\,\mathrm{d}f_{\mathrm{in}}\,\mathrm{d}f'_{\mathrm{in}}\le0\quad\text{for a.e. }t. \tag*{(62)}

On BR(t)×BR(t)B_R(t)\times B_R(t) we have ∣Xt−Xt′∣≤2R|X_t-X'_t|\le2R, hence ψ(∣Xt−Xt′∣)≥ψ(2R)\psi(|X_t-X'_t|)\ge\psi(2R). Moreover, since ∫Vt dfin=0\int V_t\,\mathrm{d}f_{\mathrm{in}}=0,

∬BR×BR∣Vt−Vt′∣2 dfin dfin′=2(1−QR)(V−TR)−2∣∫BRVt dfin∣2\iint_{B_R\times B_R}|V_t-V'_t|^2\,\mathrm{d}f_{\mathrm{in}}\,\mathrm{d}f'_{\mathrm{in}} =2(1-Q_R)(\mathcal V-T_R)-2\left|\int_{B_R}V_t\,\mathrm{d}f_{\mathrm{in}}\right|^2
≥2(1−QR)(V−TR)−2QRTR=2[(1−QR)V−TR].\ge2(1-Q_R)(\mathcal V-T_R)-2Q_RT_R =2\left[(1-Q_R)\mathcal V-T_R\right].

We have therefore proved, for every fixed R>0R>0 and almost every tt,

V˙(t)≤−2κψ(2R)[(1−QR(t))V(t)−TR(t)].(63)\dot{\mathcal V}(t)\le-2\kappa\psi(2R)\left[(1-Q_R(t))\mathcal V(t)-T_R(t)\right]. \tag*{(63)}

Choose R0≥1R_0\ge1 so large that Mxe−aR0σ≤12M_xe^{-aR_0^\sigma}\le\frac12. For R≥R0R\ge R_0, (58) ensures 1−QR(t)≥121-Q_R(t)\ge\frac12, and (60) in (35) yields

V˙(t)≤−κψ(2R)V(t)+2κDψ(2R)e−bRσ,R≥R0.(64)\dot{\mathcal V}(t)\le-\kappa\psi(2R)\mathcal V(t)+2\kappa D\psi(2R)e^{-bR^\sigma},\qquad R\ge R_0. \tag*{(64)}

This estimate is valid pointwise in the parameter RR at almost every time at which the variance identity holds. Choosing R=R(t)R=R(t) below requires no differentiation of moving spatial sets.

Step 3: Choose a growing radius to balance the two terms. Set

ϑ:=σσ+β∈(0,1),s:=1+t,R(t):=As1/(σ+β),(65)\vartheta:=\frac{\sigma}{\sigma+\beta}\in(0,1),\qquad s:=1+t,\qquad R(t):=As^{1/(\sigma+\beta)}, \tag*{(65)}

where A≥max⁡{1,R0}A\ge\max\{1,R_0\}. We have e−bR(t)σ=e−Bsϑe^{-bR(t)^\sigma}=e^{-Bs^\vartheta} with B:=bAσB:=bA^\sigma. Furthermore, because s≥1s\ge1,

4A2s2/(σ+β)≤1+4A2s2/(σ+β)≤(1+4A2)s2/(σ+β),4A^2s^{2/(\sigma+\beta)}\le1+4A^2s^{2/(\sigma+\beta)}\le(1+4A^2)s^{2/(\sigma+\beta)},

so

(1+4A2)−β/2sϑ−1≤ψ(2R(t))≤(2A)−βsϑ−1.(66)(1+4A^2)^{-\beta/2}s^{\vartheta-1}\le\psi(2R(t))\le(2A)^{-\beta}s^{\vartheta-1}. \tag*{(66)}

Using the lower bound from (66) for the negative dissipation term and the upper bound for the positive error term in (64), we obtain

V˙(t)≤−k0sϑ−1V(t)+d0sϑ−1e−Bsϑ,(67)\dot{\mathcal V}(t)\le-k_0s^{\vartheta-1}\mathcal V(t)+d_0s^{\vartheta-1}e^{-Bs^\vartheta}, \tag*{(67)}

where the constants are explicitly

k0:=κ(1+4A2)−β/2,d0:=2κD(2A)−β,B:=bAσ.(68)k_0:=\kappa(1+4A^2)^{-\beta/2},\qquad d_0:=2\kappa D(2A)^{-\beta},\qquad B:=bA^\sigma. \tag*{(68)}

The choice R(t)≍(1+t)1/(σ+β)R(t)\asymp(1+t)^{1/(\sigma+\beta)} balances the exponents tR−βtR^{-\beta} (cumulative dissipation) and RσR^\sigma (spatial-tail error).

Step 4: An explicit scalar supersolution. Choose

0<c<min⁡{B,k0ϑ}(69)0<c<\min\left\{B,\frac{k_0}{\vartheta}\right\} \tag*{(69)}

and choose

K≥max⁡{ecV(0),d0k0−cϑ}.(70)K\ge\max\left\{e^{c}\mathcal{V}(0),\frac{d_0}{k_0-c\vartheta}\right\}. \tag*{(70)}

Define W(t)=Ke−c(1+t)ϑ=Ke−csϑW(t)=Ke^{-c(1+t)^\vartheta}=Ke^{-cs^\vartheta}. Then W(0)≥V(0)W(0)\ge\mathcal{V}(0) and

W˙(t)+k0sϑ−1W(t)=(k0−cϑ)Ksϑ−1e−csϑ≥d0sϑ−1e−Bsϑ,\begin{aligned} \dot{W}(t)+k_0s^{\vartheta-1}W(t)=(k_0-c\vartheta)Ks^{\vartheta-1}e^{-cs^\vartheta} \\ &\ge d_0s^{\vartheta-1}e^{-Bs^\vartheta}, \end{aligned}

where we used c<Bc<B and (70). Therefore WW is a supersolution of the scalar inequality (67). The comparison principle gives

V(t)≤W(t)=Kexp⁡[−c(1+t)σ/(σ+β)],t≥0,\mathcal{V}(t)\le W(t)=K\exp\left[-c(1+t)^{\sigma/(\sigma+\beta)}\right],\qquad t\ge0,

which is (57).

Completion of Theorem 2.5. The exponential velocity moment in (22) yields some moment strictly higher than order two. Apply Proposition 2.8 to the uniform spatial tail proved above, obtaining (25).

Proof of Corollary 2.6. The uniform exponential moment implies every polynomial spatial moment bound. For each ss and large RR, ∣X∣s1∣X∣>R≤Cse−a∗Rσ/2ea∗∣X∣σ|X|^s1_{|X|>R}\le C_se^{-a_*R^\sigma/2}e^{a_*|X|^\sigma}, which proves uniform integrability. Since X˙t=Vt\dot{X}_t=V_t,

∥Xt−Xs∥L2(fin)≤∫stV(τ) dτ.\left\|X_t-X_s\right\|_{L^2(f_{\mathrm{in}})}\le\int_s^t\sqrt{\mathcal{V}(\tau)}\,\mathrm{d}\tau.

The latter tail integral decays like e−c′(1+s)ϑe^{-c'(1+s)^\vartheta}, after reducing c>0c>0 to absorb the polynomial prefactor. Thus XtX_t converges in L2(fin)L^2(f_{\mathrm{in}}) to a random vector X∞X_\infty with the asserted rate. The coupling (Xt,X∞)#fin(X_t,X_\infty)_{\#}f_{\mathrm{in}} gives the W2W_2 estimate; Fatou’s lemma transfers the spatial exponential moment to the limit.

A countably atomic sharpness theorem

Theorem 7.1 (Optimal spatial-tail-limited exponent for measures). Fix d≥1d\ge1, κ>0\kappa>0, 0<β<10<\beta<1, and p>0p>0. There exists a global Lagrangian probability solution of the kinetic Cucker–Smale equation, supported on a fixed line in Rd\mathbb{R}^d, with bounded initial velocities and unbounded spatial support, such that for some a,c,C>0a,c,C>0,

sup⁡t≥0∫ea∣x−xc(t)∣p dft<∞,V(t)≥cexp⁡[−C(1+t)p/(p+β)],t≥0.(71)\sup_{t\ge0}\int e^{a|x-x_c(t)|^p}\,\mathrm{d}f_t<\infty,\qquad\mathcal{V}(t)\ge c\exp\left[-C(1+t)^{p/(p+\beta)}\right],\qquad t\ge0. \tag*{(71)}

Hence the exponent p/(p+β)p/(p+\beta) in the velocity-variance upper bound is optimal for the class of kinetic probability measures having a uniform spatial exponential moment of order pp and bounded initial velocity support. The conclusion concerns atomic measure solutions; it does not assert sharpness for smooth strictly positive densities.

Proof. It suffices to construct the solution for d=1d=1 and embed the positions and velocities into one coordinate line for general dd. Set α=1−β\alpha=1-\beta, choose A>1A>1 later, and put

Rn=An,mn=εe−Rnp(n≥1),m0=1−∑n≥1mn≥12.R_n=A^n,\qquad m_n=\varepsilon e^{-R_n^p}\quad(n\ge1),\qquad m_0=1-\sum_{n\ge1}m_n\ge\frac{1}{2}.

with ε>0\varepsilon>0 sufficiently small. Take the symmetric datum

μ=m0δ(0,0)+∑n≥1mn2(δ(Rn,1)+δ(−Rn,−1)).(72)\mu=m_0\delta_{(0,0)}+\sum_{n\ge1}\frac{m_n}{2}\left(\delta_{(R_n,1)}+\delta_{(-R_n,-1)}\right). \tag*{(72)}

The central characteristic remains at the origin by symmetry. Write (xn(t),un(t))(x_n(t),u_n(t)) for the positive satellites, with (xn(0),un(0))=(Rn,1)(x_n(0),u_n(0))=(R_n,1).

Global countable dynamics. Introduce yn=xn−Rny_n=x_n-R_n, and regard (y,u)(y,u) as an unknown in ℓ∞×ℓ∞\ell^\infty\times\ell^\infty, initially (0,1)(0,1). The equations are y˙n=un\dot{y}_n=u_n and

u˙n=−κanun+κ2∑j≥1mj[ψ(∣xn−xj∣)−ψ(xn+xj)]uj(73)\dot{u}_n=-\kappa a_nu_n+\frac{\kappa}{2}\sum_{j\ge1}m_j\left[\psi(|x_n-x_j|)-\psi(x_n+x_j)\right]u_j \tag*{(73)}

where

an=m0ψ(xn)+12∑j≥1mj[ψ(∣xn−xj∣)+ψ(xn+xj)].(74)a_n=m_0\psi(x_n)+\frac{1}{2}\sum_{j\ge1}m_j\left[\psi(|x_n-x_j|)+\psi(x_n+x_j)\right]. \tag*{(74)}

Here and below the formula is used while xn≥0x_n\ge0. Since r↦ψ(∣r∣)r\mapsto\psi(|r|) is bounded and globally Lipschitz, the right-hand side, expressed with xn=Rn+ynx_n=R_n+y_n and the signed symmetric counterparts, maps ℓ∞×ℓ∞\ell^\infty\times\ell^\infty to itself and is locally Lipschitz on bounded sets, uniformly in nn: its differences are bounded by sums of the form C∑jmj(∥y−y~∥∞+∥u−u~∥∞)C\sum_jm_j(\lVert y-\widetilde{y}\rVert_\infty+\lVert u-\widetilde{u}\rVert_\infty). Thus the Banach-space Picard theorem provides a local solution even though (Rn)(R_n) itself is unbounded.

The full symmetric velocity dynamics is a weighted consensus system with nonnegative weights. The standard velocity maximum principle applied to the bounded sequence gives ∣un(t)∣≤1|u_n(t)|\le1; moreover ψ(∣xn−xj∣)≥ψ(xn+xj)\psi(|x_n-x_j|)\ge\psi(x_n+x_j) for xn,xj≥0x_n,x_j\ge0, so (73) is cooperative on the positive cone. Starting from un(0)=1u_n(0)=1, its integrating-factor formulation gives un(t)≥e−κt>0u_n(t)\ge e^{-\kappa t}>0 while all xj≥0x_j\ge0, since an≤1a_n\le1. Consequently xn(t)≥Rn>0x_n(t)\ge R_n>0, and the positivity bootstrap closes globally. Finally ∥u(t)∥∞≤1\lVert u(t)\rVert_\infty\le1 and ∥y(t)∥∞≤t\lVert y(t)\rVert_\infty\le t, precluding finite-time breakdown of the ℓ∞\ell^\infty solution. It follows that

0<un(t)≤1,xn(t)≥Rn(n≥1,t≥0).(75)0<u_n(t)\le1,\qquad x_n(t)\ge R_n\qquad(n\ge1,t\ge0). \tag*{(75)}

Uniform characteristic confinement. Let Ψ(r)=∫0rψ(∣s∣) ds\Psi(r)=\int_0^r\psi(|s|)\,\mathrm{d}s be the odd primitive. We justify the infinite primitive series appearing in the natural-velocity invariant. Since 0<ψ≤10<\psi\le1, we have ∣Ψ(s)∣≤∣s∣|\Psi(s)|\le|s|, and the specific satellite masses satisfy

∑j≥1mjRj=ε∑j≥1Aje−Ajp<∞.\sum_{j\ge1}m_jR_j=\varepsilon\sum_{j\ge1}A^je^{-A^jp}<\infty.

By (75), Rj≤xj(t)≤Rj+tR_j\le x_j(t)\le R_j+t. In particular, for fixed nn and 0≤t≤T0\le t\le T,

∣Ψ(xn(t)−xj(t))∣≤∣xn(t)∣+∣xj(t)∣≤Rn+Rj+2T.|\Psi(x_n(t)-x_j(t))|\le|x_n(t)|+|x_j(t)|\le R_n+R_j+2T.

The same bound holds for Ψ(xn(t)+xj(t))\Psi(x_n(t)+x_j(t)), so the mass-weighted primitive series (including the reflected satellites) converges absolutely and uniformly on each bounded time interval; the central-mass contribution is a single finite term. Furthermore,

∣ddtΨ(xn(t)+xj(t))∣=ψ(∣xn(t)+xj(t)∣)∣un(t)+uj(t)∣≤2.\left|\frac{\mathrm{d}}{\mathrm{d}t}\Psi(x_n(t)+x_j(t))\right|=\psi(|x_n(t)+x_j(t)|)|u_n(t)+u_j(t)|\le2.

As ∑jmj<∞\sum_j m_j < \infty, the series of derivatives converges uniformly on bounded time intervals as well. Termwise differentiation is therefore legitimate. The characteristic equations now show that the natural-velocity quantity

un(t)+κ∫Ψ(xn(t)−y) dρt(y)u_n(t) + \kappa\int\Psi(x_n(t)-y)\,d\rho_t(y)

is conserved. For x,y≥0x,y \ge0, monotonicity and oddness imply

0≤12[Ψ(x−y)+Ψ(x+y)]≤Ψ(x).0 \le\frac{1}{2}\left[\Psi(x-y)+\Psi(x+y)\right] \le\Psi(x).

The lower bound follows from x+y≥∣x−y∣x+y \ge|x-y|; the upper bound follows because the displayed average is nonincreasing as a function of y≥0y \ge0. Using un(t)≥0u_n(t) \ge0, the central mass, and the initial equality un(0)=1u_n(0)=1, we obtain

κm0Ψ(xn(t))≤1+κΨ(Rn).(76)\kappa m_0\Psi(x_n(t)) \le1+\kappa\Psi(R_n). \tag*{(76)}

There exist cβ,Cβ>0c_\beta,C_\beta>0 such that cβrα≤Ψ(r)≤Cβrαc_\beta r^\alpha\le\Psi(r) \le C_\beta r^\alpha for r≥1r \ge1. Since m0≥1/2m_0 \ge1/2 and Rn≥1R_n \ge1, (76) implies

Rn≤xn(t)≤C∗Rn,(77)R_n \le x_n(t) \le C_*R_n, \tag*{(77)}

where C∗C_* depends on β,κ\beta,\kappa but not on n,t,An,t,A or ε\varepsilon (provided m0≥1/2m_0 \ge1/2). Separation and the self-interaction term. Choose A>4C∗A>4C_*. By (77), for j<nj<n, ∣xn−xj∣≥3Rn/4|x_n-x_j| \ge3R_n/4; for j>nj>n, ∣xn−xj∣≥3Rj/4≥3Rn/4|x_n-x_j| \ge3R_j/4 \ge3R_n/4. Also xn+xj≥Rnx_n+x_j \ge R_n. Thus, for j≠nj\ne n,

ψ(∣xn−xj∣)+ψ(xn+xj)≤CRn−β.\psi(|x_n-x_j|)+\psi(x_n+x_j) \le CR_n^{-\beta}.

The j=nj=n summand in (74) has ψ(0)=1\psi(0)=1 and contributes at most mn/2m_n/2 plus CmnRn−βCm_nR_n^{-\beta}. Since mn≤εe−Rnp≤CpRn−βm_n \le\varepsilon e^{-R_n^p} \le C_pR_n^{-\beta}, this term is also absorbed. Hence

0≤an(t)≤C0Rn−β(n≥1,t≥0),(78)0 \le a_n(t) \le C_0R_n^{-\beta}\qquad(n\ge1,t\ge0), \tag*{(78)}

with C0C_0 independent of n,tn,t. The nonnegative off-diagonal term in (73) therefore gives

un(t)≥exp⁡(−κC0tRn−β).(79)u_n(t) \ge\exp(-\kappa C_0tR_n^{-\beta}). \tag*{(79)}

Uniform tails and the lower rate at every large time. The zero mean velocity implies

V(t)=∑n≥1mnun(t)2,\mathcal{V}(t)=\sum_{n\ge1}m_nu_n(t)^2,

and hence

V(t)≥εexp⁡[−Rnp−2κC0tRn−β]for each n≥1.(80)\mathcal{V}(t)\ge\varepsilon\exp\left[-R_n^p-2\kappa C_0tR_n^{-\beta}\right]\qquad\text{for each }n\ge1. \tag*{(80)}

For t≥Ap+βt\ge A^{p+\beta} choose the integer n≥1n\ge1 satisfying

An≤t1/(p+β)<An+1.A^n\le t^{1/(p+\beta)}<A^{n+1}.

Then A−1t1/(p+β)<Rn≤t1/(p+β)A^{-1}t^{1/(p+\beta)}<R_n\le t^{1/(p+\beta)}, so both terms in the exponent of (80) are bounded by C(1+t)p/(p+β)C(1+t)^{p/(p+\beta)}. Enlarging CC and decreasing the prefactor covers bounded times as well (the variance is positive there). This proves the second assertion in (71). Finally, for 0<a<C∗−p0<a<C_*^{-p}, (77) yields

sup⁡t≥0∫ea∣x∣p dft≤m0+∑n≥1mneaC∗pRnp<∞.\sup_{t\ge0}\int e^{a|x|^p}\,df_t\le m_0+\sum_{n\ge1}m_ne^{aC_*^pR_n^p}<\infty.

This proves the uniform exponential spatial moment and the theorem. □\square

Remark 7.2 (Infinite-particle background). Ha, Wang, and Xue [16] study countably infinite Cucker–Smale systems using infinite-norm inequalities and spectral criteria for clustering. The construction here instead uses a summably weighted symmetric atomic measure and the bounded ℓ∞\ell^\infty displacement–velocity system; its conclusion is a lower bound for the velocity variance of a kinetic probability solution, rather than a spectral characterization of clusters.

Discussion and open problems

The finite-moment theorem gives uniform spatial second moments, strong confinement along characteristics, integrable velocity fluctuations and a polynomial variance-decay estimate. Exponential initial position and velocity moments improve this to uniform exponential spatial tails, stretched-exponential velocity alignment, and convergence of the centered spatial law in W2W_2. The countably atomic construction proves the spatial-tail-limited rate exponent optimal in the measure-valued class, with bounded initial velocities. We do not infer sharpness in the smooth-density class or in the velocity-limited regime from that construction.

Several questions remain open and suggest directions for further investigation. First, it would be interesting to determine whether the finite-moment confinement theorem extends to the critical velocity-moment exponent, i.e. m=2/(1−β)m = 2/(1-\beta). A second direction concerns the optimality of the quantitative alignment estimates. Although our countably atomic construction establishes sharpness in the spatial-tail-limited regime, it remains unclear whether the same obstruction can be realized by smooth, strictly positive phase-space densities, or whether the predicted decay rate is optimal when the initial velocity tails determine the convergence exponent. Finally, an important challenge is to extend the confinement and tail-propagation results to critical communication decay, particularly the borderline case where the interaction weight decays inversely with distance i.e. β=1\beta= 1.

The borderline communication case β=1\beta= 1 has already been treated conditionally by Ha and Wang, who establish weak flocking with uniformly bounded centered spatial second moments and algebraic velocity alignment under an initial exponential spatial moment, bounded velocity support, and an explicit lower bound on the coupling strength. The question here is therefore not whether weak flocking can occur at β=1\beta= 1, but whether the support-independent characteristic confinement and strong spatial-tail propagation obtained above can be extended to that critical regime with unbounded velocity support and without such a coupling restriction, as well as to other borderline long-range communication kernels.

Declaration of Generative AI Use

During the preparation of this manuscript, the author used ChatGPT (OpenAI) as a research-assistance tool for exploring mathematical arguments, examining proof strategies, checking calculations, reviewing relevant literature, and improving the organization and clarity of the text. AI-generated suggestions and derivations were treated as material requiring mathematical verification, rather than as independent evidence of correctness. The author retains full responsibility for the validity of all mathematical statements, the accuracy of the references, and the final content of the manuscript.

References

  1. [1]L. Ambrosio, N. Gigli, and G. Savaré, Gradient Flows: In Metric Spaces and in the Space of Probability Measures, 2nd ed., Birkhäuser, 2008. doi:10.1007/978-3-7643-8722-8.
  2. [2]J. A. Cañizo, J. A. Carrillo, and J. Rosado, A well-posedness theory in measures for some kinetic models of collective motion, Math. Models Methods Appl. Sci. 21 (2011), 515–539. doi:10.1142/S0218202511005131.DOI
  3. [3]J. A. Carrillo, M. Fornasier, J. Rosado, and G. Toscani, Asymptotic flocking dynamics for the kinetic Cucker–Smale model, SIAM J. Math. Anal. 42 (2010), 218–236. doi:10.1137/090757290.DOI
  4. [4]Z. Chen and X. Yin, The kinetic Cucker–Smale model: Well-posedness and asymptotic behavior, SIAM J. Math. Anal. 51 (2019), 3819–3853. doi:10.1137/18M1215001.DOI
  5. [5]Y.-P. Choi, S.-Y. Ha and Z. Li, Emergent dynamics of the Cucker–Smale flocking model and its variants, in Modeling and Simulation in Science, Engineering and Technology, Springer, 2017, 299–331.arxiv.org/abs/1604.04887
  6. [7]F. Cucker and S. Smale, Emergent behavior in flocks, IEEE Trans. Automat. Control 52 (2007), 852–862. doi:10.1109/TAC.2007.895842.
  7. [8]S.-Y. Ha, Recent progress on the mean-field limit of the Cucker–Smale model for flocking, Séminaire Laurent Schwartz – EDP et applications, 2024–2025, Exposé XV, 17 pp., published 2025. doi:10.5802/slsedp.179.DOI
  8. [9]S.-Y. Ha and J.-G. Liu, A simple proof of the Cucker–Smale flocking dynamics and mean-field limit, Commun. Math. Sci. 7 (2009), 297–325. https://sites.math.duke.edu/~jliu/pdf/Ha_Liu_CMS_2009.pdf.DOI
  9. [10]S.-Y. Ha and E. Tadmor, From particle to kinetic and hydrodynamic descriptions of flocking, Kinet. Relat. Models 1 (2008), 415–435. doi:10.3934/krm.2008.1.415.DOI
  10. [11]S.-Y. Ha and X. Wang, Emergent dynamics of spatially extended relativistic kinetic Cucker–Smale model, Math. Ann. 393 (2025), 2265–2305. doi:10.1007/s00208-025-03288-2.DOI
  11. [12]S.-Y. Ha and X. Wang, On the weak flocking of the kinetic Cucker–Smale model in a fully noncompact support setting, Proc. London Math. Soc. 132 (2026), no. 6, article e70171. doi:10.1112/plms.70171.DOI
  12. [13]S.-Y. Ha and X. Wang, Weak stability and random mean-field limit of the phase-spatially extended kinetic Cucker–Smale model, arXiv preprint arXiv:2607.23906 (2026). https://arxiv.org/abs/2607.23906.
  13. [14]S.-Y. Ha and X. Wang, Weak flocking for phase-spatially extended kinetic Cucker–Smale equation in a confining force field, arXiv preprint arXiv:2608.22709 (2026). https://arxiv.org/abs/2608.22709.
  14. [15]S.-Y. Ha, X. Wang, and X. Xue, On the exponential weak flocking for the kinetic Cucker–Smale model with non-compact support, Math. Models Methods Appl. Sci. 35 (2025), 781–824. doi:10.1142/S0218202525500101.DOI
  15. [16]S.-Y. Ha, X. Wang, and X. Xue, Clustering and spectral analysis of the infinite Cucker–Smale model, Acta Math. Sin. (Engl. Ser.) 42 (2026), 821–880. doi:10.1007/s10114-026-5431-z.DOI
  16. [17]S.-Y. Ha, X. Wang, and W. Yoon, Uniform-in-time stability and mean-field limit of the Cucker–Smale-type model with noncompact support, arXiv preprint arXiv:2607.16794 (2026). https://arxiv.org/abs/2607.16794.
  17. [18]J. Haskovec and I. Markou, Asymptotic flocking in the Cucker–Smale model with reaction-type delays, Kinet. Relat. Models 13 (2020), no. 4, 795–813. doi:10.3934/krm.2020027.DOI
  18. [19]J. Haskovec and I. Markou, Exponential asymptotic flocking in the Cucker–Smale model with distributed reaction delays, Math. Biosci. Eng. 17 (2020), no. 5, 5651–5671. doi:10.3934/mbe.2020034.
  19. [20]I. Markou, Collision-avoiding in the singular Cucker–Smale model with nonlinear velocity couplings, Discrete and Continuous Dynamical Systems 38 (2018), 5245–5260. doi:10.3934/dcds.2018232.DOI
  20. [21]P. Minakowski, P. B. Mucha, J. Peszek and E. Zatorska, Singular Cucker–Smale dynamics, in Active Particles, Volume 2: Advances in Theory, Models, and Applications, Springer, 2019, 201–243.DOI
  21. [22]X. Wang and X. Xue, Formation behaviour of the kinetic Cucker–Smale model with non-compact support, Proc. R. Soc. Edinburgh Sect. A 153 (2023), 1315–1346. doi:10.1017/prm.2022.47.DOI

Paper details

vlnc:W_t3kurt5d
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

ChatGPT (OpenAI)

View statement
Looking for related papers…
Contents