Abstract
This paper provides a powerful generalization of the Golden-Thompson and Araki-Lieb-Thirring inequalities to an arbitrary number of non-commuting matrices. By leveraging complex interpolation and spectral pinching, the authors derive new, tight remainder terms for fundamental entropy laws in quantum mechanics, specifically addressing the "recoverability" of lost information.
Introduction
In matrix analysis and quantum mechanics, the "trace" of matrix exponentials is a fundamental object of study. For two commuting matrices $A$ and $B$, the relationship $\text{tr}(\exp(A+B)) = \text{tr}(\exp(A)\exp(B))$ is trivial. However, matrices representing quantum states rarely commute. The celebrated Golden-Thompson (GT) inequality provides an upper bound for the non-commuting case: $\text{tr}(\exp(A+B)) \leq \text{tr}(\exp(A)\exp(B))$.
The challenge arises when moving beyond two matrices. Straightforward extensions to three or more matrices are generally false. This "non-commutativity gap" makes it difficult to prove tight bounds for complex quantum systems. Specifically, in quantum information theory, researchers seek to understand how much information is lost when a system evolves or is partially discarded (monotonicity of relative entropy). While it was known that information is lost, explicit, tight "remainder terms" were previously lacking to quantify how well the original state could be recovered.
Methodology
The authors employ two distinct mathematical approaches to tackle the problem:
- Asymptotic Spectral Pinching: This intuitive yet technically demanding approach considers $m$-fold tensor products of matrices. As $m$ goes to infinity, "pinching maps" (which zero out off-diagonal blocks) are used to force the matrices to "almost commute" without significant information loss. This allows for the proof of existence for multivariate inequalities.
- Complex Interpolation Theory: To derive explicit, usable formulas, the authors utilize the Stein-Hirschman theorem. By treating matrix functions as holomorphic maps in the complex plane and bounding them on the edges of a "strip," they derive an integral formula for the interior. This leads to the discovery of a specific probability distribution, $\beta_0(t)$, that acts as a weight for the non-commuting rotations of the matrices.
Key Insights: The Rotated Petz Map
One of the most significant contributions is the refinement of the Petz recovery map. By "rotating" the recovery map using the complex distribution $\beta_0(t)$, the authors found a way to quantify recoverability that is tight in the commutative case. This means if the matrices were to commute, the inequality would become an equality.
Furthermore, the paper demonstrates that Lieb’s triple matrix inequality, long considered a gold standard, is actually a specific case of this more general $n$-matrix framework. The authors even provide a strengthened version of Lieb's result.
Key Results
- Multivariate GT and ALT Extensions: The authors successfully extended the Golden-Thompson and Araki-Lieb-Thirring inequalities to $n$ matrices. The new GT inequality is formulated as: $$\log \left\| \exp\left(\sum_{k=1}^n H_k\right) \right\|_p \leq \int_{-\infty}^{\infty} dt \beta_0(t) \log \left\| \prod_{k=1}^n \exp((1+it)H_k) \right\|_p$$
- Strengthened Monotonicity: They proved a new lower bound for the monotonicity of quantum relative entropy. This bound is expressed in terms of the measured relative entropy and the rotated Petz recovery map, providing a concrete answer to an open question in quantum recoverability.
- Strong Sub-Additivity (SSA) Remainder: The authors derived the first explicit remainder terms for the SSA of von Neumann entropy that are tight when the underlying states commute.
Conclusion and Future Work
This work bridges a significant gap in quantum information theory by providing a robust framework for multivariate trace inequalities. Future research may explore bridging the gap between the intuition of pinching and the explicitness of interpolation for more general cases. Additionally, these multivariate extensions could be used to derive sharper "tail bounds" for sums of random matrices, which would have significant implications for fields such as machine learning and statistics.
Summary updated