BZPEER
More

Discover

About BZPEERSearchGraph

Research tools

HomeCollectionsResearchCollaborationPublish

Network

ExpertsCommunities

Opportunities

Listings

Account

Log inCreate account
Log inCreate account

Source author record

Dejun Luo

Explore this author’s publications, collaborators and research topics. This catalog record has not been claimed by its author.

ResearcherUnclaimed source record
math.PRmath.APmath-phmath.MPArtificial IntelligenceDistributed, Parallel, and Cluster Computingmath.CAmath.DGNetworking and Internet Architecture

Catalog footprint

Research overview

21works
9topics
4close collaborators

Actions

Is this your work?

Log in to claim
Open graphBrowse papers

Published work

21 publications

preprint2026arXiv

KVServe: Service-Aware KV Cache Compression for Communication-Efficient Disaggregated LLM Serving

LLMs are widely adopted in production, pushing inference systems to their limits. Disaggregated LLM serving (e.g., PD separation and KV state disaggregation) improves scalability and cost efficiency, but it also turns KV into an explicit payload crossing network and storage boundaries, making KV a dominant end-to-end bottleneck. Existing KV compression are typically static runtime configurations, despite production service context varies over time in workload mix, bandwidth, and SLO/quality budgets. As a result, a fixed choice can be suboptimal or even increase latency. We present \emph{KVServe}, the first service-aware and adaptive KV communication compression framework for disaggregated LLM serving: KVServe (1) unifies KV compression into a modular strategy space with new components and cross-method recomposition; (2) introduces Bayesian Profiling Engine that efficiently searches this space and distills a 3D Pareto candidate set, reducing 50×50\times offline search overhead; and (3) deploys a Service-Aware Online Controller that combines an analytical latency model with a lightweight bandit to select profiles under constraints and correct offline-to-online mismatch. Integrated into vLLM and evaluated across datasets, models, GPUs and networks, KVServe achieves up to 9.13×9.13\times JCT speedup in PD-separated serving and up to 32.8×32.8\times TTFT reduction in KV-disaggregated serving.

Open access
preprint2026arXiv

Structure preservation and emergent dissipation in stochastic wave equations with transport noise

We study nonlinear wave equations perturbed by transport noise acting either on the displacement or on the velocity. Such noise models random advection and, under suitable scaling of space covariance, may generate an effective dissipative term. We establish well-posedness in both cases and analyse the associated scaling limits. When the noise acts on the displacement, the system preserves its original structure and converges to the deterministic nonlinear wave equation, whereas if it acts on the velocity, the rescaled dynamics produce an additional Laplacian damping term, leading to a stochastic derivation of a Westervelt-type acoustic model.

Open access
preprint2016arXiv

Exponential Convergence in LpL^p-Wasserstein Distance for Diffusion Processes without Uniformly Dissipative Drift

By adopting the coupling by reflection and choosing an auxiliary function which is convex near infinity, we establish the exponential convergence of diffusion semigroups (Pt)t≥0(P_t)_{t\ge0} with respect to the standard LpL^p-Wasserstein distance for all p∈[1,∞)p\in[1,\infty). In particular, we show that for the Itô stochastic differential equation $$\d X_t=\d B_t+b(X_t)\,\d t,$$ if the drift term bb satisfies that for any x,y∈Rdx,y\in\R^d, ⟨b(x)−b(y),x−y⟩≤{K1∣x−y∣2,∣x−y∣≤L;−K2∣x−y∣2,∣x−y∣>L\langle b(x)-b(y),x-y\rangle\le \begin{cases} K_1|x-y|^2,& |x-y|\le L; -K_2|x-y|^2,& |x-y|> L \end{cases} holds with some positive constants K1K_1, K2K_2 and L>0L>0, then there is a constant λ:=λ(K1,K2,L)>0λ:=λ(K_1,K_2,L)>0 such that for all p∈[1,∞)p\in[1,\infty), t>0t>0 and x,y∈Rdx,y\in\R^d, $$W_p(δ_x P_t,δ_y P_t)\leq Ce^{-λt/p} \begin{cases} |x-y|^{1/p}, & \mbox{if } |x-y|\le 1; |x-y|, & \mbox{if } |x-y|> 1. \end{cases}$$ where C:=C(K1,K2,L,p)C:=C(K_1,K_2,L,p) is a positive constant. This improves the main result in \cite{Eberle} where the exponential convergence is only proved for the L1L^1-Wasserstein distance.

Open access
preprint2015arXiv

A probabilistic proof of the fundamental gap conjecture via the coupling by reflection

Let Ω⊂RnΩ\subset\mathbb{R}^n be a strictly convex domain with smooth boundary and diameter DD. The fundamental gap conjecture claims that if V:Ωˉ→RV:\barΩ\to\mathbb{R} is convex, then the spectral gap of the Schrödinger operator −Δ+V-Δ+V with Dirichlet boundary condition is greater than 3π2D2\frac{3π^2}{D^2}. Using analytic methods, Andrews and Clutterbuck recently proved in [J. Amer. Math. Soc. 24 (2011), no. 3, 899--916] a more general spectral gap comparison theorem which implies this conjecture. In the first part of the current work, we shall give an independent probabilistic proof of their result via the coupling by reflection of the diffusion processes. Moreover, we also present in the second part a simpler probabilistic proof of the original conjecture.

Open access
preprint2014arXiv

Harnack Inequalities for SDEs with Multiplicative Noise and Non-regular Drift

The log-Harnack inequality and Harnack inequality with powers for semigroups associated to SDEs with non-degenerate diffusion coefficient and non-regular time-dependent drift coefficient are established, based on the recent papers \cite{Flandoli, Zhang11}. We consider two cases in this work: (1) the drift fulfills the LPS-type integrability, and (2) the drift is uniformly Hölder continuous with respect to the spatial variable. Finally, by using explicit heat kernel estimates for the stable process with drift, the Harnack inequality for the stochastic differential equation driven by symmetric stable process is also proved.

Open access
preprint2014arXiv

Quasi-invariance of the stochastic flow associated to Itô's SDE with singular time-dependent drift

In this paper we consider the Itô SDE dXt=dWt+b(t,Xt) dt,X0=x∈Rd,d X_t=d W_t+b(t,X_t)\,d t, \quad X_0=x\in {\mathbb R}^d, where WtW_t is a dd-dimensional standard Wiener process and the drift coefficient b:[0,T]×Rd→Rdb:[0,T]\times{\mathbb R}^d\to{\mathbb R}^d belongs to Lq(0,T;Lp(Rd))L^q(0,T;L^p({\mathbb R}^d)) with p≥2,q>2p\geq 2, q>2 and dp+2q<1\frac dp +\frac 2q<1. In 2005, Krylov and Röckner \cite{KR05} proved that the above equation has a unique strong solution XtX_t. Recently it was shown by Fedrizzi and Flandoli \cite{FF13b} that the solution XtX_t is indeed a stochastic flow of homeomorphisms on Rd{\mathbb R}^d. We prove in the present work that the Lebesgue measure is quasi-invariant under the flow XtX_t.

Open access
preprint2014arXiv

Stochastic Lagrangian flows on the group of volume-preserving homeomorphisms of the spheres

We consider stochastic differential equations on the group of volume-preserving homeomorphisms of the sphere Sd (d≥2)S^d\,(d\geq 2). The diffusion part is given by the divergence free eigenvector fields of the Laplacian acting on L2L^2-vector fields, while the drift is some other divergence free vector field. We show that the equation generates a unique flow of measure-preserving homeomorphisms when the drift has first order Sobolev regularity, and derive a formula for the distance between two Lagrangian flows. We also compute the rotation process of two particles on the sphere S2S^2 when they are close to each other.

Open access
preprint2014arXiv

Uniform Hölder Estimates on Semigroups Generated by Non-Local Operators of Variable Order

We consider the non-local operator of variable order as follows $$Lf(x)= \int_{\R^d\setminus\{0\}}\big(f(x+z)-f(x)-\<\nabla f(x),z\> \I_{\{|z|\le 1\}}\big)\frac{n(x,z)}{|z|^{d+α(x)}}\,dz.$$ Under mild conditions on α(x)α(x) and n(x,z)n(x,z), we establish the Hölder regularity for the associated semigroups. The proof is based on the probabilistic coupling method, and it successfully applies to both stable-like processes in the sense of Bass and time-change of symmetric stable processes.

Open access
preprint2013arXiv

The log-Sobolev inequality for the ground state of a Schrödinger operator on bounded convex domains

We consider the ground state ϕ0ϕ_0 of the Schrödinger operator L=−Δ+VL=-Δ+V on the bounded convex domain Ω⊂RnΩ\subset\R^n, satisfying the Dirichlet boundary condition. Assume that V∈C1(Ω)V\in C^1(Ω) and it admits an even function V~∈C1([−D/2,D/2])\tilde V\in C^1([-D/2,D/2]) as its modulus of convexity, where DD is the diameter of ΩΩ. If the first Dirichlet eigenvalue λ~0\tildeλ_0 of $-\frac{\d^2}{\d t^2}+\tilde V$ on the interval [−D/2,D/2][-D/2,D/2] satisfies λ~0>V~(0)\tildeλ_0>\tilde V(0), then the measure $\dμ=ϕ_0 \d x$ satisfies the log-Sobolev inequality on ΩΩ with the constant λ~0−V~(0)\tildeλ_0-\tilde V(0). In particular, if VV is convex, then the constant is explicitly given by π2D2\frac{π^2}{D^2}.

Open access
preprint2010arXiv

A note on Gaussian correlation inequalities for nonsymmetric sets

We consider the Gaussian correlation inequality for nonsymmetric convex sets. More precisely, if A⊂RdA\subset\mathbb{R}^d is convex and the origin 0∈A0\in A, then for any ball BB centered at the origin, it holds γd(A∩B)≥γd(A)γd(B)γ_d(A\cap B)\geq γ_d(A)γ_d(B), where γdγ_d is the standard Gaussian measure on Rd\mathbb{R}^d. This generalizes Proposition 1 in [Arch. Rational Mech. Anal. 161 (2002), 257--269].

Open access
preprint2010arXiv

Asymptotic estimates on the time derivative of entropy on a Riemannian manifold

We consider the entropy of the solution to the heat equation on a Riemannian manifold. When the manifold is compact, we provide two estimates on the rate of change of the entropy in terms of the lower bound on the Ricci curvature and the spectral gap respectively. Our explicit computation for the three dimensional hyperbolic space shows that the time derivative of the entropy is asymptotically bounded by two positive constants.

Open access
preprint2010arXiv

Quasi-invariant flow generated by Stratonovich SDE with BV drift coefficients

We generalize the results of Ambrosio [Invent. Math. 158 (2004), 227--260] on the existence, uniqueness and stability of regular Lagrangian flows of ordinary differential equations to Stratonovich stochastic differential equations with BV drift coefficients. Then we construct an explicit solution to the corresponding stochastic transport equation in terms of the stochastic flow. The approximate differentiability of the flow is also studied when the drift is a Sobolev vector field.

Open access
preprint2010arXiv

Stochastic differential equations with coefficients in Sobolev spaces

We consider Itô SDE $\d X_t=\sum_{j=1}^m A_j(X_t) \d w_t^j + A_0(X_t) \d t$ on Rd\R^d. The diffusion coefficients A1,...,AmA_1,..., A_m are supposed to be in the Sobolev space Wloc1,p(Rd)W_\text{loc}^{1,p} (\R^d) with p>dp>d, and to have linear growth; for the drift coefficient A0A_0, we consider two cases: (i) A0A_0 is continuous whose distributional divergence δ(A0)δ(A_0) w.r.t. the Gaussian measure γdγ_d exists, (ii) A0A_0 has the Sobolev regularity Wloc1,p′W_\text{loc}^{1,p'} for some p′>1p'>1. Assume $\int_{\R^d} \exp\big[λ_0\bigl(|δ(A_0)| + \sum_{j=1}^m (|δ(A_j)|^2 +|\nabla A_j|^2)\bigr)\big] \dγ_d<+\infty$ for some λ0>0λ_0>0, in the case (i), if the pathwise uniqueness of solutions holds, then the push-forward $(X_t)_# γ_d$ admits a density with respect to γdγ_d. In particular, if the coefficients are bounded Lipschitz continuous, then XtX_t leaves the Lebesgue measure $\Leb_d$ quasi-invariant. In the case (ii), we develop a method used by G. Crippa and C. De Lellis for ODE and implemented by X. Zhang for SDE, to establish the existence and uniqueness of stochastic flow of maps.

Open access

Research graph

See the researcher in context

Open full explorer

Inspect adjacent papers, topics, institutions and collaborators without losing the researcher page.

Building this map preview

BZPEER is loading the nearby papers, people, topics and institutions for this page.