Trace Inequalities for Completely Monotone Functions

• Trace inequalities for completely monotone functions are rigorous extensions of classical matrix bounds, utilizing Bernstein integral representations and scalar concavity.
• They yield refined power inequalities and norm-compression results that enhance classical McCarthy and Araki–Lieb–Thirring bounds in operator analysis.
• The approach unifies exponential kernel techniques with spectral projector methods, providing tighter quantitative controls in noncommutative settings.

Trace inequalities for completely monotone functions constitute a rigorous extension and refinement of well-known matrix trace inequalities for operator functions, particularly for positive semidefinite matrices. Building on integral representation theorems by Bernstein and leveraging advanced scalar concavity techniques, these results yield general matrix inequalities for wide classes of functions, notably including the completely monotone and Bernstein function families. Applications encompass sharpened power trace inequalities and new forms of norm-compression for block matrices, with implications for quantum information and matrix analysis (Audenaert, 2011).

1. Completely Monotone and Bernstein Function Classes

A function $f:(0,\infty)\to\mathbb{R}$ is completely monotone if $f\in C^\infty(0,\infty)$, $f(x)\geq0$, and $(-1)^n f^{(n)}(x)\geq0$ for all $n\geq1$, $x>0$. Bernstein’s theorem provides a Laplace transform representation:

$f(x)=a+\int_0^\infty e^{-xt}\,\mu(dt),$

with $a=f(0^+)\geq0$ and $\mu$ a positive measure. The “bare” completely monotone functions (denoted $f\in{\rm CM}_0$) are those with $a=0$.

Bernstein functions satisfy $f\in C^\infty(0,\infty)$, $f(x)\geq0$, and $(-1)^{n-1}f^{(n)}(x)\geq0$ for all $n\geq1$. Their Lévy–Khintchine representation is

$f(x)=ax+b+\int_{0}^{\infty}(1-e^{-xt})\,\nu(dt),$

with $a,b\geq0$ and $\nu$ a positive measure. “Bare” Bernstein functions ($f\in{\rm BF}_0$) are those with $a=b=0$. The hierarchy of Bernstein integrals is generated recursively: for integer $k\geq1$, ${\rm BF}_k$ contains functions defined as $F(x)=\int_0^x F_{k-1}(t)\,dt$ with $F_{k-1}\in{\rm BF}_{k-1}$. Explicit integral representations relate these to sums or differences of exponential functions.

Examples:

• ${\rm BF}_0$: $f(x)=1-e^{-xt}$ (concave, nonnegative, $f(0)=0$)
• ${\rm BF}_1$: $f(x)=e^{-xt}-(1-xt)$ (convex, nonnegative, $f(0)=0$)
• ${\rm BF}_2$: $f(x)=1-e^{-xt}+(xt)-(xt)^2/2$ (convex, nonnegative)

2. Scalar Additivity and Concavity Properties

For $g:[0,\infty)\to\mathbb{R}$,

• If $g\in{\rm CM}_0\cup{\rm BF}_0$, $g$ is subadditive: $g(x+y)\leq g(x)+g(y)$.
• If $g\in{\rm BF}_k$ for $k\geq1$, $g$ is superadditive: $g(x+y)\geq g(x)+g(y)$.

The function $\varphi(x)=1-e^{-x}$ exhibits geometric concavity:

$\varphi(\sqrt{xy})\geq\sqrt{\varphi(x)\varphi(y)},$

equivalently, $\log(1-e^{-e^u})$ is concave in $u$.

A refined two-point scalar inequality for exponentials is

$$e^{-(a+b)} - e^{-a} - e^{-b} \leq e^{-2\sqrt{ab}}-2e^{-\sqrt{ab}}$$

for $a,b\geq0$, with equality if and only if the function is a quadratic polynomial.

Combining these results provides a “refined one-point” inequality:

• For $g\in{\rm CM}_0\cup{\rm BF}_1$,

$$g(a+b)-g(a)-g(b)\leq g\bigl(2\sqrt{ab}\bigr)-2g\bigl(\sqrt{ab}\bigr)$$

• For $g\in{\rm BF}_0\cup{\rm BF}_2$, the reverse inequality holds.

3. Main Matrix Trace Inequality

Let $A,B\in M_d$ be positive semidefinite with spectral decompositions $A=\sum_k a_k P_k$, $B=\sum_\ell b_\ell Q_\ell$, where $P_k$, $Q_\ell$ are orthogonal projectors and $a_k, b_\ell\geq0$. For $g:[0,\infty)\to\mathbb{R}$:

• If $g\in{\rm CM}_0\cup{\rm BF}_1$,

$$\operatorname{Tr}\bigl[g(A+B)-g(A)-g(B)\bigr] \leq \sum_{k,\ell}\bigl[g(a_k+b_\ell)-g(a_k)-g(b_\ell)\bigr]\operatorname{Tr}(P_k Q_\ell)$$

$$\leq \sum_{k,\ell}\bigl[g(2\sqrt{a_kb_\ell})-2g(\sqrt{a_kb_\ell})\bigr]\operatorname{Tr}(P_k Q_\ell)$$

• If $g\in{\rm BF}_0\cup{\rm BF}_2$, the above chain of inequalities holds in the reverse sense.

Equality obtains for all $A,B$ if and only if $g(x)=1$, $x$, or $x^2$.

The proof utilizes integral representations for $g$, reduction to exponential kernels, the Golden–Thompson inequality ($$\operatorname{Tr}e^{-t(A+B)}\leq\operatorname{Tr}(e^{-tA}e^{-tB})$$), the scalar refined two-point exponential bound, and interchange of sum/integral with spectral projectors.

4. Refined Power Function Trace Inequalities

For the power function $f(x)=x^q$, classical McCarthy inequalities state:

• $$\operatorname{Tr}(A+B)^q\leq\operatorname{Tr}A^q+\operatorname{Tr}B^q$$ for $0<q\leq1$
• $$\operatorname{Tr}(A+B)^q\geq\operatorname{Tr}A^q+\operatorname{Tr}B^q$$ for $q\geq1$

Audenaert’s refinement introduces a sharp two-point correction:

Let $A,B\geq0$ and $q\in\mathbb{R}$:

• For $0<q\leq1$ or $2\leq q\leq3$,

$$\operatorname{Tr}(A+B)^q-\operatorname{Tr}A^q-\operatorname{Tr}B^q \geq (2^q-2)\operatorname{Tr}(A^{q/2}B^{q/2})$$

• For $q<0$ or (for $A,B>0$) $1\leq q\leq2$, the inequality reverses.

Connections with the Araki–Lieb–Thirring inequality provide the further variant:

$$\operatorname{Tr}(A+B)^q-\operatorname{Tr}A^q-\operatorname{Tr}B^q \geq (2^q-2)\operatorname{Tr}\Bigl((A^{1/2}BA^{1/2})^{q/2}\Bigr)$$

in the stated ranges.

For $2\leq q\leq3$, this result sharpens the superadditivity of the classical McCarthy inequality; for $0<q\leq1$ or $1\leq q\leq2$, it supplies complementary lower/upper bounds on the trace deviation.

5. Integral Representations and Proof Techniques

The proof methods employ Laplace and Lévy–Khintchine-type integral representations, reducing matrix trace inequalities for broad function classes $g$ to inequalities for the exponential kernel $g(x)=e^{-xt}$. The Golden–Thompson inequality connects the trace of the sum to traces of operator products. The scalar “geometric concavity” two-point inequality for $e^{-x}$ is then used to analyze the deviation $\operatorname{Tr}[e^{-t(A+B)}-e^{-tA}-e^{-tB}]$, ultimately providing the precise additional term at the spectral level. The entire argument leverages commutativity when present, but the bounds remain valid in the non-commutative case due to the spectral decomposition and projector structure.

6. Applications to Operator Means and Norm-Compression

Two principal applications arise:

• $p$-Power Means for Operators: By replacing $q$ with $1/p$ and considering $A^{1/q}$, $B^{1/q}$, the inequalities improve previously established bounds for $p$-power means $\operatorname{Tr}\bigl((A^p+B^p)^{1/p}\bigr)$ in certain parameter regions.
• Norm-Compression Inequality: Let $$M=\begin{pmatrix} B & C \ C^* & D \end{pmatrix}\geq 0$$ be a partitioned positive semidefinite matrix, and denote the Schatten $q$-norm $\|X\|_q = (\operatorname{Tr}|X|^q)^{1/q}$. For $1\leq q\leq2$ (and, by extension, $0<q\leq1$ or $2\leq q\leq3$ in the reversed sense),

$$\|M\|_q^q\leq (2^q-2)\|C\|_q^q + \|B\|_q^q + \|D\|_q^q$$

This generalizes earlier results by extending the range of $q$ and provides a short proof via the trace inequalities for completely monotone functions. The norm-compression bound supplies a useful tool in matrix analysis and quantum information theory.

7. Context and Significance

The trace inequalities for completely monotone and Bernstein functions unify and extend several celebrated results (such as McCarthy-type inequalities) by characterizing the range of possible defects between the trace of an operator function of a sum and the sum of the traces of the function applied to the summands. The “two-point” enhancement yields tighter quantitative control in noncommutative scenarios. These inequalities are notable for their reliance on integral transforms and their general applicability to a wide variety of operator functions beyond the usual power functions, thus enabling new bounds and comparisons for matrix means and operator norms. The methods and inequalities have independent standing in matrix analysis and are particularly relevant in quantum information contexts where such operator functionals frequently arise (Audenaert, 2011).