Challenges
Datasets
Workspaces
Papers
Discussions
Leaderboard
Log inSign up
Challenges
Datasets
Workspaces
Papers
Discussions
Leaderboard
Blog
Job Board
Q3AS

© 2026 Aqora Quantum S.A.S.

TermsPrivacyLegal Notice

arXiv:2610.39598

Sharp Tilting Inequalities

Hao-Chung Cheng, Po-Chieh Liu

math.FAmath-phquant-phSubmitted Oct 29, 2026• Announced Oct 30, 2026• 0 citations
arXivHTMLPDF

Abstract

For positive semidefinite matrices $A,B$ and $1/2\leα\le1$, we establish $\operatorname{Tr}\!\left[A\,D\log(A^α+B^α)(B^α)\right] \le s^s(1-s)^{1-s}\operatorname{Tr}[A^αB^{1-α}]$, where $s=(1-α)/α$, with the optimal coefficient. The proof uses a superoperator representation of the directional derivative of the logarithm and an Araki-type estimate for the derivative of a fractional power. Both this inequality and its symmetric-quotient counterpart extend to positive trace-class $A$ and positive, possibly unbounded, self-adjoint $B$ on a separable Hilbert space.

AI Summary

The paper asks for the optimal constant in a noncommutative version of the scalar bound $x/(1+x)\le s^s(1-s)^{1-s}x^s$. For positive semidefinite matrices $A,B$, $1/2\le\alpha\le1$, and $s=(1-\alpha)/\alpha$, it proves $\operatorname{Tr}[A\,D\log[A^\alpha+B^\alpha](B^\alpha)]\le s^s(1-s)^{1-s}\operatorname{Tr}[A^\alpha B^{1-\alpha}]$. The coefficient is optimal for every $\alpha$, improving an earlier tilting inequality. For positive definite matrices and interior $\alpha$, equality holds exactly when $B^\alpha=((1-s)/s)A^\alpha$. The same optimal bound holds for the symmetric quotient $\operatorname{Tr}[A(A^\alpha+B^\alpha)^{-1/2}B^\alpha(A^\alpha+B^\alpha)^{-1/2}]$. The proof represents the logarithmic derivative as an average of inverse left–right multiplication superoperators, reducing both quotients to a functional-calculus application of the scalar inequality. An averaged fractional-power estimate supplies the remaining comparison with the mixed trace. Its proof uses an averaged derivative identity and an Araki-type power-derivative inequality established through Epstein concavity and log-majorization results. Both inequalities extend to positive trace-class $A$ and positive, possibly unbounded, self-adjoint $B$ on a separable Hilbert space, retaining the optimal constants. Singular denominators are interpreted on their supports; in infinite dimensions, the logarithmic expression is defined by an improper integral rather than asserted to be a Fréchet derivative. The mixed trace can be infinite even for trace-class $A$ and bounded $B$, making the bound vacuous in those cases. At $\alpha=1$, $B^0$ denotes its support projection.

No sign of AI-generated textIncremental

Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.

Related papers

Tripartite Bell inequality, random matrices and trilinear forms

Gilles Pisier

In this seminar report, we present in detail the proof of a recent result due to J. Briët and T. Vidick, improving an estimate in a 2008 paper by D. Pérez-Garc\'ıa, M. Wolf, C. Palazuelos, I. Villanueva, and M. Junge, estimating the growth of the deviation in the tripartite Bell inequality. The proof requires a delicate estimate of the norms of certain trilinear (or $d$-linear) forms on Hilbert space with coefficients in the second Gaussian Wiener chaos. Let $E^n_{\vee}$ (resp. $E^n_{\min}$) denote $ \ell_1^n \otimes \ell_1^n\otimes \ell_1^n$ equipped with the injective (resp. minimal) tensor norm. Here $ \ell_1^n$ is equipped with its maximal operator space structure. The Briët-Vidick method yields that the identity map $I_n$ satisfies (for some $c>0$) $\|I_n:\ E^n_{\vee}\to E^n_{\min}\|\ge c n^{1/4} (\log n)^{-3/2}.$ Let $S^n_2$ denote the (Hilbert) space of $n\times n$-matrices equipped with the Hilbert-Schmidt norm. While a lower bound closer to $n^{1/2} $ is still open, their method produces an interesting, asymptotically almost sharp, related estimate for the map $J_n:\ S^n_2\stackrel{\vee}{\otimes} S^n_2\stackrel{\vee}{\otimes}S^n_2 \to \ell_2^{n^3} \stackrel{\vee}{\otimes} \ell_2^{n^3} $ taking $e_{i,j}\otimes e_{k,l}\otimes e_{m,n}$ to $e_{[i,k,m],[j,l,n]}$.

arXiv:1203.2509•Mar 12, 2012•5 citations

Positive definite matrices with Hermitian blocks and their partial traces

Jean-Christophe Bourin, Eunyoung Lee, Minghua Lin

Let $H$ be a positive semi-definite matrix partitioned in $β\times β$ Hermitian blocks, $H=[A_{s,t}]$, $1\le s,t,\le β$. Then, for all symmetric norms, {equation*} \| H \| \le \| \sum_{s=1}^β A_{s,s} \|. {equation*} The proof uses a nice decomposition for positive matrices and unitary congruences with the generators of a Clifford algebra. A few corollaries are given, in particular the partial trace operation increases norms of separable states on a real Hilbert space, leading to a conjecture for usual complex Hilbert spaces.

arXiv:1208.6494•Aug 31, 2012•1 citation

An attainable Gill-Massar-type bound for spin-factor models

Koichi Yamagata

We determine the exact local precision limits for single-copy estimation of smooth multiparameter quantum statistical models contained in spin factors, a class of matrix Jordan algebras whose state spaces generalize the qubit Bloch ball. At any parameter point where the symmetric logarithmic derivative (SLD) Fisher information is positive definite, we characterize the entire attainable classical Fisher-information region over all finite-outcome positive-operator-valued measurements. After SLD normalization, this region consists exactly of the real symmetric positive semidefinite matrices with trace at most one, independently of the ambient Hilbert-space dimension. This yields a sharp weighted covariance bound for every positive definite weight, attained by an explicit locally unbiased estimator based on randomized spectral measurements of SLD directions. The proof combines a statistics-preserving positive projection onto the spin factor with the two-eigenvalue structure of its effects, revealing the Jordan-algebraic origin of the tradeoff. The result extends the qubit information tradeoff to models with more than three parameters. Since the optimal measurement depends on the unknown parameter, we simulate an adaptive scheme for a five-parameter model on a four-dimensional Hilbert space and observe performance close to the optimal local benchmark.

quant-pharXiv:2609.23020•Sep 22, 2026•0 citations

Some trace inequalities for exponential and logarithmic functions

Eric A. Carlen, Élliott H. Lieb

Consider a function F(X,Y) of pairs of positive matrices with values in the positive matrices such that whenever X and Y commute F(X,Y)=XpYq. Our first main result gives conditions on F such that Tr[X log(F(Z,Y))]≤Tr[X(p log X+q log Y)] for all X,Y,Z such that TrZ=TrX. (Note that Z is absent from the right side of the inequality.) We give several examples of functions F to which the theorem applies. Our theorem allows us to give simple proofs of the well-known logarithmic inequalities of Hiai and Petz and several new generalizations of them which involve three variables X,Y,Z instead of just X,Y alone. The investigation of these logarithmic inequalities is closely connected with three quantum relative entropy functionals: The standard Umegaki quantum relative entropy D(X∥Y)=Tr[X(log X− log Y]), and two others, the Donald relative entropy DD(X∥Y), and the Belavkin–Stasewski relative entropy DBS(X∥Y). They are known to satisfy DD(X∥Y)≤D(X∥Y)≤DBS(X∥Y). We prove that the Donald relative entropy provides the sharp upper bound, independent of Z on Tr[X log(F(Z,Y))] in a number of cases in which F(Z,Y) is homogeneous of degree 1 in Z and −1 in Y. We also investigate the Legendre transforms in X of DD(X∥Y) and DBS(X∥Y), and show how our results for these lead to new refinements of the Golden–Thompson inequality.

Bulletin of Mathematical Sciences•May 29, 2018•4 citations

A Sharp Rank-r Partial-Trace Inequality

Elazar Gershuni

Let $\Phi:M_d(\C)\to M_d(\C)$ be a completely positive map with a transpose-symmetric Kraus representation. Its Choi operator determines the rank-two constant \[ \beta_2(\Phi) =\frac12\sup_{\substack{\norm z=1\\ \SR(z)\le2}} \ip z{J(\Phi)z}. \] We prove that, for every $r\ge1$, \[ \Phi\circ\tau+(r-1)\beta_2(\Phi) \left(\Delta-\frac1r\id\right) \] is $r$-positive, where $\tau$ is transposition and $\Delta(X)=\Tr(X)I$. The proof turns the negative spectrum of a partially transposed rank-one operator into the edges of the complete graph $K_r$. Regrouping by vertices costs its degree $r-1$, rather than its $\binom r2$ edges. Transpose symmetry protects one direction and is exactly what produces the term $-\id/r$. For $\Phi=\Lambda_U\otimes\Lambda_V$, with $\Lambda_E(X)=\Tr(X)I-X^T$, a sharp double-skew singular-value estimate gives $\beta_2\le1$, with equality when both local dimensions are at least two. A map identity then yields, for every operator $C$ of rank at most $r$ on $U\otimes V$, \[ \norm{\Tr_U C}_2^2+\norm{\Tr_V C}_2^2 \le r\norm C_2^2+\frac1r\abs{\Tr C}^2, \] with dimension-uniform coefficients that are sharp whenever one local dimension is at least $r$ and the other is at least two. A second family, built from the edges of an arbitrary graph, has the same local constant and attains the universal coefficient $r-1$ whenever the graph contains an $r$-clique. The Choi lifting chain through the partial-trace and graph results, together with its sharpness witnesses, is formalized in Lean~4.

•Jul 26, 2026•0 citations

More on a trace inequality in quantum information theory

Naresh Kumar Sharma

It is known that for a completely positive and trace preserving (cptp) map ${\cal N}$, $\text{Tr}$ $\exp$$\{ \log σ$ $+$ ${\cal N}^\dagger [\log {\cal N}(ρ)$ $-\log {\cal N}(σ)] \}$ $\leqslant$ $\text{Tr}$ $ρ$ when $ρ$, $σ$, ${\cal N}(ρ)$, and ${\cal N}(σ)$ are strictly positive. We state and prove a relevant version of this inequality for the hitherto unaddressed case of these matrices being nonnegative. Our treatment also provides an alternate proof for the strictly positive case.

arXiv:1512.00226•Dec 1, 2015•2 citations
For positive semidefinite matrices A,BA,BA,B and 1/2≤α≤11/2\leα\le11/2≤α≤1, we establish Tr⁡ ⁣[A Dlog⁡(Aα+Bα)(Bα)]≤ss(1−s)1−sTr⁡[AαB1−α]\operatorname{Tr}\!\left[A\,D\log(A^α+B^α)(B^α)\right] \le s^s(1-s)^{1-s}\operatorname{Tr}[A^αB^{1-α}]Tr[ADlog(Aα+Bα)(Bα)]≤ss(1−s)1−sTr[AαB1−α], where s=(1−α)/αs=(1-α)/αs=(1−α)/α, with the optimal coefficient. The proof uses a superoperator representation of the directional derivative of the logarithm and an Araki-type estimate for the derivative of a fractional power. Both this inequality and its symmetric-quotient counterpart extend to positive trace-class AAA and positive, possibly unbounded, self-adjoint BBB on a separable Hilbert space.
The paper asks for the optimal constant in a noncommutative version of the scalar bound x/(1+x)≤ss(1−s)1−sxsx/(1+x)\le s^s(1-s)^{1-s}x^sx/(1+x)≤ss(1−s)1−sxs. For positive semidefinite matrices A,BA,BA,B, 1/2≤α≤11/2\le\alpha\le11/2≤α≤1, and s=(1−α)/αs=(1-\alpha)/\alphas=(1−α)/α, it proves Tr⁡[A Dlog⁡[Aα+Bα](Bα)]≤ss(1−s)1−sTr⁡[AαB1−α]\operatorname{Tr}[A\,D\log[A^\alpha+B^\alpha](B^\alpha)]\le s^s(1-s)^{1-s}\operatorname{Tr}[A^\alpha B^{1-\alpha}]Tr[ADlog[Aα+Bα](Bα)]≤ss(1−s)1−sTr[AαB1−α]. The coefficient is optimal for every α\alphaα, improving an earlier tilting inequality. For positive definite matrices and interior α\alphaα, equality holds exactly when Bα=((1−s)/s)AαB^\alpha=((1-s)/s)A^\alphaBα=((1−s)/s)Aα.
The same optimal bound holds for the symmetric quotient Tr⁡[A(Aα+Bα)−1/2Bα(Aα+Bα)−1/2]\operatorname{Tr}[A(A^\alpha+B^\alpha)^{-1/2}B^\alpha(A^\alpha+B^\alpha)^{-1/2}]Tr[A(Aα+Bα)−1/2Bα(Aα+Bα)−1/2]. The proof represents the logarithmic derivative as an average of inverse left–right multiplication superoperators, reducing both quotients to a functional-calculus application of the scalar inequality. An averaged fractional-power estimate supplies the remaining comparison with the mixed trace. Its proof uses an averaged derivative identity and an Araki-type power-derivative inequality established through Epstein concavity and log-majorization results.
Both inequalities extend to positive trace-class AAA and positive, possibly unbounded, self-adjoint BBB on a separable Hilbert space, retaining the optimal constants. Singular denominators are interpreted on their supports; in infinite dimensions, the logarithmic expression is defined by an improper integral rather than asserted to be a Fréchet derivative. The mixed trace can be infinite even for trace-class AAA and bounded BBB, making the bound vacuous in those cases. At α=1\alpha=1α=1, B0B^0B0 denotes its support projection.
In this seminar report, we present in detail the proof of a recent result due to J. Briët and T. Vidick, improving an estimate in a 2008 paper by D. Pérez-Garc\'ıa, M. Wolf, C. Palazuelos, I. Villanueva, and M. Junge, estimating the growth of the deviation in the tripartite Bell inequality. The proof requires a delicate estimate of the norms of certain trilinear (or ddd-linear) forms on Hilbert space with coefficients in the second Gaussian Wiener chaos. Let E∨nE^n_{\vee}E∨n​ (resp. Emin⁡nE^n_{\min}Eminn​) denote $ \ell_1^n \otimes \ell_1^n\otimes \ell_1^n$ equipped with the injective (resp. minimal) tensor norm. Here $ \ell_1^n$ is equipped with its maximal operator space structure. The Briët-Vidick method yields that the identity map InI_nIn​ satisfies (for some c>0c>0c>0) ∥In: E∨n→Emin⁡n∥≥cn1/4(log⁡n)−3/2.\|I_n:\ E^n_{\vee}\to E^n_{\min}\|\ge c n^{1/4} (\log n)^{-3/2}.∥In​: E∨n​→Eminn​∥≥cn1/4(logn)−3/2. Let S2nS^n_2S2n​ denote the (Hilbert) space of n×nn\times nn×n-matrices equipped with the Hilbert-Schmidt norm. While a lower bound closer to $n^{1/2} $ is still open, their method produces an interesting, asymptotically almost sharp, related estimate for the map $J_n:\ S^n_2\stackrel{\vee}{\otimes} S^n_2\stackrel{\vee}{\otimes}S^n_2 \to \ell_2^{n^3} \stackrel{\vee}{\otimes} \ell_2^{n^3} $ taking ei,j⊗ek,l⊗em,ne_{i,j}\otimes e_{k,l}\otimes e_{m,n}ei,j​⊗ek,l​⊗em,n​ to e[i,k,m],[j,l,n]e_{[i,k,m],[j,l,n]}e[i,k,m],[j,l,n]​.
Let HHH be a positive semi-definite matrix partitioned in β×ββ\times ββ×β Hermitian blocks, H=[As,t]H=[A_{s,t}]H=[As,t​], 1≤s,t,≤β1\le s,t,\le β1≤s,t,≤β. Then, for all symmetric norms, {equation*} \| H \| \le \| \sum_{s=1}^β A_{s,s} \|. {equation*} The proof uses a nice decomposition for positive matrices and unitary congruences with the generators of a Clifford algebra. A few corollaries are given, in particular the partial trace operation increases norms of separable states on a real Hilbert space, leading to a conjecture for usual complex Hilbert spaces.
We determine the exact local precision limits for single-copy estimation of smooth multiparameter quantum statistical models contained in spin factors, a class of matrix Jordan algebras whose state spaces generalize the qubit Bloch ball. At any parameter point where the symmetric logarithmic derivative (SLD) Fisher information is positive definite, we characterize the entire attainable classical Fisher-information region over all finite-outcome positive-operator-valued measurements. After SLD normalization, this region consists exactly of the real symmetric positive semidefinite matrices with trace at most one, independently of the ambient Hilbert-space dimension. This yields a sharp weighted covariance bound for every positive definite weight, attained by an explicit locally unbiased estimator based on randomized spectral measurements of SLD directions. The proof combines a statistics-preserving positive projection onto the spin factor with the two-eigenvalue structure of its effects, revealing the Jordan-algebraic origin of the tradeoff. The result extends the qubit information tradeoff to models with more than three parameters. Since the optimal measurement depends on the unknown parameter, we simulate an adaptive scheme for a five-parameter model on a four-dimensional Hilbert space and observe performance close to the optimal local benchmark.
Consider a function F(X,Y) of pairs of positive matrices with values in the positive matrices such that whenever X and Y commute F(X,Y)=XpYq. Our first main result gives conditions on F such that Tr[X log(F(Z,Y))]≤Tr[X(p log X+q log Y)] for all X,Y,Z such that TrZ=TrX. (Note that Z is absent from the right side of the inequality.) We give several examples of functions F to which the theorem applies. Our theorem allows us to give simple proofs of the well-known logarithmic inequalities of Hiai and Petz and several new generalizations of them which involve three variables X,Y,Z instead of just X,Y alone. The investigation of these logarithmic inequalities is closely connected with three quantum relative entropy functionals: The standard Umegaki quantum relative entropy D(X∥Y)=Tr[X(log X− log Y]), and two others, the Donald relative entropy DD(X∥Y), and the Belavkin–Stasewski relative entropy DBS(X∥Y). They are known to satisfy DD(X∥Y)≤D(X∥Y)≤DBS(X∥Y). We prove that the Donald relative entropy provides the sharp upper bound, independent of Z on Tr[X log(F(Z,Y))] in a number of cases in which F(Z,Y) is homogeneous of degree 1 in Z and −1 in Y. We also investigate the Legendre transforms in X of DD(X∥Y) and DBS(X∥Y), and show how our results for these lead to new refinements of the Golden–Thompson inequality.
Let $\Phi:M_d(\C)\to M_d(\C)$ be a completely positive map with a transpose-symmetric Kraus representation. Its Choi operator determines the rank-two constant \[ \beta_2(\Phi) =\frac12\sup_{\substack{\norm z=1\\ \SR(z)\le2}} \ip z{J(\Phi)z}. \] We prove that, for every r≥1r\ge1r≥1, \[ \Phi\circ\tau+(r-1)\beta_2(\Phi) \left(\Delta-\frac1r\id\right) \] is rrr-positive, where τ\tauτ is transposition and $\Delta(X)=\Tr(X)I$. The proof turns the negative spectrum of a partially transposed rank-one operator into the edges of the complete graph KrK_rKr​. Regrouping by vertices costs its degree r−1r-1r−1, rather than its (r2)\binom r2(2r​) edges. Transpose symmetry protects one direction and is exactly what produces the term $-\id/r$. For Φ=ΛU⊗ΛV\Phi=\Lambda_U\otimes\Lambda_VΦ=ΛU​⊗ΛV​, with $\Lambda_E(X)=\Tr(X)I-X^T$, a sharp double-skew singular-value estimate gives β2≤1\beta_2\le1β2​≤1, with equality when both local dimensions are at least two. A map identity then yields, for every operator CCC of rank at most rrr on U⊗VU\otimes VU⊗V, \[ \norm{\Tr_U C}_2^2+\norm{\Tr_V C}_2^2 \le r\norm C_2^2+\frac1r\abs{\Tr C}^2, \] with dimension-uniform coefficients that are sharp whenever one local dimension is at least rrr and the other is at least two. A second family, built from the edges of an arbitrary graph, has the same local constant and attains the universal coefficient r−1r-1r−1 whenever the graph contains an rrr-clique. The Choi lifting chain through the partial-trace and graph results, together with its sharpness witnesses, is formalized in Lean~4.
It is known that for a completely positive and trace preserving (cptp) map N{\cal N}N, Tr\text{Tr}Tr exp⁡\expexp{log⁡σ\{ \log σ{logσ +++ N†[log⁡N(ρ){\cal N}^\dagger [\log {\cal N}(ρ)N†[logN(ρ) −log⁡N(σ)]}-\log {\cal N}(σ)] \}−logN(σ)]} ⩽\leqslant⩽ Tr\text{Tr}Tr ρρρ when ρρρ, σσσ, N(ρ){\cal N}(ρ)N(ρ), and N(σ){\cal N}(σ)N(σ) are strictly positive. We state and prove a relevant version of this inequality for the hitherto unaddressed case of these matrices being nonnegative. Our treatment also provides an alternate proof for the strictly positive case.