---
title: 'Log-Semiconvex Functions: Theory & Applications'
url: https://www.emergentmind.com/topics/log-semiconvex-functions
type: topic
---

# Log-Semiconvex Functions: Theory & Applications

A log-semiconvex function is a smooth, positive-valued function on $\mathbb{R}^n$ whose logarithm exhibits a quantified lower bound on its curvature, controlled by a parameter $\beta \ge 0$. These functions interpolate between log-convexity ($\beta = 0$) and non-convex behaviors, providing a key analytic class for deviation inequalities under Gaussian measures, with deep connections to semigroup regularization, the Ornstein–Uhlenbeck (OU) flow, and the Gaussian analogues of Talagrand’s conjecture [1706.08688].

## 1. Definition and Characterization

Let $g:\mathbb{R}^n\to(0,\infty)$ be smooth. The function $g$ is log-semiconvex with constant $\beta \ge 0$ (or $\beta$-log-semiconvex) if
\[
\mathrm{Hess}(\log g(x)) \succeq -\beta\,\mathrm{Id}, \quad \text{for all } x\in\mathbb{R}^n,
\]
where $\mathrm{Hess}$ is the Hessian and $\succeq$ denotes matrix inequality in the Loewner order. Equivalently, writing $g = e^f$ for some $f:\mathbb{R}^n\rightarrow\mathbb{R}$, this is $\mathrm{Hess}(f)(x) \succeq -\beta\,\mathrm{Id}$. The case $\beta=0$ recovers log-convexity, i.e., convexity of $f$.

## 2. Canonical Examples

Three principal classes under the standard Gaussian measure $\gamma_n$ are documented:

| Function                                    | $\beta$ value           | Notable features                                              |
|----------------------------------------------|-------------------------|---------------------------------------------------------------|
| $g(x) = \exp(a\cdot x + b)$                  | $0$                     | Log-convex (linear exponent, $\beta=0$)                       |
| $g(x) = P_s g$, Ornstein–Uhlenbeck smoothing | $1/(2s)$ for $s>0$      | Smoothing of $g$ preserves log-semiconvexity                  |
| $g(x) = \exp(a\cdot x - |x|^2/(2\theta))$    | $\theta^{-1}$           | “Tilted Gaussians”; Hessian $=-\theta^{-1}\mathrm{Id}$        |

Ornstein–Uhlenbeck smoothing: For $g\in L^1(\gamma_n)$ nonnegative, the OU semigroup $P_sg$ satisfies
\[
\mathrm{Hess}(\log P_sg) \succeq -\frac{1}{2s}\mathrm{Id},\quad s>0,
\]
and is thus $\beta$–log–semiconvex with $\beta=1/(2s)$.

## 3. Deviation Inequalities

A fundamental property of log-semiconvex functions is the existence of sharp deviation inequalities under $\gamma_n$. For $g \ge 0$, $\int g\,d\gamma_n = 1$, and $\mathrm{Hess}(\log g) \succeq -\beta\,\mathrm{Id}$,
\[
\gamma_n\{ g \ge t \} \le \frac{C_\beta}{t\sqrt{\log t}}, \qquad \forall t > 1,
\]
where $C_\beta = \alpha \max\{1,\beta\}$ for universal $\alpha > 0$. Setting $g = e^f$, $t = e^u$,
\[
\gamma_n\{ f \ge u \} \le C_\beta \frac{e^{-u}}{\sqrt{u}},\quad \forall u > 0.
\]

For $\beta=0$ (log-convexity), a sharp form is available:
\[
\gamma_n\{f\ge t\} \le \overline{\Phi}(\sqrt{2t}),
\]
where $\overline{\Phi}(s) = \int_s^\infty \frac{e^{-u^2/2}}{\sqrt{2\pi}} du$ is the upper tail of the standard normal. The standard Gaussian estimate $\overline{\Phi}(s)\le (s\sqrt{2\pi})^{-1}e^{-s^2/2}$ implies the optimality in exponential–root behavior:
\[
\gamma_n\{f\ge t\} \leq \frac{e^{-t}}{2\sqrt{\pi t}}.
\]

## 4. Proof Approaches and Analytical Tools

Two distinctive arguments provide the convex-case (zero semiconvexity) deviation bounds:

**A) Ehrhard-Concavity and Fenchel–Moreau Duality:**  
Define sublevel sets $A_u = \{f \leq u\}$ with $\varphi(u) = \Phi^{-1}(\gamma_n(A_u))$. Convexity of $f$ and Ehrhard’s inequality yield concavity and monotonicity of $\varphi$. Analysis via integration by parts and Fenchel–Legendre conjugation shows
\[
\varphi(u) \ge \inf_{t\le 0} \{ -ut-1/(2t) \} = \sqrt{2u},
\]
establishing the sharp tail via the normal CDF.

**B) Monotone Rearrangement and One-Dimensional Bound:**  
Transport $\gamma_n$ via $f$ to obtain push-forward $\mu_f$ with distribution function $F_f$. Analysis of $T_f = F_f^{-1} \circ \Phi$ using convexity/concavity properties gives
\[
T_f(x) \le \frac{1}{2}x^2 \implies F_f(u) \ge \Phi(\sqrt{2u}),
\]
so that $\gamma_n\{f\ge u\} \le \overline{\Phi}(\sqrt{2u})$.

A general “semi-convex comparison” lemma states: for $f$ with $\int e^f\,d\gamma_n = 1$ and $\mathrm{Hess}(f)\succeq -\beta\,\mathrm{Id}$,
\[
f(x) \le \frac{n}{2}\ln(1+\beta) + \frac{1}{2}|x|^2.
\]
This lemma underpins the full $\beta$–log–semiconvex deviation estimates.

## 5. Connection to Talagrand’s Conjecture and Ornstein–Uhlenbeck Regularization

The classical Talagrand conjecture in discrete settings concerns the $L^1$ smoothing properties of biased convolution on the Boolean hypercube. Its continuous analogue involves regularization by the Ornstein–Uhlenbeck semigroup $P_s$. While hypercontractivity is trivial at $p=1$, it is shown that
\[
\mathrm{Hess}(\log P_sg) \succeq -\frac{1}{2s}\mathrm{Id}.
\]
Combined with general tail bounds for $\beta$–log–semiconvex functions, this yields
\[
\gamma_n\{P_sg \ge t\} \le \frac{C_{1/(2s)}}{t\sqrt{\log t}} \to 0 \quad \text{as } t\to\infty,
\]
proving that $P_s$ improves tail behavior even starting from $L^1$. Lehec [1706.08688] established the sharp asymptotic $1/(t\sqrt{\log t})$ decay with $\beta$-independent constants.

## 6. Sharpness and Further Corollaries

The inequality $\overline{\Phi}(\sqrt{2t})$ is exactly sharp, being achieved for linear functions $f_t(x) = \sqrt{2t}\, x_1 - t$. Preservation of log-convexity under $P_s$ yields, for $\int g\,d\gamma_n=1$ log-convex:
\[
\gamma_n\{P_sg \ge t\} \le \overline{\Phi}\big(\sqrt{2\ln t}\bigr),\quad t\ge 1,
\]
independently of $s$. Furthermore, by applying monotone push-forward maps (e.g., cumulative distribution transforms to the exponential or chi-square), results extend to additional structured measures. These conclusions strictly enhance previous deviation inequalities and are summarized in Corollary 1.2 and Corollary 1.4 [1706.08688].

## 7. Geometric and Probabilistic Implications

Ehrhard’s inequality (Gaussian Brunn–Minkowski) is pivotal, not only ensuring concavity of the quantile transform $\varphi$ but also classical Gaussian facts such as median $\le$ mean for convex $f$. Attempts to extend purely geometric rearrangement methods from the convex ($\beta = 0$) case to the general log-semiconvex case ($\beta > 0$) encounter obstacles; local minima in $f$ can disrupt required semiconvexity of the induced monotone transport $T_f$, as elucidated by van Handel. Consequently, sharp tail bounds for $\beta$–log–semiconvex functions for $\beta > 0$ necessarily depend on semigroup or stochastic localization techniques rather than geometric rearrangement [1706.08688].

Source: https://www.emergentmind.com/topics/log-semiconvex-functions