New Characterizations of Matrix $Φ$-Entropies, Poincaré and Sobolev Inequalities and an Upper Bound to Holevo Quantity.

Hao–Chung Cheng, Min-Hsiu HsieharXiv 2015 · arXiv

Abstract

We derive new characterizations of matrix $\Phi$-entropies, extending the framework introduced in previous literature. These characterizations facilitate a deeper understanding of matrix $\Phi$-entropy properties and serve as a robust tool for establishing matrix concentration inequalities for functions of independent random variables. Key results include a matrix Poincaré inequality, a new proof for the matrix Efron-Stein inequality, and logarithmic Sobolev inequalities for matrix-valued functions. Furthermore, we apply these findings to quantum information theory, establishing a new upper bound for the Holevo quantity under Markov evolution.

Introduction

In high-dimensional statistics and quantum computing, understanding the concentration of random variables is paramount. While classical $\Phi$-entropies—generalizing concepts like variance and Shannon entropy—provide a robust framework for proving concentration inequalities for scalar values, the matrix setting introduces significant complexity due to non-commutativity ($AB \neq BA$). This work fills a critical gap by providing a comprehensive "manual" for matrix $\Phi$-entropies, proving that the equivalent characterizations used in the classical scalar world have direct functional analogues in the matrix domain.

Bridging Classical and Matrix $\Phi$-Entropies

The core contribution of this research is the discovery of a "one-to-one" correspondence between classical and matrix-valued $\Phi$-entropies. The authors identify ten distinct, equivalent properties that define these entropy classes. These range from joint convexity and subadditivity to specific functional inequalities.

Table 1: Comparison of equivalent characterizations for Classical (C1) and Matrix (C2) $\Phi$-Entropy classes.

The methodology relies on advanced operator theory, including Fréchet derivatives to handle matrix-valued calculus and the Löwner partial ordering for comparing Hermitian matrices. A pivotal insight is that the subadditivity property—where the entropy of a joint system is bounded by the sum of its parts—serves as the "engine" for deriving more complex concentration tools.

Matrix Concentration Inequalities

By establishing the subadditivity of matrix $\Phi$-entropies, the authors derive several fundamental inequalities:

  • Matrix Poincaré Inequality: Theorem 4.2 provides a generalized bound for the variance of a matrix-valued function based on the expectation of its gradient's norm.
  • Log-Sobolev Inequalities: Using the matrix Bonami-Beckner inequality, the authors derive Sobolev inequalities for functions on Boolean hypercubes and Gaussian distributions.
  • Defective Constants: For matrix dimensions $d > 1$, these inequalities include a $\log(d)$ term. This "defect" is a necessary consequence of the matrix structure, distinguishing them from the "tight" scalar versions.

Applications in Quantum Information

The theoretical framework extends directly to quantum information theory. The authors identify that the Holevo quantity—a measure of the classical information capacity of a quantum state—is a specific instance of matrix $\Phi$-entropy where $\Phi(x) = x \log x$. This connection allows for the establishment of a new upper bound for the Holevo quantity of a classical-quantum ensemble undergoing Markov evolution, effectively acting as a "strong data processing inequality."

Conclusion and Future Directions

While this work provides a definitive mapping between classical and matrix entropies, several avenues for future research remain. The relationship between this framework and non-commutative $L_p$ spaces is yet to be fully elucidated. Additionally, while the current theory handles multivariate functions of commuting matrices, the development of a theory for fully non-commutative multivariate matrix functions remains an open challenge.

Summary updated

Abstract

We derive new characterizations of the matrix $\Phi$-entropies introduced in [Electron.~J.~Probab., 19(20): 1--30, 2014}]. These characterizations help to better understand the properties of matrix $\Phi$-entropies, and are a powerful tool for establishing matrix concentration inequalities for matrix-valued functions of independent random variables. In particular, we use the subadditivity property to prove a Poincar\'e inequality for the matrix $\Phi$-entropies. We also provide a new proof for the matrix Efron-Stein inequality. Furthermore, we derive logarithmic Sobolev inequalities for matrix-valued functions defined on Boolean hypercubes and with Gaussian distributions. Our proof relies on the powerful matrix Bonami-Beckner inequality. Finally, the Holevo quantity in quantum information theory is closely related to the matrix $\Phi$-entropies. This allows us to upper bound the Holevo quantity of a classical-quantum ensemble that undergoes a special Markov evolution.
BibTeX
@misc{cheng2015new,
  author = {Hao–Chung Cheng and Min-Hsiu Hsieh},
  title = {{New Characterizations of Matrix $Φ$-Entropies, Poincaré and Sobolev Inequalities and an Upper Bound to Holevo Quantity.}},
  year = {2015},
  eprint = {1506.06801},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1506.06801},
}
Figure