Log-Semiconvexity: Theory and Applications

• Log-semiconvexity is a relaxation of log-concavity, defined via potentials U for which U(x) + (m/2)||x||^2 is convex, enabling the modeling of nonconvex and nonsmooth density functions.
• It permits potentials to grow convexly up to a quadratic term, capturing complex behaviors like multimodality and discontinuous gradients.
• This concept underlies improved convergence guarantees in score-based generative models, extending Wasserstein-2 analyses to a broader class of distributions.

Log-semiconvexity formalizes a structural relaxation of log-concavity for density functions on $\mathbb{R}^d$, capturing potentials $U$ such that $e^{-U(x)}$ defines a broad, practically relevant class of probability measures that are not necessarily smooth, convex, or even differentiable everywhere. In contrast to log-concavity, which corresponds to convexity of $U$, log-semiconvexity admits potentials growing “convexly up to a quadratic”—essential for modeling nonconvex, multimodal, or nonsmooth data distributions. This property underlies recent advances in convergence guarantees for generative models, specifically in extending rigorous Wasserstein-2 analysis to distributions with discontinuous gradients and nonconvex structure (Bruno et al., 6 May 2025).

1. Definition and Characterizations

Let $U: \mathbb{R}^d \to \mathbb{R}\cup\{+\infty\}$ be lower semi-continuous. The potential $U$ is called $m$-semiconvex (or log-semiconvex) for some $m\in\mathbb{R}$ if the mapping

$x \mapsto U(x) + \frac{m}{2}\lVert x\rVert^2$

is convex on $\mathbb{R}^d$. Equivalently, several characterizations exist:

• Second-derivative characterization: If $U$0, then $U$1 is $U$2-semiconvex iff $U$3 for all $U$4.
• Subgradient characterization: For any $U$5 and $U$6, $U$7,

$U$8

The sign and value of $U$9 interpolate between several important regimes:

• $e^{-U(x)}$0: $e^{-U(x)}$1 is convex, so $e^{-U(x)}$2 is log-concave.
• $e^{-U(x)}$3: $e^{-U(x)}$4 is strongly convex ($e^{-U(x)}$5-strong), so $e^{-U(x)}$6 is strongly log-concave.
• $e^{-U(x)}$7: $e^{-U(x)}$8 need not be convex; it grows convexly up to a quadratic.

2. Structural Implications and Properties

Semiconvexity is strictly weaker than strong convexity, permitting a much broader class of potentials, including nonconvex and nondifferentiable cases. Log-semiconvex potentials may possess discontinuous gradients on sets of measure zero, yet remain locally Lipschitz due to the semiconvexity condition via subdifferential calculus.

A central refinement is to consider $e^{-U(x)}$9 that is globally semiconvex but becomes strongly convex beyond a certain radius. Specifically, within a ball of radius $U$0 around some $U$1, $U$2 is $U$3-semiconvex; outside, $U$4 becomes $U$5-strongly convex:

• Inside: $U$6 $U$7-semiconvexity.
• Outside: $U$8 $U$9-strong convexity.

This split structure accommodates distributions with nonconvex or non-smooth “core” regions but well-behaved tails—a recurring scenario in applied modeling of complex data.

3. Relation to Convexity, Log-Concavity, and Weak Convexity

Log-semiconvexity encompasses log-concavity as a special case and extends it in two principal ways:

• For $U: \mathbb{R}^d \to \mathbb{R}\cup\{+\infty\}$0, log-semiconvexity coincides with log-concavity.
• For $U: \mathbb{R}^d \to \mathbb{R}\cup\{+\infty\}$1, distributions are strongly log-concave and inherit associated functional inequalities, like log-Sobolev.

A closely related concept is weak convexity, defined via a profile function $U: \mathbb{R}^d \to \mathbb{R}\cup\{+\infty\}$2. Every $U: \mathbb{R}^d \to \mathbb{R}\cup\{+\infty\}$3 that is semiconvex within a ball and strongly convex outside is weakly convex with a suitable profile, and conversely, weak convexity under certain profiles implies semiconvexity plus strong convexity at infinity (see Proposition 3.13 in (Bruno et al., 6 May 2025)). This relationship is pivotal for extending analytic guarantees to a wider spectrum of data distributions.

4. Representative Examples

Several canonical families of potentials illustrate log-semiconvexity’s scope:

Example	Potential $U: \mathbb{R}^d \to \mathbb{R}\cup\{+\infty\}$4	Semiconvexity/Convexity Structure
Symmetric modified half-normal	$U: \mathbb{R}^d \to \mathbb{R}\cup\{+\infty\}$5	Continuous, differentiable a.e., $U: \mathbb{R}^d \to \mathbb{R}\cup\{+\infty\}$6, $U: \mathbb{R}^d \to \mathbb{R}\cup\{+\infty\}$7
Gaussian mixtures	$U: \mathbb{R}^d \to \mathbb{R}\cup\{+\infty\}$8	Weakly convex, semiconvex+strongly convex at infinity
Double-well	$U: \mathbb{R}^d \to \mathbb{R}\cup\{+\infty\}$9	$U$0-semiconvex, strongly convex at infinity
Elastic-net	$U$1	Nonsmooth (convex $U$2 part), strongly convex for large $U$3
Max-type	$U$4	Nonconvex, covered by log-semiconvexity assumptions

These examples encompass both smooth and non-smooth, unimodal and multimodal, and nonconvex density structures.

5. Analytic Consequences: Contractivity and Score Properties

Log-semiconvexity, especially when paired with strong convexity at infinity, yields strong control over process contractivity and the behavior of associated score functions. Specifically, for the Ornstein–Uhlenbeck forward process, the score function $U$5 satisfies a monotonicity (one-sided Lipschitz) condition: $U$6 with $U$7 explicit and time-dependent, interpolating from $U$8 at $U$9 to $m$0 as $m$1. For large $m$2, the score is contractive ($m$3), ensuring uniqueness and stability properties crucial for generative modeling analysis. For small $m$4, the parameter $m$5 reflects the degree of nonconvexity permitted by the local semiconvexity bound. Such structural properties are cornerstone elements in non-asymptotic Wasserstein-2 convergence analysis for score-based generative models (Bruno et al., 6 May 2025).

6. Role in Score-Based Generative Models and Implications

Log-semiconvexity broadens the admissible class for which explicit, dimension-optimal, and non-asymptotic convergence guarantees can be established for score-based generative models (SGMs) that rely on stochastic differential equations and reverse-diffusion sampling. The framework enables rigorous analysis for data distributions not satisfying classical smoothness or log-concavity, thus better reflecting empirical data encountered in vision, audio, reinforcement learning, and biology. These advances facilitate provable performance bounds for SGMs on multimodal, nonconvex, and nonsmooth data—bridging empirical effectiveness with theoretical rigor (Bruno et al., 6 May 2025).