Concavity of Tsallis Entropy and Tsallis Entropy Power along Heat Flow

1. Problem Statement

The paper investigates the temporal evolution of Tsallis entropy ($S_q$) and Tsallis entropy power ($N_{q,\mu}$) along the heat flow (diffusion process) in arbitrary dimensions ($d \geq 1$).

• Context: While the concavity of Shannon entropy power along the heat flow (Costa's Entropy Power Inequality) is a classical result, the behavior of non-additive entropies like Tsallis entropy in higher dimensions remains largely unexplored.
• Gap: Previous studies on the concavity of Tsallis entropy were restricted to the one-dimensional setting ($d=1$) and often relied on computer-assisted methods (Sum-of-Squares decompositions) that do not scale well to higher dimensions due to the explosion in the size of algebraic constraints.
• Objective: To establish the concavity of Tsallis entropy ($d^2/dt^2 S_q \leq 0$) and Tsallis entropy power in arbitrary dimensions for a nontrivial range of the entropic index $q$, using a fully analytic approach.

2. Methodology

The author introduces a novel, fully analytic framework that avoids computer-assisted optimization. The core components of the methodology are:

• Nonlinear Transformation:
  The heat equation $\partial_t \phi_t = \Delta \phi_t$ is transformed via $u_t = \phi_t^p$ (where $p = 1/(1+\delta)$). This transforms the linear heat equation into a nonlinear equation:
  $$\partial_t u_t = \Delta u_t + \delta \frac{\|\nabla u_t\|^2}{u_t}$$
  This transformation simplifies the calculation of temporal derivatives of the entropy functional.

• Functional Inequalities:
  Instead of relying on Sum-of-Squares (SOS) decompositions, the proof relies on establishing specific functional inequalities involving the Laplacian ($\Delta$) and the gradient ($\nabla$).

  • The core challenge reduces to estimating the term $\int \Delta u_t \frac{\|\nabla u_t\|^2}{u_t} dx$ in terms of $\int (\Delta u_t)^2 dx$.
  • The author derives a new functional inequality (Lemma 19) relating the $L^4$ norm of the gradient to the $L^2$ norm of the Laplacian:
    $$\int \frac{\|\nabla u\|^4}{u^2} dx \leq (2\sqrt{5} + 6) \int \|\Delta u\|^2 dx$$
    (under specific integral symmetry conditions).

• Rigorous Justification of Integration by Parts:
  A significant portion of the work (Section III) is dedicated to rigorously justifying the vanishing of boundary terms at infinity for integration-by-parts operations. The author uses trace theorems and $L^p-L^1$ estimates to prove that these boundary terms vanish for specific ranges of $\delta$ (and thus $q$), a step often glossed over in previous literature.

3. Key Contributions and Results

A. Concavity of Tsallis Entropy

The paper proves that the second time derivative of Tsallis entropy is non-positive ($d^2/dt^2 S_q \leq 0$) along the heat flow for the following ranges of the entropic index $q$:

• One-dimensional ($d=1$): $q \in (0, 3]$.
• Higher-dimensional ($d \geq 2$): $q \in [1, \frac{2(\sqrt{5}+1)}{\sqrt{5}}] \approx [1, 2.894]$.

Note: The paper highlights that the range $(0, 1)$ for $q$ generally fails to yield concavity in dimensions $d \geq 3$, making the lower bound $q=1$ optimal for higher dimensions.

B. Generalized de Bruijn Identity and $q$-Fisher Information

As a consequence of the analysis, the paper recovers the generalized de Bruijn identity:
$$\frac{d}{dt} S_q(\phi_t) = \frac{4}{q} \int \|\nabla \phi_t^{q/2}\|^2 dx =: I_q(\phi_t)$$
where $I_q$ is the $q$-Fisher information. The paper further establishes the monotonicity of this $q$-Fisher information along the heat flow ($d/dt I_q \leq 0$) for the derived ranges of $q$ in dimensions $d \geq 2$.

C. Concavity of Tsallis Entropy Power

The paper defines Tsallis entropy power as $N_{q,\mu}(\rho) = e^{\frac{\mu}{d} S_q(\rho)}$. It proves the concavity of this quantity ($d^2/dt^2 N_{q,\mu} \leq 0$) under two scenarios:

1. Initial Data Constraint: For sufficiently small initial values of $\int \phi_0^q dx$ (depending on $d, q, \mu$).
2. Asymptotic Behavior: For any general initial probability measure, the concavity holds for all $t$ greater than a certain threshold $C(q, \mu, d)$.

D. New Functional Inequality

The derivation yields a new functional inequality (Lemma 19) relating the gradient and Laplacian terms, which is of independent interest to the fields of functional analysis and PDEs.

4. Significance and Impact

• Extension to High Dimensions: This is the first work to rigorously establish the concavity of Tsallis entropy and its power in arbitrary dimensions without relying on computer-assisted methods. Previous SOS-based methods became intractable as dimension $d$ increased.
• Analytic Rigor: The paper provides a mathematically rigorous justification for integration-by-parts identities in the context of heat flow, addressing gaps in previous works where boundary term vanishing was assumed without proof.
• Information-Theoretic Implications:
  • The results extend classical information-theoretic inequalities (like the Entropy Power Inequality) beyond the additive Shannon framework to non-additive settings.
  • The monotonicity of $q$-Fisher information provides new tools for analyzing systems with non-Gaussian behavior, long-range dependence, or heavy tails.
  • These findings have potential applications in coding theory, statistical estimation, and hypothesis testing where classical Gaussian assumptions fail.
• Methodological Shift: The approach demonstrates that nonlinear transformations combined with functional inequalities can solve high-dimensional problems in entropy evolution that were previously thought to require computational algebraic geometry.

5. Limitations and Future Work

• Optimality: The authors do not claim the derived ranges for $q$ are optimal. Specifically, the upper bound for $d \geq 2$ depends on the constants in the derived functional inequality, which may not be sharp.
• Higher-Order Derivatives: The paper focuses on second-order derivatives. Extending this to higher-order derivatives (to test the "Completely Monotone Conjecture") would require new, likely more complex, functional inequalities.
• Other Diffusions: The framework is currently applied to the standard heat flow. Extending it to nonlinear diffusion equations (e.g., porous medium equation) or underdamped Langevin flows remains an open direction.

In summary, this paper provides a breakthrough in the analysis of non-additive entropies by establishing their concavity properties in high dimensions through a novel analytic framework, thereby bridging a significant gap between classical information theory and generalized entropy settings.