### Pinching

Let $A$ be a Hermitian matrix, and $A = \sum_j \lambda_j P_j$ be the spectral decomposition of $A$.
The *pinching map* defined by $A$ is given by

for any Hermitian matrix $X$.

**Theorem 1.** Let $A$ be a positive semi-definite matrix and $B$ be a Hermitian matrix.
The following statements hold.

- $\mathcal{P}_B (A)$ commutes with $B$.
- $\mathrm{Tr} ( \mathcal{P}_B (A) B ) = \mathrm{Tr} ( A B )$.
*(Pinching inequality)*$\vert \mathrm{spec} (B) \vert \, \mathcal{P}_B (A) \geq A$, where $\mathrm{spec} (B)$ denotes the set of eigenvalues of $B$.

The first two statements are easy to check. The earliest reference on the pinching inequality I can find is the classic book by Jacques Dixmier. A simple proof of the pinching inequality can be found in the textbook by Masahito Hayashi.

One main issue in matrix analysis is non-commutativity. The first statement in Theorem 1 hints that pinching can be an useful tool to deal with this issue. In the next section, the pinching trick is illustrated using the Golden-Thompson inequality as an example.

### A proof of the Golden-Thompson Inequality

The Golden-Thompson inequality says that

for any two Hermitian matrices $A$ and $B$. Obviously, if $A$ commutes with $B$, the Golden-Thompson inequality holds with an equality; however, in general one needs to take non-commutativity into consideration. Below we present a very elegant proof using the pinching trick from a recent paper by D. Sutter et al.

The key observation is that $\vert \mathrm{spec} ( A^{\otimes n} ) \vert$ does not grow rapidly with $n$ for any Hermitian matrix $A$.

**Lemma 1.** One has $\vert \mathrm{spec} ( A^{\otimes n} ) \vert = O ( \mathsf{poly} (n) )$ for any Hermitian matrix $A$.

*Proof (Golden-Thompson inequality).*

Let $X$ and $Y$ be two positive definite matrices. Then one can write

for any positive integer $n$. By the pinching inequality, one has

By the first two statements in Theorem 1 and Lemma 1, one has

Then one obtains the Golden-Thompson Inequality by letting $n \to \infty$.