[Paper Review] Regularization in $L_1$ for the Ornstein--Uhlenbeck semigroup

Joseph Lehec|arXiv (Cornell University)|May 15, 2015

Stochastic processes and financial applications4 references3 citations

TL;DR

This paper simplifies and improves upon prior work by Eldan and Lee, establishing a dimension-independent tail bound for the Ornstein–Uhlenbeck semigroup in $L^1$. It removes the $\log\log r$ factor in the decay rate, proving that $\gamma_n(\{Q_tf > r\}) \leq C \frac{\max(1,t^{-1})}{r\sqrt{\log r}}$ for all $r > 1$, with a universal constant $C$, under the assumption that $f \geq 0$ and $\int f\,d\gamma_n = 1$. The result is sharp in terms of $r$-dependence, as shown by dimension-1 examples.

ABSTRACT

Let $γ_n$ be the standard Gaussian measure on $\mathbb R^n$ and let $(Q_t)$ be the Ornstein--Ulhenbeck semigroup. Eldan and Lee recently established that for every non--negative function $f$ of integral $1$ and any time $t$ the following tail inequality holds true: \[ γ_n ( \{ Q_t f > r \} ) \leq C_t \, \frac{ (\log \log r)^4 }{r \sqrt{\log r}} , \quad \forall r>1 \] where $C_t$ is a constant depending on $t$ but not on the dimension. The purpose of the present paper is to simplify parts of their argument and to remove the $(\log \log r)^4$ factor.

Motivation & Objective

• To improve upon Eldan and Lee's $L^1$ regularization result for the Ornstein–Uhlenbeck semigroup by removing the $\log\log r$ factor in the tail bound.
• To simplify the proof techniques of Eldan and Lee while preserving the dimension-independent decay rate.
• To establish a sharp tail estimate of the form $\gamma_n(\{Q_tf > r\}) \leq C \frac{\max(1,t^{-1})}{r\sqrt{\log r}}$ for non-negative $f$ with $\int f\,d\gamma_n = 1$.
• To confirm the optimality of the $1/\sqrt{\log r}$ decay rate via explicit counterexamples in dimension one.

Proposed method

• Uses a key lemma showing that $\nabla^2 \log(Q_tf) \geq -\frac{1}{2t}\,\text{id}$ pointwise, linking the semigroup to log-concave-like regularity.
• Reduces the problem to analyzing functions $f$ satisfying $\nabla^2 \log f \geq -\beta\,\text{id}$, which models the regularity induced by $Q_t$.
• Applies a stochastic calculus approach via Girsanov's theorem to control the tail probability $\mathsf{P}(f(X_1) > r)$ by constructing a time-changed Brownian motion and a martingale change of measure.
• Uses a deviation bound via Lemma 8: if $\mathsf{E}[e^Z] \leq 1$, then $\mathsf{P}(Z \leq -2) \leq -\mathsf{E}[Z]$, to control rare events in the likelihood ratio.
• Employs a decomposition of the likelihood ratio into three terms, with the critical term handled by a modified Girsanov-type argument and moment estimates.
• Applies Tchebychev and Markov inequalities to control the tail probabilities of stochastic integrals and quadratic variations in the drift correction terms.

Experimental results

Research questions

• RQ1 Can the $\log\log r$ factor in Eldan and Lee's $L^1$ regularization bound for the Ornstein–Uhlenbeck semigroup be removed?
• RQ2 Is the $1/\sqrt{\log r}$ decay rate in the tail bound optimal, both in terms of $r$-dependence and dimension independence?
• RQ3 Can the proof of the tail estimate be simplified without losing the sharpness of the bound?
• RQ4 Does the regularity condition $\nabla^2 \log f \geq -\beta\,\text{id}$ imply a universal $1/(r\sqrt{\log r})$ tail decay under $L^1$ normalization?
• RQ5 Can the sharpness of the bound be confirmed via explicit examples in dimension one?

Key findings

• The paper establishes a sharp tail bound: $\gamma_n(\{Q_tf > r\}) \leq C \frac{\max(1,t^{-1})}{r\sqrt{\log r}}$ for all $r > 1$, with a universal constant $C$, under the assumption that $f \geq 0$ and $\int f\,d\gamma_n = 1$.
• The $\log\log r$ factor present in Eldan and Lee's original result is completely removed, improving the decay rate to the optimal $1/\sqrt{\log r}$.
• The bound is dimension-independent, with the constant $C$ universal and independent of $n$, the dimension of the space.
• The optimality of the $1/\sqrt{\log r}$ decay is confirmed by a construction in dimension one: for $f_\alpha(x) = e^{\alpha x - \alpha^2/2}$ with $\alpha = \sqrt{2\log r}$, the tail probability is bounded below by $c'/(r\sqrt{\log r})$.
• The result is sharp not only in $r$ but also in the time parameter: the $\max(1,t^{-1})$ factor captures the correct dependence on $t$.
• The proof technique is simplified compared to Eldan and Lee, while retaining the core stochastic calculus argument based on Girsanov's theorem and exponential martingale estimates.