betaFeedback

WorksMathematics

Reverse Transportation and Harnack Inequalities for Underdamped Langevin in the Non-convex Setting

Abstract

We establish a dimension-free long-time reverse transportation, respectively Harnack inequality for the underdamped Langevin dynamics with non-convex potential. Specifically, we prove that for any two initial points z,z′z,z' and PtP_t the underdamped Langevin semigroup, it holds \begin{equation} \mathrm{KL}(δ_z P_t|δ_{z'}P_t) \leq C\mathrm{e}^{-χt}|z-z'|(\mathcal{R}\vee|z-z'|) \end{equation} for all t>δt>δ where δ>0δ>0 is arbitrarily small and C=C(δ),R,χ>0C=C(δ),\mathcal{R},χ>0 are dimension free constants. In particular, our proof extends recent results which are restricted to either the convex underdamped or the non-convex but overdamped setting.

Introduction

In this article we are interested in the underdamped Langevin stochastic differential equation (SDE), that is, the stochastic process Zt=(Xt,Vt)Z_t=(X_t,V_t) satisfying

dXt=VtdtdVt={−γVt−∇U(Xt)}dt+2γdWt(1)\begin{aligned} \mathrm{d}X_t &= V_t\mathrm{d}t \\ \mathrm{d}V_t &= \{-\gamma V_t-\nabla U(X_t)\}\mathrm{d}t+\sqrt{2\gamma}\mathrm{d}W_t \tag*{(1)} \end{aligned}

with Brownian motion (Wt)t(W_t)_t, a friction parameter γ>0\gamma>0 and a potential U:Rd→RU:\mathbb{R}^d\to\mathbb{R}. Specifically, for possibly non-convex UU we derive a reverse transportation inequality for (ULD). That is, denoting the semigroup induced by (ULD) as PtP_t, we prove the existence of dimension-free constants C,χ,R>0C,\chi,\mathcal{R}>0 such that

KL⁡(δzPt∣δz′Pt)≤Ce−χt∣z−z′∣(R∨∣z−z′∣),(2)\operatorname{KL}(\delta_zP_t\mid\delta_{z'}P_t) \le C e^{-\chi t}|z-z'|(\mathcal{R}\vee|z-z'|), \tag*{(2)}

where KL⁡\operatorname{KL} denotes the Kullback–Leibler divergence. We may represent δzPt\delta_zP_t and δz′Pt\delta_{z'}P_t as the laws of ZtZ_t respectively Zt′Z'_t which solve ULD with initial condition Z0=zZ_0=z, Z0′=z′Z'_0=z' and z,z′∈R2dz,z'\in\mathbb{R}^{2d}, that is,

dXt=VtdtdVt={−γVt−∇U(Xt)}dt+2γdWt,(X0,V0)=zdXt′=Vt′dtdVt′={−γVt′−∇U(Xt′)}dt+2γdWt(X0′,V0′)=z′.(3)\begin{aligned} \mathrm{d}X_t &= V_t\mathrm{d}t \\ \mathrm{d}V_t &= \{-\gamma V_t-\nabla U(X_t)\}\mathrm{d}t+\sqrt{2\gamma}\mathrm{d}W_t, \\ (X_0,V_0) &= z \end{aligned} \qquad \begin{aligned} \mathrm{d}X'_t &= V'_t\mathrm{d}t \\ \mathrm{d}V'_t &= \{-\gamma V'_t-\nabla U(X'_t)\}\mathrm{d}t+\sqrt{2\gamma}\mathrm{d}W_t \\ (X'_0,V'_0) &= z'. \tag*{(3)} \end{aligned}

Moreover, (2) is equivalent to a log-Harnack inequality, that is, for any sufficiently well-behaved f:R2d→(0,∞)f:\mathbb{R}^{2d}\to(0,\infty)

Pt(log⁡f)(z)≤log⁡Ptf(z′)+Ce−χt∣z−z′∣(R∨∣z−z′∣).(4)P_t(\log f)(z)\le\log P_tf(z')+Ce^{-\chi t}|z-z'|(R\vee|z-z'|). \tag*{(4)}

We refer to section 4.6 for a short proof via Donsker-Varadhan variational formula.

A main obstruction in proving (2) is the fact that the initial distributions δz,δz′\delta_z,\delta_{z'} are mutually singular whenever z≠z′z\ne z' prohibiting many arguments based on a direct application of Girsanov’s theorem to compare the two [SDEs] in (3).

As a remedy, one may consider coupling techniques using a third auxiliary process, a technique referred to as shifted Girsanov, cf. [13, 1, 2, 3]. More specifically, we consider an auxiliary bridge process (Ztaux)t(Z_t^{\mathrm{aux}})_t interpolating between z′z' and ZTZ_T, i.e., Z0aux=z′Z_0^{\mathrm{aux}}=z' and ZTaux=ZTZ_T^{\mathrm{aux}}=Z_T at a prescribed time T>0T>0. This allows us to write, KL⁡(ZT∣ZT′)=KL⁡(ZTaux∣ZT′aux)\operatorname{KL}(Z_T\mid Z'_T)=\operatorname{KL}(Z_T^{\mathrm{aux}}\mid Z_T^{\prime\mathrm{aux}}). By designing the process (Ztaux)t(Z_t^{\mathrm{aux}})_t appropriately, we can ensure that the path measures of (Ztaux)t(Z_t^{\mathrm{aux}})_t and (Zt′)t(Z'_t)_t are, indeed, mutually absolutely continuous hence allowing to estimate KL⁡(ZT′∣ZT)\operatorname{KL}(Z_T'\mid Z_T) using Girsanov.

Related Work and Contributions

Underdamped Langevin Due to its accelerating properties there is a vast literature on underdamped Langevin dynamics. The main technical difficulty of underdamped Langevin is the degeneracy of the Brownian motion, which led to Villani’s development of hypocoercivity [Villani, 2006]. In the machine learning community the research on underdamped Langevin was largely incited by [4]. For works considering the convex setting and synchronous coupling, see [6, 4, 12]. In the non-convex setting, reflection coupling with carefully designed distance functions has been fruitfully used to derive contraction results in Wasserstein distances [8, 7] (Schuh 2024). Furthermore, there is a vast literature on different discretizations of underdamped Langevin dynamics. Most notably, its structure is amenable to splitting methods where individual components of the drift are integrated alternatingly [11, 12]. A discretization scheme that has been particularly successful is the randomized midpoint scheme [2] (Shen and Lee 2019). Of all the cited works, the most closely related is undoubtedly [2]. There the authors employ the shifted Girsanov argument that is also used in the present article and leads to improved scaling in the condition number κ\kappa.

Shifted Girsanov and reverse-Transport- and Harnack Inequalities Dimension-free Harnack inequalities were first shown for heat semigroups on manifolds via semigroup methods in [Wang, 1997] and afterward via coupling techniques in [3]. By now there is an extensive body of literature regarding Harnack inequalities and their applications and we refer to [Wang, 2010a, 2013, 2010b, Wang and Yuan, 2011] for some overviews. In fact, [3] is the work pioneering the idea of using a third auxiliary process together with Girsanov’s theorem to derive Harnack type inequalities. More recently, the works of [1] and [2] used a shifted Girsanov arguments to establish reverse transport inequalities for the overdamped and underdamped Langevin dynamics. In [13] the techniques from [2] are applied to the overdamped Langevin diffusion in the non-logconcave setting. Finally, [9] derived reverse transportation inequalities for the underdamped Langevin diffusion via a chain of implications (cf. [Wang, 2013, Section 1.3]). However, in contrast to our results, their bounds do not converge to 0 as T→∞T \to\infty.

Contributions As mentioned above, our work can be seen as an extension of some of the results in [2, 13]. In particular, [2] provide a corresponding convergence result only for convex potentials2 and [13] for non-convex potentials but for the overdamped Langevin dynamics. Conversely the results in the present article cover the non-convex and underdamped setting. It should be mentioned, however, that [2, 13] provide additional results in other contexts. In particular, in [2] the authors also consider the discrete setting and in [13] results are also obtained for the, more general, Rényi divergence.

Notation and Assumptions

Notation Throughout xx will be used for the position variable, vv for the velocity, and z=(x,v)z = (x,v) for the position-velocity pair and we note that several different variations of these symbols will appear (x,X,X′,Xaux(x, X, X', X^{\mathrm{aux}} etc.). The set of probability measures on Rd\mathbb{R}^d is denoted as P(Rd)\mathcal{P}(\mathbb{R}^d). We write μ≪ν\mu\ll\nu when μ\mu is absolutely continuous w.r.t. ν\nu, and with dμdν\frac{\mathrm{d}\mu}{\mathrm{d}\nu} for the Radon-Nikodym derivative of μ\mu w.r.t. ν\nu. We define the Kullback-Leibler divergence as

KL:P(Rd)×P(Rd)→[0,∞]\mathrm{KL} : \mathcal{P}(\mathbb{R}^d) \times\mathcal{P}(\mathbb{R}^d) \to[0,\infty]
(μ,ν)↦KL(μ∣ν)={∫log⁡(dμdν)dμdν dνif μ≪ν∞else.(5)(\mu,\nu) \mapsto\mathrm{KL}(\mu|\nu) = \begin{cases} \int\log\left(\frac{\mathrm{d}\mu}{\mathrm{d}\nu}\right)\frac{\mathrm{d}\mu}{\mathrm{d}\nu}\,\mathrm{d}\nu& \text{if } \mu\ll\nu\\ \infty& \text{else.} \end{cases} \tag*{(5)}

We will denote for simplicity for two random variables X,YX,Y

KL(X∣Y):=KL(law⁡(X)∣law⁡(Y)).(6)\mathrm{KL}(X|Y) := \mathrm{KL}(\operatorname{law}(X)|\operatorname{law}(Y)). \tag*{(6)}

The symbol O~\widetilde{O} is used to denote complexity results omitting logarithmic factors. Our main assumptions about the potential UU are LL-smoothness and strong convexity outside a ball:

Assumption 3.1. The potential U∈C2(Rd)U \in C^2(\mathbb{R}^d) and there exist L,m,R>0L,m,R > 0 such that we have

LI⪰∇2U(x)⪰−LI,x∈Rd(7)LI \succeq\nabla^2 U(x) \succeq-LI,\qquad x \in\mathbb{R}^d \tag*{(7)}
∇2U(x)⪰mI,∣x∣≥R.(8)\nabla^2 U(x) \succeq mI,\qquad|x| \ge R. \tag*{(8)}

Proof of the main result

Set-up

To prove the above stated reverse transportation inequality we rely on the shifted composition principle introduced in [3] and recently rediscovered in [1, 2]. We explain and motivate the approach in the following.

Our goal is to bound the KL divergence between the law of two solutions of (ULD) initialized at z0,z0′∈R2dz_0,z'_0 \in\mathbb{R}^{2d}, respectively

{dXt=Vtdt,dVt={−γVt−∇U(Xt)}dt+2γdWt,Z0=z0{dXt′=Vt′dt,dVt′={−γVt′−∇U(Xt′)}dt+2γdWt,Z0′=z0′.(9)\begin{cases} \mathrm{d}X_t = V_t\mathrm{d}t,\\ \mathrm{d}V_t = \{-\gamma V_t-\nabla U(X_t)\}\mathrm{d}t+\sqrt{2\gamma}\mathrm{d}W_t,\\ Z_0=z_0 \end{cases} \qquad \begin{cases} \mathrm{d}X'_t = V'_t\mathrm{d}t,\\ \mathrm{d}V'_t = \{-\gamma V'_t-\nabla U(X'_t)\}\mathrm{d}t+\sqrt{2\gamma}\mathrm{d}W_t,\\ Z'_0=z'_0. \tag*{(9)} \end{cases}

Recall, however, that a naïve application of Girsanov’s theorem to these processes as-is will not lead to meaningful bounds as the different initializations render the respective path measures mutually singular. As a remedy, we introduce a deliberately designed interpolating auxiliary process ZauxZ^{\mathrm{aux}} which satisfies on one hand Z0aux=z0′Z^{\mathrm{aux}}_0=z'_0 so that we may apply Girsanov’s theorem to compare ZauxZ^{\mathrm{aux}} and Z′Z'. On the other hand, we design ZauxZ^{\mathrm{aux}} such that at a finite time T>0T>0 it holds law⁡(ZTaux)=law⁡(ZT)\operatorname{law}(Z^{\mathrm{aux}}_T)=\operatorname{law}(Z_T) which implies KL⁡(ZT∣ZT′)=KL⁡(ZTaux∣ZT′)\operatorname{KL}(Z_T\mathbin{|}Z'_T)=\operatorname{KL}(Z^{\mathrm{aux}}_T\mathbin{|}Z'_T). In other words, the process ZauxZ^{\mathrm{aux}} constitutes a bridge from Z0′Z'_0 to ZTZ_T. Let us now make this construction more concrete:

For z0,z0′∈R2dz_0,z'_0\in\mathbb{R}^{2d} we define the coupled SDE,

{dXt=Vtdt,dVt={−γVt−∇U(Xt)}dt+2γdWt,dXtaux=Vtauxdt,dVtaux={−γVtaux−∇U(Xtaux)+ηtx(Xt−Xtaux)+ηtv(Vt−Vtaux)}dt+2γdW‾t,Z0=z0,Z0aux=z0′\begin{cases} \mathrm{d}X_t = V_t\mathrm{d}t,\\ \mathrm{d}V_t = \{-\gamma V_t-\nabla U(X_t)\}\mathrm{d}t+\sqrt{2\gamma}\mathrm{d}W_t,\\ \mathrm{d}X^{\mathrm{aux}}_t = V^{\mathrm{aux}}_t\mathrm{d}t,\\ \mathrm{d}V^{\mathrm{aux}}_t = \left\{-\gamma V^{\mathrm{aux}}_t-\nabla U(X^{\mathrm{aux}}_t)+\eta^x_t(X_t-X^{\mathrm{aux}}_t)+\eta^v_t(V_t-V^{\mathrm{aux}}_t)\right\}\mathrm{d}t+\sqrt{2\gamma}\mathrm{d}\overline{W}_t,\\ Z_0=z_0,\\ Z^{\mathrm{aux}}_0=z'_0 \end{cases}

where (Wt)t(W_t)_t and (W‾t)t(\overline{W}_t)_t are two Brownian motions coupled in a deliberate way, which will be explained below. Crucially, the shift parameters ηtv,ηtx\eta^v_t,\eta^x_t will be chosen to satisfy ηtv,ηtx→∞\eta^v_t,\eta^x_t\to\infty as t→Tt\to T ensuring that ZTaux=ZTZ^{\mathrm{aux}}_T=Z_T at a finite time TT. In addition, we have a second solution of (ULD) initialized at z0′z'_0

{dXt′=Vt′dt,dVt′={−γVt′−∇U(Xt′)}dt+2γdWt′,Z0′=z0′(10)\begin{cases} \mathrm{d}X'_t=V'_t\mathrm{d}t,\\ \mathrm{d}V'_t=\{-\gamma V'_t-\nabla U(X'_t)\}\mathrm{d}t+\sqrt{2\gamma}\mathrm{d}W'_t,\\ Z'_0=z'_0 \tag*{(10)} \end{cases}

where we note that, technically, (10) is not defined on the same probability space as (AUX) (see section A for details). Applying Girsanov’s theorem to compare ZauxZ^{\mathrm{aux}} with Z′Z' then yields

KL⁡(ZT∣ZT′)=KL⁡(ZTaux∣ZT′)≤KL⁡((Ztaux)t∈[0,T]∣(Zt′)t∈[0,T])=14γ∫0TE[∣ηtx(Xt−Xtaux)+ηtv(Vt−Vtaux)∣2]dt(11)\begin{aligned} \operatorname{KL}(Z_T\mathbin{|}Z'_T)=\operatorname{KL}(Z^{\mathrm{aux}}_T\mathbin{|}Z'_T)\leq\operatorname{KL}\left((Z^{\mathrm{aux}}_t)_{t\in[0,T]}\mathbin{\big|}(Z'_t)_{t\in[0,T]}\right) \\ &=\frac{1}{4\gamma}\int_0^T\mathbb{E}\left[\left|\eta^x_t(X_t-X^{\mathrm{aux}}_t)+\eta^v_t(V_t-V^{\mathrm{aux}}_t)\right|^2\right]\mathrm{d}t \tag*{(11)} \end{aligned}

where we used the hitting condition law⁡(ZTaux)=law⁡(ZT)\operatorname{law}(Z_T^{\mathrm{aux}})=\operatorname{law}(Z_T) and the data processing inequality. To now obtain the desired bounds, (2) respectively (4), it remains to estimate the above integral.

Proof outline The proof consists of the following steps:

  1. Inspired by [13] we define a time t^∈[0,T]\hat{t}\in[0,T] at which a regime transition happens. For t≥t^t\ge\hat{t} we will choose ηtx,ηtv≫0\eta_t^x,\eta_t^v\gg0. In this regime the shifts are strong, and a synchronous coupling, that is, dWt=dW‾t\mathrm{d}W_t=\mathrm{d}\overline{W}_t in (AUX) is sufficient to show contraction and prove the hitting condition ZTaux=ZTZ_T^{\mathrm{aux}}=Z_T.

  1. Conversely, for t<t^t<\hat{t} we set ηtx=ηtv=0\eta_t^x=\eta_t^v=0. In this regime, ZtauxZ_t^{\mathrm{aux}} and ZtZ_t are two conventional ULD processes. Thus, we employ a reflection coupling for WtW_t and W‾t\overline{W}_t (cf. [8] (Schuh 2024)) which allows us to derive contraction in a deliberately designed distance without any shifts.

  1. The result follows by concatenating the bounds in the two regimes.

The remainder of the article contains the proof details.

Coupling construction

It is well-known in the literature that purely synchronous coupling is unsuitable to derive contraction results for (ULD) in the non-convex setting (cf. [8]). Thus, we rely on a combination of synchronous and reflection coupling. More precisely, building on [8] (Schuh 2024) we first introduce new coordinates. Specifically, let us denote

Pt=Xt+γ−1Vt(12)P_t=X_t+\gamma^{-1}V_t \tag*{(12)}

and analogously for the auxiliary process. Then (ULD) reads as

dXt=−γXtdt+γPtdtdPt=−γ−1∇U(Xt)dt+2γ−1dWt.(13)\begin{aligned} \mathrm{d}X_t&=-\gamma X_t\mathrm{d}t+\gamma P_t\mathrm{d}t\\ \mathrm{d}P_t&=-\gamma^{-1}\nabla U(X_t)\mathrm{d}t+\sqrt{2\gamma^{-1}}\mathrm{d}W_t. \tag*{(13)} \end{aligned}

Now we define the processes (Wt)t(W_t)_t and (W‾t)t(\overline{W}_t)_t in (AUX) using a combination of synchronous and reflection coupling as

Wt=∫0trc⁡(Zs,Zsaux)dWsrc+∫0tsc⁡(Zs,Zsaux)dWssc,(14)W_t=\int_0^t\operatorname{rc}(Z_s,Z_s^{\mathrm{aux}})\mathrm{d}W_s^{\mathrm{rc}}+\int_0^t\operatorname{sc}(Z_s,Z_s^{\mathrm{aux}})\mathrm{d}W_s^{\mathrm{sc}}, \tag*{(14)}
W‾t={∫0trc⁡(Zs,Zsaux)RsdWsrc+∫0tsc⁡(Zs,Zsaux)dWssc,for t<t^,Wt−Wt^+W‾t^,for t≥t^.(15)\overline{W}_t= \begin{cases} \int_0^t\operatorname{rc}(Z_s,Z_s^{\mathrm{aux}})R_s\mathrm{d}W_s^{\mathrm{rc}}+\int_0^t\operatorname{sc}(Z_s,Z_s^{\mathrm{aux}})\mathrm{d}W_s^{\mathrm{sc}}, & \text{for }t<\hat{t},\\ W_t-W_{\hat{t}}+\overline{W}_{\hat{t}}, & \text{for }t\ge\hat{t}. \end{cases} \tag*{(15)}

with (Wtrc)t(W_t^{\mathrm{rc}})_t, (Wtsc)t(W_t^{\mathrm{sc}})_t two independent Brownian motions and the reflector Rs=(I−2esesT)R_s=(I-2e_se_s^{\mathsf T}) with et=Qt/∣Qt∣e_t=Q_t/\lvert Q_t\rvert and Qt=Pt−PtauxQ_t=P_t-P_t^{\mathrm{aux}}. The two smooth cut-off functions sc,rc:R2d×R2d→[0,1]\mathrm{sc},\mathrm{rc}:\mathbb{R}^{2d}\times\mathbb{R}^{2d}\to[0,1] satisfy sc2+rc2=1\mathrm{sc}^2+\mathrm{rc}^2=1. One may then check that WtW_t and W‾t\overline{W}_t are, indeed, Brownian motions by Lévy’s characterization [10 Theorem 3.16]. While for t>t^t>\hat{t} we have standard synchronous coupling, the construction for t<t^t<\hat{t} is exactly the one from [Schuh, 2024] and we refer to this work for the details and limit ourselves to a high-level explanation here.

Essentially, the coupling should be synchronous whenever Qt=0Q_t=0 or Zt−ZtauxZ_t-Z_t^{\mathrm{aux}} is large. The reason therefore is that in the first case, Pt=PtauxP_t=P_t^{\mathrm{aux}} and thus the first equation in (13) yields a contraction also in the XX-component. In the second case, on the other hand, the convexity outside a ball of UU provides contraction with synchronous coupling. Thus, the reflection coupling is active only in the intermediate regime, where Zt−ZtauxZ_t-Z_t^{\mathrm{aux}} is small but not zero.

Contraction for t>t^t > \hat{t}

We begin by investigating the regime t>t^t > \hat{t}. In this setting, we define a slightly different twisted coordinate system than (12). Specifically, we set

\begin{bmatrix} X_t \\ \widetilde{P}_t \end{bmatrix} := \begin{bmatrix} X_t \\ X_t + \frac{2}{\gamma_t}V_t \end{bmatrix} = A_t \begin{bmatrix} X_t \\ V_t \end{bmatrix}, \qquad A_t := \begin{bmatrix} 1 & 0 \\ 1 & \frac{2}{\gamma_t} \end{bmatrix}, \qquad \gamma_t = \gamma+ \eta_t^v. \tag*{(16)}\end{bmatrix}

The definition of the time-dependent friction parameter γt=γ+ηtv\gamma_t = \gamma+ \eta_t^v follows naturally from collecting the VtauxV_t^{\mathrm{aux}}-terms in (AUX). Note that P~\widetilde{P} differs from PP by the factor 2 in the bottom right entry of AtA_t. Moreover, we introduce the notation

Dt:=[Xt−XtauxVt−Vtaux],Tt:=AtDt(17)D_t := \begin{bmatrix} X_t - X_t^{\mathrm{aux}} \\ V_t - V_t^{\mathrm{aux}} \end{bmatrix}, \qquad T_t := A_tD_t \tag*{(17)}

for the difference processes in regular and twisted coordinates. Next, we fix the precise definition of the shift parameters.

Definition 4.1 (Shift parameters). We define the shift parameters for c0>0c_0 > 0, as

ηtv=1t≥t^c0ωeω(T−t)−1,ηtx=γtηtv2(18)\eta_t^v = \mathbf{1}_{t \ge\hat{t}}\frac{c_0\omega}{e^{\omega(T-t)} - 1}, \qquad \eta_t^x = \frac{\gamma_t\eta_t^v}{2} \tag*{(18)}

where

ω={−L3γif γ≥32L (high friction)−L3if γ<32L (low friction)(19)\omega= \begin{cases} -\frac{L}{3\gamma} & \text{if } \gamma\ge\sqrt{32L}\ \text{(high friction)} \\ -\frac{\sqrt{L}}{3} & \text{if } \gamma< \sqrt{32L}\ \text{(low friction)} \end{cases} \tag*{(19)}

Note that, indeed, ηtv,ηtx→∞\eta_t^v,\eta_t^x \to\infty as t→Tt \to T. The following result is shown in [2]. We include the proof for completeness.

Lemma 4.1. There exist absolute constants c′,c0>0c',c_0 > 0 such that for any s,t∈[t^,T]s,t \in[\hat{t},T], s<ts < t it holds

E[∣Tt∣2]≤(1−e−ω(T−t)1−e−ω(T−s))c′c0E[∣Ts∣2].(20)\mathbb{E}[|T_t|^2] \le\left(\frac{1-e^{-\omega(T-t)}}{1-e^{-\omega(T-s)}}\right)^{c'c_0}\mathbb{E}[|T_s|^2]. \tag*{(20)}

Moreover, choosing c0>0c_0 > 0 sufficiently large it holds E[∣Zt−Ztaux∣2]→0\mathbb{E}[|Z_t-Z_t^{\mathrm{aux}}|^2] \to0 as t→Tt \to T.

Proof. Recall that for t>t^t > \hat{t} we have a synchronous coupling (cf. (14)) and it holds

d∣Tt∣2=2⟨Tt,dTt⟩(21)\mathrm{d}|T_t|^2 = 2\langle T_t,\mathrm{d}T_t\rangle \tag*{(21)}

where, since Tt=AtDtT_t = A_tD_t,

dTt=A˙tDtdt+AtdDt.(22)\mathrm{d}T_t = \dot{A}_tD_t\mathrm{d}t + A_t\mathrm{d}D_t. \tag*{(22)}

For the second term we find,

dDt=[0I−ηtxI−Ht−γtI]Dtdt=MtDtdt(23)\mathrm{d}D_t = \begin{bmatrix} 0 & I \\ -\eta_t^x I - H_t & -\gamma_t I \end{bmatrix} D_t\mathrm{d}t = M_tD_t\mathrm{d}t \tag*{(23)}

where we used that by the fundamental theorem of calculus

∇xU(Xt)−∇xU(Xtaux)=∫01∇2U(Xtaux+λ(Xt−Xtaux))dλ⏟=:Ht(Xt−Xtaux).(24)\nabla_xU(X_t)-\nabla_xU(X_t^{\mathrm{aux}}) = \underbrace{\int_0^1 \nabla^2U\left(X_t^{\mathrm{aux}}+\lambda(X_t-X_t^{\mathrm{aux}})\right)\mathrm{d}\lambda}_{=:H_t} \left(X_t-X_t^{\mathrm{aux}}\right). \tag*{(24)}

Inserting in (21) yields

d∣Tt∣2=2⟨Tt,(A˙t+AtMt)At−1Tt⟩dt=−2⟨Tt,QtTt⟩dt\mathrm{d}|T_t|^2=2\langle T_t,(\dot{A}_t+A_tM_t)A_t^{-1}T_t\rangle\mathrm{d}t=-2\langle T_t,Q_tT_t\rangle\mathrm{d}t

where we introduce the notation Qt=−sym⁡((A˙t+AtMt)At−1)Q_t=-\operatorname{sym}((\dot{A}_t+A_tM_t)A_t^{-1}) with the symmetrization sym⁡(M):=(M+MT)/2\operatorname{sym}(M):=(M+M^{\mathsf{T}})/2 for any quadratic matrix MM. By [2 Lemma 3.13] for c0≥24c_0\ge24, it holds for the smallest eigenvalue of QQ

λmin⁡(Qt)≥(ηtv)248(25)\lambda_{\min}(Q_t)\ge\frac{(\eta_t^v)^2}{48} \tag*{(25)}

where we note that this bound is proven in [2 Lemma 3.13] in both the high and low friction settings. Inserting the lower bound and the definition of ηtv\eta_t^v yields with c′=148c'=\frac{1}{48}

d∣Tt∣2≤−c′c0ωeω(T−t)−1∣Tt∣2dt.(26)\mathrm{d}|T_t|^2\le-\frac{c'c_0\omega}{e^{\omega(T-t)}-1}|T_t|^2\mathrm{d}t. \tag*{(26)}

By Grönwall it follows for any t^≤s<t\hat{t}\le s<t

E[∣Tt∣2]≤(1−e−ω(T−t)1−e−ω(T−s))c′c0E[∣Ts∣2].(27)\mathbb{E}[|T_t|^2]\le\left(\frac{1-e^{-\omega(T-t)}}{1-e^{-\omega(T-s)}}\right)^{c'c_0}\mathbb{E}[|T_s|^2]. \tag*{(27)}

To proof the hitting condition, note that since E[∣Xt−Xtaux∣2]≤E[∣Tt∣2]→0\mathbb{E}[|X_t-X_t^{\mathrm{aux}}|^2]\le\mathbb{E}[|T_t|^2]\to0 as t→Tt\to T we directly obtain convergence of the xx-component. For the vv-component we have for c′c0>2c'c_0>2

E[∣Vt−Vtaux∣2]≲γt2E[∣Tt∣2]≲1(eω(T−t)−1)2(1−e−ω(T−t)1−e−ω(T−t^))c′c0E[∣Tt^∣2]\mathbb{E}[|V_t-V_t^{\mathrm{aux}}|^2]\lesssim\gamma_t^2\mathbb{E}[|T_t|^2]\lesssim\frac{1}{(e^{\omega(T-t)}-1)^2}\left(\frac{1-e^{-\omega(T-t)}}{1-e^{-\omega(T-\hat{t})}}\right)^{c'c_0}\mathbb{E}[|T_{\hat{t}}|^2]
≲1e2ω(T−t)(1−e−ω(T−t))c′c0−2(1−e−ω(T−t^))c′c0E[∣Tt^∣2]→0(28)\lesssim\frac{1}{e^{2\omega(T-t)}}\frac{(1-e^{-\omega(T-t)})^{c'c_0-2}}{(1-e^{-\omega(T-\hat{t})})^{c'c_0}}\mathbb{E}[|T_{\hat{t}}|^2]\to0 \tag*{(28)}

as t→Tt\to T.

Lemma 4.2. It holds

KL⁡(ZT∥ZT′)≲(∣ω∣3(e−4ω(T−t^)−1)(1−e−ω(T−t^))4+∣ω∣γ2(e−2ω(T−t^)−1)(1−e−ω(T−t^))2)E[∣Tt^∣2].(29)\operatorname{KL}(Z_T\Vert Z'_T)\lesssim\left(\frac{|\omega|^3(e^{-4\omega(T-\hat{t})}-1)}{(1-e^{-\omega(T-\hat{t})})^4}+\frac{|\omega|\gamma^2(e^{-2\omega(T-\hat{t})}-1)}{(1-e^{-\omega(T-\hat{t})})^2}\right)\mathbb{E}[|T_{\hat{t}}|^2]. \tag*{(29)}

Proof. By lemma 4.1 Zt−Ztaux→0Z_t-Z_t^{\mathrm{aux}}\to0 as t→Tt\to T in L2(P)L^2(\mathbb{P}). By L2(P)L^2(\mathbb{P})-continuity of (Zt)t(Z_t)_t it follows

E[∣ZT−Ztaux∣2]1/2≤E[∣ZT−Zt∣2]1/2+E[∣Zt−Ztaux∣2]1/2→0.(30)\mathbb{E}[|Z_T-Z_t^{\mathrm{aux}}|^2]^{1/2}\le\mathbb{E}[|Z_T-Z_t|^2]^{1/2}+\mathbb{E}[|Z_t-Z_t^{\mathrm{aux}}|^2]^{1/2}\to0. \tag*{(30)}

Thus, it holds Ztaux→ZTZ_t^{\mathrm{aux}}\to Z_T and Zt′→ZT′Z'_t\to Z'_T both in L2L^2. Indeed, L2L^2 convergence implies weak convergence of the laws and therefore, by (joint) weak lower semicontinuity of the KL we do, in fact, have

KL⁡(ZT∥ZT′)≤lim inf⁡t→TKL⁡(Ztaux∥Zt′)≤14γ∫0TE[∣ηtx(Xt−Xtaux)+ηtv(Vt−Vtaux)∣2]dt(31)\operatorname{KL}(Z_T\Vert Z'_T)\le\liminf_{t\to T}\operatorname{KL}(Z_t^{\mathrm{aux}}\Vert Z'_t)\le\frac{1}{4\gamma}\int_0^T\mathbb{E}\left[|\eta_t^x(X_t-X_t^{\mathrm{aux}})+\eta_t^v(V_t-V_t^{\mathrm{aux}})|^2\right]\mathrm{d}t \tag*{(31)}

where we used Girsanov’s theorem in the last inequality. Using lemma 4.1 we obtain

KL⁡(ZT∥ZT^)≤14γ∫t^T(ηtν)2γt2E[∣Tt∣2] dt(32)\operatorname{KL}(Z_T\|Z_{\hat{T}}) \leq\frac{1}{4\gamma}\int_{\hat{t}}^T (\eta_t^\nu)^2\gamma_t^2\mathbb{E}[|T_t|^2]\,\mathrm{d}t \tag*{(32)}
≤14γ∫t^T(ηtν)2γt2(1−e−ω(T−t)1−e−ω(T−t^))c′c0dt E[∣Tt^∣2](33)\leq\frac{1}{4\gamma}\int_{\hat{t}}^T (\eta_t^\nu)^2\gamma_t^2\left(\frac{1-e^{-\omega(T-t)}}{1-e^{-\omega(T-\hat{t})}}\right)^{c'c_0}\mathrm{d}t\,\mathbb{E}[|T_{\hat{t}}|^2] \tag*{(33)}

The integral can be estimated as follows,

∫t^T(ηtν)2γt2(1−e−ω(T−t)1−e−ω(T−t^))c′c0dt(34)\int_{\hat{t}}^T (\eta_t^\nu)^2\gamma_t^2\left(\frac{1-e^{-\omega(T-t)}}{1-e^{-\omega(T-\hat{t})}}\right)^{c'c_0}\mathrm{d}t \tag*{(34)}
≲∫t^Tω4(eω(T−s)−1)4(1−e−ω(T−s)1−e−ω(T−t^))c′c0ds+∫t^Tω2γ2(eω(T−s)−1)2(1−e−ω(T−s)1−e−ω(T−t^))c′c0ds(35)\lesssim\int_{\hat{t}}^T \frac{\omega^4}{(e^{\omega(T-s)}-1)^4}\left(\frac{1-e^{-\omega(T-s)}}{1-e^{-\omega(T-\hat{t})}}\right)^{c'c_0}\mathrm{d}s+\int_{\hat{t}}^T \frac{\omega^2\gamma^2}{(e^{\omega(T-s)}-1)^2}\left(\frac{1-e^{-\omega(T-s)}}{1-e^{-\omega(T-\hat{t})}}\right)^{c'c_0}\mathrm{d}s \tag*{(35)}

noting that 1−e−ω(T−s)1−e−ω(T−t^)∈[0,1]\frac{1-e^{-\omega(T-s)}}{1-e^{-\omega(T-\hat{t})}}\in[0,1] and assuming c0>0c_0>0 large enough so that c′c0≥4c'c_0\geq4

≲∫t^Tω4(eω(T−s)−1)4(1−e−ω(T−s)1−e−ω(T−t^))4ds+∫t^Tω2γ2(eω(T−s)−1)2(1−e−ω(T−s)1−e−ω(T−t^))2ds(36)\lesssim\int_{\hat{t}}^T \frac{\omega^4}{(e^{\omega(T-s)}-1)^4}\left(\frac{1-e^{-\omega(T-s)}}{1-e^{-\omega(T-\hat{t})}}\right)^4\mathrm{d}s+\int_{\hat{t}}^T \frac{\omega^2\gamma^2}{(e^{\omega(T-s)}-1)^2}\left(\frac{1-e^{-\omega(T-s)}}{1-e^{-\omega(T-\hat{t})}}\right)^2\mathrm{d}s \tag*{(36)}

using that 1−1/uu−1=1u\frac{1-1/u}{u-1}=\frac{1}{u}

≲∫t^Tω4e4ω(T−s)1(1−e−ω(T−t^))4 ds+∫t^Tω2γ2e2ω(T−s)1(1−e−ω(T−t^))2 ds(37)\lesssim\int_{\hat{t}}^T \frac{\omega^4}{e^{4\omega(T-s)}}\frac{1}{(1-e^{-\omega(T-\hat{t})})^4}\,\mathrm{d}s+\int_{\hat{t}}^T \frac{\omega^2\gamma^2}{e^{2\omega(T-s)}}\frac{1}{(1-e^{-\omega(T-\hat{t})})^2}\,\mathrm{d}s \tag*{(37)}
≲∣ω∣3(e−4ω(T−t^)−1)(1−e−ω(T−t^))4+∣ω∣γ2(e−2ω(T−t^)−1)(1−e−ω(T−t^))2(38)\lesssim\frac{|\omega|^3(e^{-4\omega(T-\hat{t})}-1)}{(1-e^{-\omega(T-\hat{t})})^4}+\frac{|\omega|\gamma^2(e^{-2\omega(T-\hat{t})}-1)}{(1-e^{-\omega(T-\hat{t})})^2} \tag*{(38)}

yielding the desired result.

Remark 4.1. Note that, if we were to set t^=0\hat{t}=0 the bound in (29) would not tend to zero as T→∞T\to\infty since ω<0\omega<0. The exponential decay therefore has to be established in [0,t^][0,\hat{t}].

Remark 4.2. Note that the above arguments – especially the application of Girsanov to bound the KL – rely on strong existence and uniqueness of solutions of the occurring SDEs which is not trivially given due to the singular coefficients ηsν,ηsx\eta_s^\nu,\eta_s^x. However, for any t<Tt<T, strong existence and uniqueness on [0,t][0,t] follow from standard results due to Lipschitz-continuity. While ηsν,ηsx\eta_s^\nu,\eta_s^x exhibit a discontinuity at s=t^s=\hat{t}, we have strong existence and uniqueness, on [0,t^)[0,\hat{t}) as well as [t^,t][\hat{t},t] leading to a unique solution on [0,t][0,t] by concatenation. This is sufficient for the limit argument in (31).

Contraction for t<t^t<\hat{t}

In view of (29) it remains to bound E[∣Tt^∣2]\mathbb{E}[|T_{\hat{t}}|^2]. Given that ηtν=ηtx=0\eta_t^\nu=\eta_t^x=0 for t∈[0,t^)t\in[0,\hat{t}) this means we have to obtain a contraction result for regular (ULD). Readily available results, however, do not directly yield satisfying rates. Indeed, we can identify three groups of contraction results:

  1. Strongly convex: For instance [4, 6] provide contraction in the L2L^2-norm under the assumption of a strongly logconcave target using synchronous coupling. However, since we consider the non-logconcave setting, we cannot apply these results.

  1. Non-logconcave: In L2L^2 exponential decay in [8] is established using reflection and synchronous coupling even in the non-logconcave case. However, the convergence deteriorates exponentially in the dimension.

  1. Lastly, using similar techniques as in [8], in [Schuh, 2024] dimension-free exponential decay was proven in the L1L^1-norm.

Our goal is to apply the above mentioned L1L^1-convergence result from [Schuh, 2024] which we state explicitly in the following.

Lemma 4.3 (Adapted from Theorem 5 and Remark 3 in [Schuh, 2024]). Let assumption 3.1 hold. Let further γ2≥2(L+m)2m\gamma^2 \ge\frac{2(L+m)^2}{m} then it holds for some dimension free constants χ,C>0\chi,C>0 and all 0≤s≤t<t^0\le s\le t<\hat{t}

EP[∣Tt∣]≤Ce−χ(t−s)EP[∣Ts∣].(39)\mathbb{E}_{\mathbb{P}}[|T_t|]\le Ce^{-\chi(t-s)}\mathbb{E}_{\mathbb{P}}[|T_s|]. \tag*{(39)}

In order to apply lemma 4.3 we have to go from the L2L^2 to the L1L^1 norm. By Hölder’s inequality we have

E[∣Tt^∣2]≤E[∣Tt^∣]∥Tt^∥L∞(P).(40)\mathbb{E}[|T_{\hat{t}}|^2]\le\mathbb{E}[|T_{\hat{t}}|]\|T_{\hat{t}}\|_{L^\infty(\mathbb{P})}. \tag*{(40)}

Thus, to conclude, we are left to derive an a.s. uniform bound of Tt^T_{\hat{t}}.

Lemma 4.4. Let assumption 3.1 hold and assume γ≥2L\gamma\ge\sqrt{2L}. Then, for t∈[0,t^]t\in[0,\hat{t}] a.s.

∣Tt∣≲(1∨γ−1)(R∨∣z0−z0′∣).(41)|T_t|\lesssim(1\vee\gamma^{-1})(\mathcal{R}\vee|z_0-z'_0|). \tag*{(41)}

where

R≍max⁡{LmR,γ2Lmax⁡(1,L/mγ)(m−1/2+(L/m)3/2R)}.(42)\mathcal{R}\asymp\max\left\{\frac{L}{m}R,\frac{\gamma^2}{L}\max\left(1,\frac{\sqrt{L/m}}{\gamma}\right)\left(m^{-1/2}+(L/m)^{3/2}R\right)\right\}. \tag*{(42)}

Proof. Carefully tracking constants in [Schuh, 2024], one can check that rc=0\mathrm{rc}=0 if

∣Tt∣≳(γ2/L)max⁡(1,L/mγ)(m−1/2+(L/m)3/2R).(43)|T_t|\gtrsim(\gamma^2/L)\max\left(1,\frac{\sqrt{L/m}}{\gamma}\right)\left(m^{-1/2}+(L/m)^{3/2}R\right). \tag*{(43)}

By Itô’s lemma for such TtT_t due to the purely synchronous coupling

dTt=[−γ2(Pˉt−Pˉtaux)+γ2(Xt−Xtaux)−2γ(∇U(Xt)−∇U(Xtaux))]dt(44)\mathrm{d}T_t=\left[-\frac{\gamma}{2}(\bar{P}_t-\bar{P}^{\mathrm{aux}}_t)+\frac{\gamma}{2}(X_t-X^{\mathrm{aux}}_t)-\frac{2}{\gamma}\left(\nabla U(X_t)-\nabla U(X^{\mathrm{aux}}_t)\right)\right]\mathrm{d}t \tag*{(44)}

and, thus,

d∣Tt∣2=2Tt⊤[−γ2γ2γ2−γ2]Ttdt−4γ(Pˉt−Pˉtaux)⊤Ht(Xt−Xtaux)dt(45)\mathrm{d}|T_t|^2=2T_t^\top\begin{bmatrix}-\frac{\gamma}{2}&\frac{\gamma}{2}\\\frac{\gamma}{2}&-\frac{\gamma}{2}\end{bmatrix}T_t\mathrm{d}t-\frac{4}{\gamma}(\bar{P}_t-\bar{P}^{\mathrm{aux}}_t)^\top\mathcal{H}_t(X_t-X^{\mathrm{aux}}_t)\mathrm{d}t \tag*{(45)}

where

Ht:=∫01∇2U(Xtaux+s(Xt−Xtaux))ds.(46)\mathcal{H}_t:=\int_0^1\nabla^2U\left(X^{\mathrm{aux}}_t+s(X_t-X^{\mathrm{aux}}_t)\right)\mathrm{d}s. \tag*{(46)}

Now we make a case distinction. Let R~>0\widetilde{R}>0 be such that for ∣Xt−Xtaux∣>R~|X_t-X^{\mathrm{aux}}_t|>\widetilde{R} it holds Ht⪰m2I\mathcal{H}_t\succeq\frac{m}{2}I. This can be ensured for R~≍LmR\widetilde{R}\asymp\frac{L}{m}R, cf. lemma B.1. Then, for ∣Xt−Xtaux∣≤R~|X_t-X^{\mathrm{aux}}_t|\le\widetilde{R} the second row of (44) yields

d∣Pˉt−Pˉtaux∣2≤∣Pˉt−Pˉtaux∣(−γ∣Pˉt−Pˉtaux∣+(4Lγ+γ)R~).(47)\mathrm{d}|\bar{P}_t-\bar{P}^{\mathrm{aux}}_t|^2\le|\bar{P}_t-\bar{P}^{\mathrm{aux}}_t|\left(-\gamma|\bar{P}_t-\bar{P}^{\mathrm{aux}}_t|+\left(\frac{4L}{\gamma}+\gamma\right)\widetilde{R}\right). \tag*{(47)}

and in particular, t↦∣Pˉt−Pˉtaux∣2t \mapsto|\bar{P}_t-\bar{P}^{\mathrm{aux}}_t|^2 decreases whenever ∣Pˉt−Pˉtaux∣≥(4Lγ2+1)R~|\bar{P}_t-\bar{P}^{\mathrm{aux}}_t| \ge\left(\frac{4L}{\gamma^2}+1\right)\widetilde{R}. On the other hand-side, for ∣Xt−Xtaux∣≥R~|X_t-X^{\mathrm{aux}}_t| \ge\widetilde{R}, (45) gives

d∣Tt∣2≤2Tt⊤[−γ2γ2γ2−γ2−1γHt]Ttdt=2Tt⊤[−γ2γ2∗−γ2−1γHt]Ttdt.(48)\begin{aligned} \mathrm{d}|T_t|^2 \leq2T_t^\top \begin{bmatrix} -\frac{\gamma}{2} & \frac{\gamma}{2} \\ \frac{\gamma}{2} & -\frac{\gamma}{2}-\frac{1}{\gamma}\mathcal{H}_t \end{bmatrix} T_t\mathrm{d}t \\ &=2T_t^\top \begin{bmatrix} -\frac{\gamma}{2} & \frac{\gamma}{2} \\ * & -\frac{\gamma}{2}-\frac{1}{\gamma}\mathcal{H}_t \end{bmatrix} T_t\mathrm{d}t. \tag*{(48)} \end{aligned}

where we symmetrized the matrix for the last equality. The eigenvalues of the above matrix are the solutions λ\lambda of the equation

0=(γ2+λ)2−(γ2−μγ)2(49)0=\left(\frac{\gamma}{2}+\lambda\right)^2-\left(\frac{\gamma}{2}-\frac{\mu}{\gamma}\right)^2 \tag*{(49)}

with μ\mu an eigenvalue of Ht\mathcal{H}_t. An explicit computation reveals that if γ2≥2μ\gamma^2\geq2\mu the largest eigenvalue is −μγ-\frac{\mu}{\gamma} and since μ>m2\mu>\frac{m}{2} it holds

d∣Tt∣2≤−mγ∣Tt∣2dt.(50)\mathrm{d}|T_t|^2\leq-\frac{m}{\gamma}|T_t|^2\mathrm{d}t. \tag*{(50)}

In total, since γ≥2L\gamma\geq\sqrt{2L} it follows that t↦∣Tt∣2t\mapsto|T_t|^2 is decreasing when

∣Tt∣2≳R~2≳L2m2R2.(51)|T_t|^2\gtrsim\widetilde{R}^2\gtrsim\frac{L^2}{m^2}R^2. \tag*{(51)}

Combining both cases, noting a.s. path-continuity, accounting for the initialization distance and equivalence of the twisted norm to the Euclidean norm yields the desired result.

Putting everything together

Theorem 4.1. Let assumption 3.1 be satisfied and γ2>2(L+m)2m\gamma^2>\frac{2(L+m)^2}{m}. Fix δ>0\delta>0. There exist dimension free constants CC, RR, χ\chi depending only on the problem parameters LL, mm, RR and CC in addition on δ\delta such that for T>δT>\delta

KL⁡(ZT∥ZT′)≤C(R∨∣z0−z0′∣)e−χT∣z0−z0′∣.(52)\operatorname{KL}(Z_T\|Z'_T)\leq C(R\vee|z_0-z'_0|)e^{-\chi T}|z_0-z'_0|. \tag*{(52)}

The constants RR is polynomial in the problem parameters.

Proof. Note that

∣Tt^∣2=∣Xt^−Xt^aux∣2+∣Xt^−Xt^aux+2γt^(Vt^−Vt^aux)∣2(53)|T_{\hat{t}}|^2=|X_{\hat{t}}-X^{\mathrm{aux}}_{\hat{t}}|^2+|X_{\hat{t}}-X^{\mathrm{aux}}_{\hat{t}}+\frac{2}{\gamma_{\hat{t}}}(V_{\hat{t}}-V^{\mathrm{aux}}_{\hat{t}})|^2 \tag*{(53)}
≲∣Xt^−Xt^aux∣2+∣Xt^−Xt^aux+2γ(Vt^−Vt^aux)∣2\lesssim|X_{\hat{t}}-X^{\mathrm{aux}}_{\hat{t}}|^2+|X_{\hat{t}}-X^{\mathrm{aux}}_{\hat{t}}+\frac{2}{\gamma}(V_{\hat{t}}-V^{\mathrm{aux}}_{\hat{t}})|^2

since γt^≥γ\gamma_{\hat{t}}\geq\gamma. Then, combining lemmas 4.2 to 4.4 we find

KL⁡(ZT∥ZT′)≲(∣ω∣3(e−4ω(T−t^)−1)γ(e−ω(T−t^)−1)4+∣ω∣γ(e−2ω(T−t^)−1)(e−ω(T−t^)−1)2)(1∨γ−1)2(R∨∣z0−z0′∣)Ce−χt^∣z0−z0′∣.(54)\operatorname{KL}(Z_T\|Z'_T)\lesssim\left(\frac{|\omega|^3(e^{-4\omega(T-\hat{t})}-1)}{\gamma(e^{-\omega(T-\hat{t})}-1)^4}+\frac{|\omega|\gamma(e^{-2\omega(T-\hat{t})}-1)}{(e^{-\omega(T-\hat{t})}-1)^2}\right)(1\vee\gamma^{-1})^2(R\vee|z_0-z'_0|)Ce^{-\chi\hat{t}}|z_0-z'_0|. \tag*{(54)}

Using that 1+t≤et1+t\leq e^t, T−t^=δT-\hat{t}=\delta and overwriting the constant CC we obtain

KL⁡(ZT∥ZT′)≲(1δ3+1δ)C(R∨∣z0−z0′∣)∣z0−z0′∣e−χT.(55)\operatorname{KL}(Z_T\|Z'_T)\lesssim\left(\frac{1}{\delta^3}+\frac{1}{\delta}\right)C(R\vee|z_0-z'_0|)|z_0-z'_0|e^{-\chi T}. \tag*{(55)}

Remark 4.3. Note that the rate χ\chi is the exact same convergence rate of [Schuh, 2024]. While RR is only polynomial, C,χC,\chi deteriorate exponentially in the problem parameters. All constants are independent of the dimension dd.

Harnack inequalities

It is well known that reverse transportation inequalities, such as in theorem 4.1, are dual to Harnack inequalities. We give a proof for completeness.

Lemma 4.5 (Logarithmic Harnack Inequality). Let PtP_t be the semigroup of (ULD) and z,z′∈R2dz,z' \in\mathbb{R}^{2d}. Then theorem 4.1 is equivalent to

Ptlog⁡f(z)≤C(R∨∣z−z′∣)e−χt∣z−z′∣+log⁡Ptf(z′),(56)P_t \log f(z) \le C(\mathcal{R} \vee|z-z'|)e^{-\chi t}|z-z'|+\log P_t f(z'), \tag*{(56)}

for all measurable, bounded, and strictly positive ff.

Proof. “⇒\Rightarrow”: By the Donsker–Varadhan theorem [5 Theorem 1.5.7] we have

Eμ[g]≤KL⁡(μ∣ν)+log⁡Eν[eg],(57)\mathbb{E}_{\mu}[g] \le\operatorname{KL}(\mu\mid\nu)+\log\mathbb{E}_{\nu}[e^g], \tag*{(57)}

for probability measures μ,ν\mu,\nu and any bounded and measurable test function gg. Now let z,z′∈R2dz,z' \in\mathbb{R}^{2d} and set μ=δzPt\mu=\delta_zP_t, ν=δz′Pt\nu=\delta_{z'}P_t and set g=log⁡fg=\log f for smooth, bounded ff satisfying f≥ε>0f\ge\varepsilon>0 for some ε\varepsilon. Inserting into (57) and using theorem 4.1 to bound the KL term yields

Ptlog⁡f(z)=∫log⁡f d(δzPt)≤KL⁡(δzPt∣δz′Pt)+log⁡∫f d(δz′Pt)(58)P_t\log f(z)=\int\log f\,\mathrm{d}(\delta_zP_t)\le\operatorname{KL}(\delta_zP_t\mid\delta_{z'}P_t)+\log\int f\,\mathrm{d}(\delta_{z'}P_t) \tag*{(58)}
≤KL⁡(δzPt∣δz′Pt)+log⁡Ptf(z′).(59)\le\operatorname{KL}(\delta_zP_t\mid\delta_{z'}P_t)+\log P_tf(z'). \tag*{(59)}

The result then follows by inserting the upper bound of the KL from theorem 4.1. For arbitrary bounded, measurable and strictly positive ff, we may approximate ff by fn=f∨1nf_n=f\vee\frac{1}{n}. Then, by monotonicity we have

Ptlog⁡f(z)≤Ptlog⁡fn(z)≤KL⁡(δzPt∣δz′Pt)+log⁡∫fnd(δz′Pt)(60)P_t\log f(z)\le P_t\log f_n(z)\le\operatorname{KL}(\delta_zP_t\mid\delta_{z'}P_t)+\log\int f_n\mathrm{d}(\delta_{z'}P_t) \tag*{(60)}
→KL⁡(δzPt∣δz′Pt)+log⁡Ptf(z′).(61)\to\operatorname{KL}(\delta_zP_t\mid\delta_{z'}P_t)+\log P_tf(z'). \tag*{(61)}

where the convergence follows from the monotone convergence theorem. “⇐\Leftarrow”: Inserting f=egf=e^g with gg bounded and measurable into (56) yields

EδzPt[g]−log⁡Eδz′Pt[eg]≤C(R∨∣z−z′∣)e−χt∣z−z′∣.(62)\mathbb{E}_{\delta_zP_t}[g]-\log\mathbb{E}_{\delta_{z'}P_t}[e^g]\le C(\mathcal{R}\vee|z-z'|)e^{-\chi t}|z-z'|. \tag*{(62)}

Taking the supremum over gg yields the result again by Donsker–Varadhan. □

Conclusion, limitations and future work

We provided a proof for a long-term reverse transportation, respectively, Harnack inequality for the underdamped Langevin dynamics in the non-logconcave setting. The proof relies on a coupling construction with an interpolating auxiliary process. In particular, we retain the Wasserstein-1 convergence rates for the underdamped Langevin dynamics from [Schuh, 2024].

Limitations and Future work An interesting direction for future work is the adaptation of the results for convergence in Rényi divergence. As pointed out in [13] in this case it seems necessary to rely on a shift construction with bounded shifts. Moreover, we plan to combine our results with [14] to obtain convergence results for annealed underdamped Langevin dynamics in the non-convex setting.

Acknowledgements

This research was funded in whole or in part by the Austrian Science Fund (FWF) [10.55776/PAT4839425, 10.55776/F100800]. For open access purposes, the author has applied a CC BY public copyright license to any author-accepted manuscript version arising from this submission.

Rigorous arguments for the application of Girsanov’s theorem

Let (Ω,A,(Ft)t≥0,P)(\Omega,\mathcal{A},(\mathcal{F}_t)_{t\geq0},\mathbb{P}) denote a filtered probability space satisfying the usual assumptions. For z0,z0′∈R2dz_0,z'_0 \in\mathbb{R}^{2d} and t^∈[0,T]\hat{t}\in[0,T] we define the coupled SDE,

dXt=VtdtdVt={−γVt−∇U(Xt)} dt+2γ dWtdXtaux=Vtaux dtdVtaux={−γVtaux−∇U(Xtaux)+ηtx(Xt−Xtaux)+ηtv(Vt−Vtaux)} dt+2γ dW‾t.(63)\begin{aligned} \mathrm{d}X_t &= V_t\mathrm{d}t \\ \mathrm{d}V_t &= \{-\gamma V_t-\nabla U(X_t)\}\,\mathrm{d}t+\sqrt{2\gamma}\,\mathrm{d}W_t \\ \mathrm{d}X_t^{\mathrm{aux}} &= V_t^{\mathrm{aux}}\,\mathrm{d}t \\ \mathrm{d}V_t^{\mathrm{aux}} &= \{-\gamma V_t^{\mathrm{aux}}-\nabla U(X_t^{\mathrm{aux}})+\eta_t^x(X_t-X_t^{\mathrm{aux}})+\eta_t^v(V_t-V_t^{\mathrm{aux}})\}\,\mathrm{d}t+\sqrt{2\gamma}\,\mathrm{d}\overline{W}_t. \tag*{(63)} \end{aligned}

Z0=z0, Z0aux=z0′Z_0=z_0,\ Z_0^{\mathrm{aux}}=z'_0

where the shift parameters ηtx,ηtv\eta_t^x,\eta_t^v are to be defined later. For the coupling of the Brownian motions (Wt)t(W_t)_t, (W‾t)t(\overline{W}_t)_t we refer to § 4.2, respectively, [Schuh, 2024].

Next, define the measure Q\mathbb{Q} via Nt:=−ηtx(Xt−Xtaux)−ηtv(Vt−Vtaux)N_t:=-\eta_t^x(X_t-X_t^{\mathrm{aux}})-\eta_t^v(V_t-V_t^{\mathrm{aux}}) and

dQdP=e12γ∫0tNsdW‾s−14γ∫0t∣Ns∣2ds.(64)\frac{\mathrm{d}\mathbb{Q}}{\mathrm{d}\mathbb{P}}=\mathrm{e}^{\frac{1}{\sqrt{2\gamma}}\int_0^tN_s\mathrm{d}\overline{W}_s-\frac{1}{4\gamma}\int_0^t|N_s|^2\mathrm{d}s}. \tag*{(64)}

respectively,

dPdQ=e−12γ∫0tNsdW‾s+14γ∫0t∣Ns∣2ds=:Mt((Zs,Zsaux)s≤t).(65)\frac{\mathrm{d}\mathbb{P}}{\mathrm{d}\mathbb{Q}}=\mathrm{e}^{-\frac{1}{\sqrt{2\gamma}}\int_0^tN_s\mathrm{d}\overline{W}_s+\frac{1}{4\gamma}\int_0^t|N_s|^2\mathrm{d}s}=:M_t((Z_s,Z_s^{\mathrm{aux}})_{s\leq t}). \tag*{(65)}

where MtM_t is a function purely of (Zs,Zsaux)s≤t(Z_s,Z_s^{\mathrm{aux}})_{s\leq t} since also the Brownian motion (W‾t)t(\overline{W}_t)_t can be expressed in terms of this process. By Girsanov’s theorem (Bt)t(B_t)_t defined via B0=0B_0=0 and

2γ dBt={ηtx(Xt−Xtaux)+ηtv(Vt−Vtaux)}dt+2γ dW‾t.(66)\sqrt{2\gamma}\,\mathrm{d}B_t=\{\eta_t^x(X_t-X_t^{\mathrm{aux}})+\eta_t^v(V_t-V_t^{\mathrm{aux}})\}\mathrm{d}t+\sqrt{2\gamma}\,\mathrm{d}\overline{W}_t. \tag*{(66)}

is a Q\mathbb{Q}-Brownian motion and (AUX) satisfies Q\mathbb{Q}-a.s.

{dXtaux=Vtaux dtdVtaux={−γVtaux−∇U(Xtaux)}dt+2γ dBt,(X0aux,V0aux)=(x0′,v0′)\left\{ \begin{aligned} \mathrm{d}X_t^{\mathrm{aux}} &= V_t^{\mathrm{aux}}\,\mathrm{d}t \\ \mathrm{d}V_t^{\mathrm{aux}} &= \left\{-\gamma V_t^{\mathrm{aux}}-\nabla U(X_t^{\mathrm{aux}})\right\}\mathrm{d}t+\sqrt{2\gamma}\,\mathrm{d}B_t, \\ (X_0^{\mathrm{aux}},V_0^{\mathrm{aux}})&=(x'_0,v'_0) \end{aligned} \right.

The Q\mathbb{Q}-distribution of the first components (Xt,Vt)t(X_t,V_t)_{t} in (AUX) is not directly tractable as it is unclear how the Brownian motion (Wt)t(W_t)_t behaves under the change of measure, but it is also irrelevant for our purposes. If we denote as μSHIFTz0\mu_{\mathrm{SHIFT}}^{z_0} the path measure of (AUX) in P\mathbb{P} and as μUDz0\mu_{\mathrm{UD}}^{z_0} the one in Q\mathbb{Q} we have

dμSHIFTz0μUDz0((zs,zsaux)s≤t)=Mt((zs,zsaux)s≤t)(67)\frac{\mathrm{d}\mu_{\mathrm{SHIFT}}^{z_0}}{\mu_{\mathrm{UD}}^{z_0}}\left((z_s,z_s^{\mathrm{aux}})_{s\leq t}\right)=M_t\left((z_s,z_s^{\mathrm{aux}})_{s\leq t}\right) \tag*{(67)}

It follows

KL⁡(Zaux∣Z)=EP[log⁡dμSHIFTz0μUDz0((Zs,Zsaux)s≤t)]=∫0tEP[14γ∣ηtx(Xt−Xtaux)+ηtv(Vt−Vtaux)∣2]ds.(68)\begin{aligned} \operatorname{KL}(Z^{\mathrm{aux}}\mid Z)=\mathbb{E}_{\mathbb{P}}\left[\log\frac{\mathrm{d}\mu_{\mathrm{SHIFT}}^{z_0}}{\mu_{\mathrm{UD}}^{z_0}}\left((Z_s,Z_s^{\mathrm{aux}})_{s\leq t}\right)\right] \\ &=\int_0^t\mathbb{E}_{\mathbb{P}}\left[\frac{1}{4\gamma}\left|\eta_t^x(X_t-X_t^{\mathrm{aux}})+\eta_t^v(V_t-V_t^{\mathrm{aux}})\right|^2\right]\mathrm{d}s. \tag*{(68)} \end{aligned}

Remark A.1. The application of Girsanov’s theorem can be justified using standard localization techniques.

Helper Results

Lemma B.1. Let assumption 3.1 hold. Define

H(x,y):=∫01∇2U(y+t(x−y))dt.(69)\mathcal{H}(x,y):=\int_0^1\nabla^2U\left(y+t(x-y)\right)\mathrm{d}t. \tag*{(69)}

Then ∣x−y∣≥R′|x-y|\geq R' implies H(x,y)⪰m′:=R′−2RR′m−L 2RR′\mathcal{H}(x,y)\succeq m' := \frac{R'-2R}{R'}m-\frac{L\,2R}{R'}. In particular, if ∣x−y∣≥4(1+Lm)R|x-y|\geq4(1+\frac{L}{m})R, then H(x,y)⪰m2\mathcal{H}(x,y)\succeq\frac{m}{2}.

Proof. Let w∈Rd∖{0}w\in\mathbb{R}^d\setminus\{0\}. Indeed, by assumption 3.1

wT∫01∇2U(y+t(x−y))dt w≥∣w∣2∫01(1∣y+t(x−y)∣≥Rm−1∣y+t(x−y)∣<RL)dt≥((1−λ)m−λL)∣w∣2(70)\begin{aligned} w^{\mathsf{T}}\int_0^1\nabla^2U\left(y+t(x-y)\right)\mathrm{d}t\,w\geq|w|^2\int_0^1\left(\mathbf{1}_{|y+t(x-y)|\geq R}m-\mathbf{1}_{|y+t(x-y)|<R}L\right)\mathrm{d}t \\ &\geq\left((1-\lambda)m-\lambda L\right)|w|^2 \tag*{(70)} \end{aligned}

where we define λ∈[0,1]\lambda\in[0,1] is defined as

λ=∣{t∈[0,1]∣∣y+t(x−y)∣<R}∣(71)\lambda=\left|\left\{t\in[0,1]\mid|y+t(x-y)|<R\right\}\right| \tag*{(71)}

with ∣⋅∣|\cdot| the Lebesgue measure. One can easily verify that for ∣x−y∣≥R′|x-y|\geq R', λ≤1∧2RR′\lambda\leq1\wedge\frac{2R}{R'}, implying the desired result.

References

  1. [1]J. M. Altschuler and S. Chewi. Shifted composition i: Harnack and reverse transport inequalities. IEEE Trans. Inf. Theory., 71(1):90–113, Jan. 2025. ISSN 0018-9448.arxiv.org/abs/2311.14520
  2. [2]J. M. Altschuler, S. Chewi, and M. S. Zhang. Shifted composition iv: toward ballistic acceleration for log-concave sampling. In Proceedings of the 58th Annual ACM Symposium on Theory of Computing, pages 1739–1750, 2026.
  3. [3]M. Arnaudon, A. Thalmaier, and F.-Y. Wang. Harnack inequality and heat kernel estimates on manifolds with curvature unbounded below. Bulletin des Sciences Mathématiques, 130(3):223–233, 2006. ISSN 0007-4497.DOI
  4. [4]X. Cheng, N. S. Chatterji, P. L. Bartlett, and M. I. Jordan. Underdamped langevin mcmc: A non-asymptotic analysis. In Conference on learning theory, pages 300–323. PMLR, 2018.arxiv.org/abs/1707.03663
  5. [5]S. Chewi. Log-concave sampling. Forthcoming, 2026. Available online at https://chewisinho.github.io/.
  6. [6]A. S. Dalalyan and L. Riou-Durand. On sampling from a log-concave density using kinetic langevin diffusions. ArXiv, abs/1807.09382, 2018.
  7. [7]A. Eberle and F. Lörler. Non-reversible lifts of reversible diffusion processes and relaxation times. Probability Theory and Related Fields, 194(1):173–203, Feb 2026. ISSN 1432-2064.
  8. [8]A. Eberle, A. Guillin, and R. Zimmer. Couplings and quantitative contraction rates for langevin dynamics. The Annals of Probability, 2019.arxiv.org/abs/1703.01617
  9. [9]A. Guillin and F.-Y. Wang. Degenerate fokker-planck equations : Bismut formula, gradient estimate and harnack inequality, 2012. URL https://arxiv.org/abs/1103.2817.
  10. [10]I. Karatzas and S. Shreve. Brownian motion and stochastic calculus. springer, 2014.
  11. [11]B. Leimkuhler and C. Matthews. Rational construction of stochastic numerical methods for molecular sampling. Applied Mathematics Research Express, 2013(1):34–56, 06 2012.
  12. [12]B. J. Leimkuhler, D. Paulin, and P. A. Whalley. Contraction and convergence rates for discretized kinetic langevin dynamics. SIAM Journal on Numerical Analysis, 62(3):1226–1258, 2024.DOI
  13. [13]J. Lu and Y. Wang. Long-time reverse transportation inequalities for non-globally-dissipative langevin dynamics, 2026.arxiv.org/abs/2512.18598
  14. [14]L. Nagler, A. Falk, and A. Habring. Annealed underdamped langevin dynamics, 2026. URL https://arxiv.org/abs/2609.32446.

Paper details

vlnc:W_nga48e4r
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

Not recorded for this paper

View statement
Looking for related papers…
Contents