1
Introduction
The question
Let φ \varphi φ be a positive linear functional on a C∗ ^* ∗ -algebra. Gardner [125 ] proved that φ \varphi φ is a trace, φ ( X Y ) = φ ( Y X ) \varphi(XY)=\varphi(YX) φ ( X Y ) = φ ( Y X ) , if and only if
∣ φ ( X ) ∣ ≤ φ ( ∣ X ∣ ) for all X . \begin{equation}
|\varphi(X)|\le\varphi(|X|)\qquad\text{for all }X . \tag*{(1)}
\end{equation} ∣ φ ( X ) ∣ ≤ φ ( ∣ X ∣ ) for all X . (1)
Since then many inequalities that hold for every trace have been shown to hold only for traces: the triangle inequality for φ ( ∣ ⋅ ∣ ) \varphi(|\cdot|) φ ( ∣ ⋅ ∣ ) , the inequalities of Petz–Zemánek, monotonicity of A ↦ φ ( A 2 ) A\mapsto\varphi(A^2) A ↦ φ ( A 2 ) , Young’s, Golden–Thompson’s and Peierls–Bogoliubov’s inequalities, the anticommutator inequality of Bikchentaev and Abed, and a family of inequalities in which only projections enter. Centrality of a positive element admits similar characterisations (Novikov and Tikhonov [152 ] ). § 2 reviews these results. Each of them answers a yes/no question. For a functional that is not a trace the inequality fails for some operators; but how strongly, and does the size of the failure measure how far the functional is from being a trace? This paper answers these questions for 38 characterisations of traces and for the characterisations of central elements of [152 ] .
The known proofs are qualitative. They reduce to M 2 ( C ) M_2(\mathbb{C}) M 2 ( C ) with φ = T r ( S ⋅ ) \varphi=\mathop{\mathrm{Tr}}(S\,\cdot\,) φ = Tr ( S ⋅ ) , S = d i a g ( 1 / 2 + s , 1 / 2 − s ) S=\mathop{\mathrm{diag}}(1/2+s,1/2-s) S = diag ( 1/2 + s , 1/2 − s ) , and exhibit a one-parameter family of test operators for which the inequality fails at first order in the parameter as soon as s ≠ 0 s\ne0 s = 0 ; no constant results. For Gardner’s inequality on M n ( C ) M_n(\mathbb{C}) M n ( C ) the sharp constant is contained in an inequality of Bourin [113 Cor. 1.4, Rem. 1.9] , which is not formulated as a trace characterisation (Remark 5.2 ), and the mixed Schwarz inequality of Kato [139 ] , in the form ∣ φ ( x ) ∣ 2 ≤ φ ( ∣ x ∣ ) φ ( ∣ x ∗ ∣ ) |\varphi(x)|^2\le\varphi(|x|)\varphi(|x^*|) ∣ φ ( x ) ∣ 2 ≤ φ ( ∣ x ∣ ) φ ( ∣ x ∗ ∣ ) (see [120 ] ), gives the non-sharp bound ∣ φ ( x ) ∣ ≤ κ ( ρ ) 1 / 2 φ ( ∣ x ∣ ) |\varphi(x)|\le\kappa(\rho)^{1/2}\varphi(|x|) ∣ φ ( x ) ∣ ≤ κ ( ρ ) 1/2 φ ( ∣ x ∣ ) for φ = T r ( ρ ⋅ ) \varphi=\mathop{\mathrm{Tr}}(\rho\,\cdot\,) φ = Tr ( ρ ⋅ ) , where κ ( ρ ) \kappa(\rho) κ ( ρ ) is the condition number. Apart from this we are not aware of quantitative versions of these characterisations in the literature.
Contrast and noncentrality
On M n M_n M n every positive functional is φ ρ = T r ( ρ ⋅ ) \varphi_\rho=\mathop{\mathrm{Tr}}(\rho\,\cdot\,) φ ρ = Tr ( ρ ⋅ ) with a density ρ ≥ 0 \rho\ge0 ρ ≥ 0 , and φ ρ \varphi_\rho φ ρ is a trace if and only if ρ \rho ρ is a multiple of 1 1 1 . The natural scale-invariant distance of ρ \rho ρ from the scalars is its Michelson contrast
δ = O M C ( ρ ) = λ max ( ρ ) − λ min ( ρ ) λ max ( ρ ) + λ min ( ρ ) = κ ( ρ ) − 1 κ ( ρ ) + 1 ∈ [ 0 , 1 ] , \delta=\mathop{\mathrm{OMC}}(\rho)=\frac{\lambda_{\max}(\rho)-\lambda_{\min}(\rho)}{\lambda_{\max}(\rho)+\lambda_{\min}(\rho)}
=\frac{\kappa(\rho)-1}{\kappa(\rho)+1}\in[0,1], δ = OMC ( ρ ) = λ m a x ( ρ ) + λ m i n ( ρ ) λ m a x ( ρ ) − λ m i n ( ρ ) = κ ( ρ ) + 1 κ ( ρ ) − 1 ∈ [ 0 , 1 ] ,
which was studied for operators in [85 ] , together with its logarithmic form O L C ( ρ ) = 1 2 ln κ ( ρ ) = artanh δ \mathop{\mathrm{OLC}}(\rho)=\frac12\ln\kappa(\rho)=\operatorname{artanh}\delta OLC ( ρ ) = 2 1 ln κ ( ρ ) = artanh δ [150 ] . Gibbs states show that this is a physically meaningful scale. For a Hamiltonian H H H with extreme eigenvalues E min ≤ E max E_{\min}\le E_{\max} E m i n ≤ E m a x and the inverse temperature β \beta β , the density ρ = e − β H / T r e − β H \rho=e^{-\beta H}/\mathop{\mathrm{Tr}}e^{-\beta H} ρ = e − β H / Tr e − β H has
O M C ( ρ ) = tanh β ( E max − E min ) 2 , O L C ( ρ ) = β ( E max − E min ) 2 , \mathop{\mathrm{OMC}}(\rho)=\tanh\frac{\beta(E_{\max}-E_{\min})}2 ,\qquad \mathop{\mathrm{OLC}}(\rho)=\frac{\beta(E_{\max}-E_{\min})}2, OMC ( ρ ) = tanh 2 β ( E m a x − E m i n ) , OLC ( ρ ) = 2 β ( E m a x − E m i n ) ,
so δ → 0 \delta\to0 δ → 0 at infinite temperature, where the Gibbs state becomes the normalised trace, and δ → 1 \delta\to1 δ → 1 at zero temperature. Every bound in terms of δ \delta δ is therefore a law in the temperature: for Gardner’s inequality we shall find the largest relative violation cosh ( β ( E max − E min ) / 2 ) − 1 \cosh(\beta(E_{\max}-E_{\min})/2)-1 cosh ( β ( E m a x − E m i n ) /2 ) − 1 . In the language of modular theory, φ ρ \varphi_\rho φ ρ (for invertible ρ \rho ρ ) is a trace if and only if its modular group σ t ( X ) = ρ i t X ρ − i t \sigma_t(X)=\rho^{it}X\rho^{-it} σ t ( X ) = ρ i t X ρ − i t is trivial; the derivation X ↦ [ log ρ , X ] X\mapsto[\log \rho,X] X ↦ [ log ρ , X ] that generates it has the spectrum { log w i − log w j } \{\log w_i-\log w_j\} { log w i − log w j } , where w i w_i w i are the eigenvalues of ρ \rho ρ , whose largest point is log κ ( ρ ) = 2 O L C ( ρ ) \log\kappa(\rho)=2\mathop{\mathrm{OLC}}(\rho) log κ ( ρ ) = 2 OLC ( ρ ) . The contrast thus measures how far the modular dynamics is from being trivial.
Beyond factors the contrast is the wrong parameter. Let M \mathcal{M} M be a finite von Neumann algebra with a faithful normal tracial state τ \tau τ and φ = τ ( ρ ⋅ ) \varphi=\tau(\rho\,\cdot\,) φ = τ ( ρ ⋅ ) with ρ ∈ M \rho\in\mathcal{M} ρ ∈ M positive and invertible. If ρ \rho ρ is central, then φ \varphi φ is a trace and nothing is violated, however large O M C ( ρ ) \mathop{\mathrm{OMC}}(\rho) OMC ( ρ ) is; already on M 2 ⊕ M 2 M_2\oplus M_2 M 2 ⊕ M 2 the density 1 ⊕ 4 1\oplus 4 1 ⊕ 4 has contrast 3 / 5 3/5 3/5 and gives a trace. The relevant parameter is the noncentrality of ρ \rho ρ , the distance of ρ \rho ρ from the centre Z \mathcal{Z} Z rather than from the scalars. It is defined through the central envelopes of ρ \rho ρ : ⌈ ρ ⌉ \lceil \rho\rceil ⌈ ρ ⌉ is the least central element ≥ ρ \ge\rho ≥ ρ and ⌊ ρ ⌋ \lfloor \rho\rfloor ⌊ ρ ⌋ the greatest central element ≤ ρ \le\rho ≤ ρ (they exist by Lemma 4.1 ; they are the largest and the smallest elements of Halpern’s central spectrum [131 ] ), and
κ Z ( ρ ) = ∥ ⌈ ρ ⌉ ⌊ ρ ⌋ − 1 ∥ , L N C ( ρ ) = 1 2 ln κ Z ( ρ ) , δ Z = N C ( ρ ) = tanh L N C ( ρ ) = κ Z ( ρ ) − 1 κ Z ( ρ ) + 1 , \kappa_\mathcal{Z}(\rho)=\bigl\|\lceil \rho\rceil\,\lfloor \rho\rfloor^{-1}\bigr\|,\qquad
\mathop{\mathrm{LNC}}(\rho)=\tfrac12\ln\kappa_\mathcal{Z}(\rho),\qquad \delta_\mathcal{Z}=\mathop{\mathrm{NC}}(\rho)=\tanh\mathop{\mathrm{LNC}}(\rho)=\frac{\kappa_\mathcal{Z}(\rho)-1}{\kappa_\mathcal{Z}(\rho)+1}, κ Z ( ρ ) = ⌈ ρ ⌉ ⌊ ρ ⌋ − 1 , LNC ( ρ ) = 2 1 ln κ Z ( ρ ) , δ Z = NC ( ρ ) = tanh LNC ( ρ ) = κ Z ( ρ ) + 1 κ Z ( ρ ) − 1 ,
in analogy with O M C = tanh O L C \mathop{\mathrm{OMC}}=\tanh\mathop{\mathrm{OLC}} OMC = tanh OLC . Thus κ Z ( ρ ) = 1 \kappa_\mathcal{Z}(\rho)=1 κ Z ( ρ ) = 1 if and only if ρ \rho ρ is central, κ Z ( ρ ) ≤ κ ( ρ ) \kappa_\mathcal{Z}(\rho)\le\kappa(\rho) κ Z ( ρ ) ≤ κ ( ρ ) , κ Z = κ \kappa_\mathcal{Z}=\kappa κ Z = κ in factors, and κ Z ( ρ ) = min z κ ( z ρ ) \kappa_\mathcal{Z}(\rho)=\min_z\kappa(z\rho) κ Z ( ρ ) = min z κ ( z ρ ) over positive invertible central z z z (§ 4 ). On a direct sum ⨁ c M k c \bigoplus_cM_{k_c} ⨁ c M k c the noncentrality of ρ = ⨁ c ρ c \rho=\bigoplus_c\rho_c ρ = ⨁ c ρ c is the largest contrast inside the blocks, max c O M C ( ρ c ) \max_c\mathop{\mathrm{OMC}}(\rho_c) max c OMC ( ρ c ) .
The constants are expressed through the Kantorovich constant
G = G ( ρ ) = cosh O L C ( ρ ) = 1 + κ ( ρ ) 2 κ ( ρ ) = 1 1 − δ 2 = sec ψ , G Z = cosh L N C ( ρ ) = 1 1 − δ Z 2 = sec ψ Z , G=G(\rho)=\cosh\mathop{\mathrm{OLC}}(\rho)=\frac{1+\kappa(\rho)}{2\sqrt{\kappa(\rho)}}=\frac1{\sqrt{1-\delta^2}}=\sec\psi,
\qquad G_\mathcal{Z}=\cosh\mathop{\mathrm{LNC}}(\rho)=\frac1{\sqrt{1-\delta_\mathcal{Z}^2}}=\sec\psi_\mathcal{Z}, G = G ( ρ ) = cosh OLC ( ρ ) = 2 κ ( ρ ) 1 + κ ( ρ ) = 1 − δ 2 1 = sec ψ , G Z = cosh LNC ( ρ ) = 1 − δ Z 2 1 = sec ψ Z ,
where ψ = arcsin δ \psi=\arcsin\delta ψ = arcsin δ is the angle of ρ \rho ρ in the sense of Krein and Gustafson (the largest angle between a vector x x x and its image ρ x \rho x ρ x ) and ψ Z = arcsin δ Z \psi_\mathcal{Z}=\arcsin\delta_\mathcal{Z} ψ Z = arcsin δ Z ; G 2 G^2 G 2 is the best constant in Kantorovich’s inequality. Table 2 in § 3 collects the notation.
Envelopes and violation constants
In [85 ] we plotted the violation of Gardner’s inequality against O M C ( ρ ) \mathop{\mathrm{OMC}}(\rho) OMC ( ρ ) for random real 2 × 2 2\times2 2 × 2 and 3 × 3 3\times3 3 × 3 matrices [85 Sect. 5.1, Fig. 1] and asked whether such violations are bounded by functions of the contrasts [85 Sect. 7] . To make the question precise, write an inequality as L ≤ R \mathcal{L}\le\mathcal{R} L ≤ R , where L \mathcal{L} L and R \mathcal{R} R depend on φ \varphi φ and on operator variables. The violation constant c ( ρ ) c(\rho) c ( ρ ) is the best constant c c c in the weakened inequality L ≤ c R \mathcal{L}\le c\,\mathcal{R} L ≤ c R (for instance ∣ φ ( X ) ∣ ≤ c φ ( ∣ X ∣ ) |\varphi(X)|\le c\,\varphi(|X|) ∣ φ ( X ) ∣ ≤ c φ ( ∣ X ∣ ) ); it equals 1 1 1 exactly when φ \varphi φ is a trace. On M n M_n M n we also fix a normalisation N > 0 \mathcal{N}>0 N > 0 that makes the quotient ( L − R ) / N (\mathcal{L}-\mathcal{R})/\mathcal{N} ( L − R ) / N invariant under the scaling of ρ \rho ρ and of the variables (mostly N = R \mathcal{N}=\mathcal{R} N = R , and then ( L − R ) / N = L / R − 1 (\mathcal{L}-\mathcal{R})/\mathcal{N}=\mathcal{L}/\mathcal{R}-1 ( L − R ) / N = L / R − 1 ), and call
V n ( δ ) = sup { L − R N : ρ ≥ 0 in M n , O M C ( ρ ) = δ , all operator variables free } V_n(\delta)=\sup\Bigl\{\frac{\mathcal{L}-\mathcal{R}}{\mathcal{N}}:\ \rho\ge0\ \text{in}\ M_n,\ \mathop{\mathrm{OMC}}(\rho)=\delta,\ \text{all
operator variables free}\Bigr\} V n ( δ ) = sup { N L − R : ρ ≥ 0 in M n , OMC ( ρ ) = δ , all operator variables free }
the violation envelope. By the characterisation theorems V n ( δ ) > 0 V_n(\delta)>0 V n ( δ ) > 0 for δ > 0 \delta>0 δ > 0 , and V n ( 0 ) = 0 V_n(0)=0 V n ( 0 ) = 0 . A violation needs more than a non-tracial functional: if the operators are scalars, nothing can go wrong. So we also prescribe the contrast δ ′ \delta' δ ′ of the operator variables (of their moduli, for operators that are not positive) and write V n ( δ , δ ′ ) V_n(\delta,\delta') V n ( δ , δ ′ ) .
Gardner’s inequality with two contrasts
For Gardner’s inequality both contrasts can be prescribed, and the answer is a function of their product alone.
Theorem A 1. Let ρ > 0 \rho>0 ρ > 0 and X X X be in M n M_n M n , n ≥ 2 n\ge2 n ≥ 2 , with O M C ( ρ ) = δ \mathop{\mathrm{OMC}}(\rho)=\delta OMC ( ρ ) = δ and O M C ( ∣ X ∣ ) = δ ′ \mathop{\mathrm{OMC}}(|X|)=\delta' OMC ( ∣ X ∣ ) = δ ′ , δ δ ′ < 1 \delta\delta'<1 δ δ ′ < 1 . Then ∣ φ ρ ( X ) ∣ ≤ 1 1 − δ 2 δ ′ 2 φ ρ ( ∣ X ∣ ) , |\varphi_\rho(X)|\le\frac{1}{\sqrt{1-\delta^2\delta'^2}}\,\varphi_\rho(|X|), ∣ φ ρ ( X ) ∣ ≤ 1 − δ 2 δ ′2 1 φ ρ ( ∣ X ∣ ) , and for even n n n the constant is attained, so V n ( δ , δ ′ ) = ( 1 − δ 2 δ ′ 2 ) − 1 / 2 − 1 V_n(\delta,\delta')=(1-\delta^2\delta'^2)^{-1/2}-1 V n ( δ , δ ′ ) = ( 1 − δ 2 δ ′2 ) − 1/2 − 1 . For odd n n n the envelope is given by an explicit one-dimensional maximum; it is smaller when δ ′ < 1 \delta'<1 δ ′ < 1 and tends to the even value as n → ∞ n\to\infty n → ∞ . The same upper bound holds for the triangle inequality φ ( ∣ X + Y ∣ ) ≤ φ ( ∣ X ∣ ) + φ ( ∣ Y ∣ ) \varphi(|X+Y|)\le\varphi(|X|)+\varphi(|Y|) φ ( ∣ X + Y ∣ ) ≤ φ ( ∣ X ∣ ) + φ ( ∣ Y ∣ ) and for φ ( A + A ∗ ) ≤ 2 φ ( ∣ A ∗ ∣ ) \varphi(A+A^*)\le2\varphi(|A^*|) φ ( A + A ∗ ) ≤ 2 φ ( ∣ A ∗ ∣ ) (Theorem 6.1 ).
With the operator free (δ ′ = 1 \delta'=1 δ ′ = 1 ) the constant is G = sec ψ G=\sec\psi G = sec ψ , and in a finite von Neumann algebra it is G Z = cosh L N C ( ρ ) G_\mathcal{Z}=\cosh\mathop{\mathrm{LNC}}(\rho) G Z = cosh LNC ( ρ ) (Theorem 5.1 ).
Figure 1. (a) The angle of a positive operator: A x / λ ˉ ( A ) Ax/\bar\lambda(A) A x / λ ˉ ( A ) lies in the ball of radius O M C ( A ) ∣ x ∣ \mathop{\mathrm{OMC}}(A)|x| OMC ( A ) ∣ x ∣ about x x x , so ∠ ( x , A x ) ≤ ψ A = arcsin O M C ( A ) \angle(x,Ax)\le\psi_A=\arcsin\mathop{\mathrm{OMC}}(A) ∠ ( x , A x ) ≤ ψ A = arcsin OMC ( A ) . (b) For a Jordan product the angles add; if ψ A + ψ B < π / 2 \psi_A+\psi_B<\pi/2 ψ A + ψ B < π /2 , then A B x ABx A B x and B A x BAx B A x stay in a cone around x x x that lies in the half-space of x x x , and A B + B A > 0 AB+BA>0 A B + B A > 0 (Theorem 9.2 ). (c) On a two-dimensional block, Gardner’s quotient is ∣ z ∣ / R e z = sec arg z |z|/\mathrm{Re}\,z=\sec\arg z ∣ z ∣/ Re z = sec arg z for a point z z z on the circle of radius δ δ ′ \delta\delta' δ δ ′ about 1 1 1 (Remark 6.2 ); the largest argument is arcsin ( δ δ ′ ) \arcsin(\delta\delta') arcsin ( δ δ ′ ) . In the figure the contrast is written Δ ( ⋅ ) \Delta(\cdot) Δ ( ⋅ ) , so Δ ( A ) = O M C ( A ) \Delta(A)=\mathop{\mathrm{OMC}}(A) Δ ( A ) = OMC ( A ) .
The theorem has a simple geometric meaning (Figure 1 ). Krein [143 ] and Gustafson [127 ] introduced the angle ψ ( A ) \psi(A) ψ ( A ) of an operator, the largest angle between a vector x x x and its image A x Ax A x ; for A > 0 A>0 A > 0 , sin ψ ( A ) = O M C ( A ) \sin\psi(A)=\mathop{\mathrm{OMC}}(A) sin ψ ( A ) = OMC ( A ) , since A A A divided by the midpoint λ ˉ ( A ) \bar\lambda(A) λ ˉ ( A ) of its spectrum is 1 + E 1+E 1 + E with ∥ E ∥ = O M C ( A ) \|E\|=\mathop{\mathrm{OMC}}(A) ∥ E ∥ = OMC ( A ) . If ∣ X ∣ = v v ∗ |X|=vv^* ∣ X ∣ = v v ∗ has rank one (δ ′ = 1 \delta'=1 δ ′ = 1 ), Gardner’s quotient is ∣ ⟨ ρ v , U v ⟩ ∣ / ⟨ ρ v , v ⟩ ≤ ∥ ρ v ∥ / ⟨ ρ v , v ⟩ = sec ∠ ( v , ρ v ) ≤ sec ψ ( ρ ) |\langle \rho v,Uv\rangle|/\langle \rho v,v\rangle\le\|\rho
v\|/\langle \rho v,v\rangle=\sec\angle(v,\rho v)\le\sec\psi(\rho) ∣ ⟨ ρ v , U v ⟩ ∣/ ⟨ ρ v , v ⟩ ≤ ∥ ρ v ∥/ ⟨ ρ v , v ⟩ = sec ∠ ( v , ρ v ) ≤ sec ψ ( ρ ) , which is the Kantorovich constant of ρ \rho ρ . For general ∣ X ∣ |X| ∣ X ∣ , the proof reduces to two-dimensional blocks on which ρ \rho ρ and ∣ X ∣ |X| ∣ X ∣ are, up to scale, 1 + δ S 1+\delta S 1 + δ S and 1 + δ ′ S ′ 1+\delta'S' 1 + δ ′ S ′ with symmetries S , S ′ S,S' S , S ′ ; there T r ( ∣ X ∣ ρ ) \mathop{\mathrm{Tr}}(|X|\rho) Tr ( ∣ X ∣ ρ ) and the trace norm ∥ ∣ X ∣ ρ ∥ 1 \||X|\rho\|_1 ∥∣ X ∣ ρ ∥ 1 are 2 R e z 2\mathrm{Re}\,z 2 Re z and 2 ∣ z ∣ 2|z| 2∣ z ∣ for z = 1 + δ δ ′ e i α z=1+\delta\delta'e^{i\alpha} z = 1 + δ δ ′ e i α , where e ± i α e^{\pm i\alpha} e ± i α are the eigenvalues of the rotation S ′ S S'S S ′ S . Gardner’s quotient is sec arg z \sec\arg z sec arg z , and the circle described by z z z is seen from the origin under the angle θ \theta θ with sin θ = δ δ ′ = sin ψ ( ρ ) sin ψ ( ∣ X ∣ ) \sin\theta=\delta\delta'=\sin\psi(\rho)\sin\psi(|X|) sin θ = δ δ ′ = sin ψ ( ρ ) sin ψ ( ∣ X ∣ ) . The product of the contrasts is the radius of this circle, and the constant of Theorem A is sec θ \sec\theta sec θ .
The catalogue
Table 1 summarises the constants. Two general statements organise them. The first (Theorem 4.15 ) is a transfer principle: if an inequality is compatible with central rescaling, its upper bound is proved with the ambient condition number in arbitrary finite von Neumann algebras, and a matching example exists in M 2 ( C ) M_2(\mathbb{C}) M 2 ( C ) , then the violation constant is the same function of κ Z \kappa_\mathcal{Z} κ Z in every finite von Neumann algebra. The matrix envelopes of § 5 supply such examples and, conversely, the transfer principle turns several matrix results into results on von Neumann algebras. The constants that appear are G G G , its square, the condition number, the Specht ratio and coth 2 ( L N C / 2 ) \coth^2(\mathop{\mathrm{LNC}}/2) coth 2 ( LNC /2 ) ; on M n M_n M n the finite proved envelopes vanish at the trace to the orders δ \delta δ , δ 2 \delta^2 δ 2 and δ 4 \delta^4 δ 4 (item 37 is + ∞ +\infty + ∞ for every non-central ρ \rho ρ , i.e. non-scalar on M n M_n M n ), and the weighted monotonicity (item 25) has a proved lower bound of order δ 8 \delta^8 δ 8 , while only the trivial upper bound V n ≤ κ − 1 = 2 δ / ( 1 − δ ) V_n\le\kappa-1=2\delta/(1-\delta) V n ≤ κ − 1 = 2 δ / ( 1 − δ ) (constant ≤ κ Z \le\kappa_\mathcal{Z} ≤ κ Z in finite von Neumann algebras) is known for it; the computed envelope (δ ≥ 0.3 \delta\ge0.3 δ ≥ 0.3 ) does not exceed the lower bound (Theorem B and Figure 2 ). So the characterisations differ greatly in sensitivity: two inequalities that are equivalent for traces may fail at different rates. In particular, the monotonicity of A ↦ φ ( A 2 ) A\mapsto\varphi(A^2) A ↦ φ ( A 2 ) fails only to the order δ 4 \delta^4 δ 4 : its constant is ( G Z + 1 ) 2 / ( 4 G Z ) (G_\mathcal{Z}+1)^2/(4G_\mathcal{Z}) ( G Z + 1 ) 2 / ( 4 G Z ) (Theorem 5.27 ), and on M n M_n M n an explicit two-parameter family of rank-one perturbations attains it (Remark 5.28 ).
Of the 38 trace characterisations, 18 have an exact constant in every finite von Neumann algebra (items 1–5, 8, 10, 17, 18, 21–23, 26, 28, 29, 31, 33, and item 37, whose constant is + ∞ +\infty + ∞ ), five have an exact envelope on matrix algebras only (items 11, 19 and 24, item 9 for n = 2 n=2 n = 2 and item 20 for even n n n ), 13 have proved bounds (items 6, 7, 12, 14, 16, 25, 27, 30, 32, 34–36 and 38), and items 13 and 15 have numerical values only.
Returning to the unnormalised quantities of [85 ] , we find the sharp bound of ∣ T r ( X Z ) ∣ − T r ( X ∣ Z ∣ ) |\mathop{\mathrm{Tr}}(XZ)|-\mathop{\mathrm{Tr}}(X|Z|) ∣ Tr ( X Z ) ∣ − Tr ( X ∣ Z ∣ ) with both contrasts prescribed and those of the two other trace functionals sampled there (Theorem C), which answers the question of [85 ] .
Table 1. Violation constants. Items refer to Table 3 . Third column: the constant c ( ρ ) c(\rho) c ( ρ ) in every finite von Neumann algebra (“exact” when proved to be the stated function of κ Z \kappa_\mathcal{Z} κ Z ); in factors κ Z = κ \kappa_\mathcal{Z}=\kappa κ Z = κ and G Z = G G_\mathcal{Z}=G G Z = G . Fourth column: the envelope V n ( δ ) V_n(\delta) V n ( δ ) on M n M_n M n with the normalisation N \mathcal{N} N of Table 3 (V = c − 1 V=c-1 V = c − 1 when N = R \mathcal{N}=\mathcal{R} N = R ). y p \mathsf y_p y p , t p \mathsf t_p t p , m p \mathsf m_p m p , b 2 b_2 b 2 , g 25 g_{25} g 25 and s s s are defined in the cited results. Items 39–42: every von Neumann algebra (item 42 exact for finite M \mathcal{M} M ).
Noncentrality and the comparison theorem
The second general statement (Theorem 8.1 ) compares contrast and noncentrality: the contrast always gives a valid but weaker bound; it gives the exact constant if and only if max ⌈ ρ ⌉ \max\lceil \rho\rceil max ⌈ ρ ⌉ and min ⌊ ρ ⌋ \min\lfloor \rho\rfloor min ⌊ ρ ⌋ are attained at a common point of the spectrum of the centre; the loss is unbounded; and in every finite von Neumann algebra that is neither a factor nor abelian (every such von Neumann algebra, finite or not, for the constants of § 7 other than Theorem (?) ) the violation constant is not a function of the contrast, while for the exact cases it is a function of the noncentrality. On a direct sum ⨁ c M k c \bigoplus_cM_{k_c} ⨁ c M k c this means that only the contrast inside the blocks counts (Corollary 8.2 ): a density that varies only between the blocks violates none of the inequalities except the two that characterise the canonical trace, items 11 and 24.
Central elements
The dual problem, in which the functional is arbitrary and an element is to be recognised as central, has the same answer. For a ≥ 0 a\ge0 a ≥ 0 invertible the best constant λ \lambda λ in p a p ≤ λ a pap\le\lambda a p a p ≤ λa over all projections p p p is cosh 2 L N C ( a ) \cosh^2\mathop{\mathrm{LNC}}(a) cosh 2 LNC ( a ) , and the best constant in ∣ φ ( a ) ∣ ≤ C ∣ φ ∣ ( a ) |\varphi(a)|\le
C|\varphi|(a) ∣ φ ( a ) ∣ ≤ C ∣ φ ∣ ( a ) over all normal functionals is cosh L N C ( a ) \cosh\mathop{\mathrm{LNC}}(a) cosh LNC ( a ) , in every von Neumann algebra (§ 7 ).
Composite operators, thresholds and numerics
§ 9 records how contrasts behave under the composites through which the operators enter the inequalities: the rapidities O L C = artanh O M C \mathop{\mathrm{OLC}}=\operatorname{artanh}\mathop{\mathrm{OMC}} OLC = artanh OMC add under products and means (the triangle inequality of Hilbert’s projective metric), the angles arcsin O M C \arcsin\mathop{\mathrm{OMC}} arcsin OMC add under Jordan products (Gustafson’s criterion), and the commutator test φ ( A B ) = φ ( B A ) \varphi(AB)=\varphi(BA) φ ( A B ) = φ ( B A ) fails by exactly 2 O M C ( ρ ) O M C ( A ) O M C ( B ) 2\mathop{\mathrm{OMC}}(\rho)\mathop{\mathrm{OMC}}(A)\mathop{\mathrm{OMC}}(B) 2 OMC ( ρ ) OMC ( A ) OMC ( B ) in even dimension. These rules prove sharp thresholds: if O M C ( A ) = O M C ( B ) ≤ 2 2 / 3 \mathop{\mathrm{OMC}}(A)=\mathop{\mathrm{OMC}}(B)\le2\sqrt2/3 OMC ( A ) = OMC ( B ) ≤ 2 2 /3 , then ( A + B ) 1 / 2 ≤ A 1 / 2 + B 1 / 2 (A+B)^{1/2}\le A^{1/2}+B^{1/2} ( A + B ) 1/2 ≤ A 1/2 + B 1/2 as operators, so no positive functional violates subadditivity of the square root, while above 2 2 / 3 2\sqrt2/3 2 2 /3 suitable operators violate it for every density of contrast close enough to 1 1 1 ; and the power-mean inequality φ ( A 2 ) 1 / 2 ≤ φ ( A ) \varphi(A^2)^{1/2}\le\varphi(A) φ ( A 2 ) 1/2 ≤ φ ( A ) is never violated when O M C ( A ) ≤ ( n − 1 ) / n \mathop{\mathrm{OMC}}(A)\le\sqrt{(n-1)/n} OMC ( A ) ≤ ( n − 1 ) / n , sharply (§ 9.1 and Figure 6 ).
For the inequalities without a closed form we computed the envelopes on M 2 M_2 M 2 , M 3 M_3 M 3 and M 4 M_4 M 4 by constrained optimisation (§ 10 ), using the open-source library of [149 ] . Every reported value is recomputed in 40-digit arithmetic at an exactly admissible point, so it is a rigorous lower bound for the supremum, not a certified value of it. Where the envelope is known, the search reproduces it to 1 × 10 − 7 1\times 10^{-7} 1 × 1 0 − 7 –3 × 10 − 4 3\times 10^{-4} 3 × 1 0 − 4 ; it reaches the proved value of item 10 and the proved lower bound of item 12, and comes within 3 ⋅ 10 − 3 3\cdot10^{-3} 3 ⋅ 1 0 − 3 of the proved lower bound of item 16 (§ 10.6 ), and elsewhere it supports the conjectures of § 11 . Random sampling, the method of [85 ] , rarely comes close to the envelope: for monotonicity and subadditivity fewer than one random sample in a thousand violates the inequality at all. No proof in this paper depends on a floating-point computation.
What is new and what is classical
The characterisation theorems are due to the cited authors. The constants G G G , κ \kappa κ and the Specht ratio, and the inequalities behind them, are classical: the Kantorovich inequality [137 ] in the operator form of Greub and Rheinboldt [126 ] (equivalently, the Krein inequality and Gustafson’s operator angle [127 , 143 ] ), Wielandt’s inequality in the Bauer–Householder form [92 ] , Hölder’s inequality [121 ] , the Araki–Lieb–Thirring inequality [142 ] and Specht’s reverse arithmetic–geometric mean inequality [158 ] . In factors several of the upper bounds below are short consequences of these inequalities; on M n ( C ) M_n(\mathbb{C}) M n ( C ) the Gardner constant and the operator inequality P ρ P ≤ G 2 ρ P\rho P\le
G^2\rho P ρP ≤ G 2 ρ are contained in inequalities of Bourin [113 ] , and Bourin’s reverse of Araki’s inequality [112 ] gives our upper bound for monotonicity on approximately finite-dimensional algebras. Several statements of § 9 are also classical: the bounds for congruences are the submultiplicativity of the condition number, the positivity criterion for Jordan products is Gustafson’s [127 ] , and the norm bound for commutators of positive operators follows from Kittaneh’s inequality [141 ] . In § 4 , the central envelopes are the extreme elements of Halpern’s central spectrum [131 ] , and the identity ∥ A d A ∥ = κ Z ( ∣ A ∣ ) \|\mathop{\mathrm{Ad}}_A\|=\kappa_\mathcal{Z}(|A|) ∥ Ad A ∥ = κ Z ( ∣ A ∣ ) for the norm of an inner automorphism follows from the description of norms of elementary operators in [116 , 157 , 165 ] (Remark 4.9 ).
What we add is that the classical constants are the exact violation constants; the determination of these constants in finite von Neumann algebras and their central versions; the identification of the noncentrality as the governing parameter; the dependence on the product of the contrasts in Theorem A; the exact envelopes of the projection inequalities 17–23 (items 19 and 20 on M n M_n M n , item 20 for even n n n ), of the Bikchentaev–Abed inequality 28 and, for n = 2 n=2 n = 2 , of item 9, and of the monotonicity (item 10); the constructions with violations of the orders δ 4 \delta^4 δ 4 and δ 8 \delta^8 δ 8 ; the sharp bounds of Theorem C; the sharp trace version of the commutator bound; and the threshold results of § 9.1 . For the Petz–Zemánek inequality with k ≥ 3 k\ge3 k ≥ 3 the situation is different: the constant depends on more of the spectrum than its extreme points (Propositions 5.11 and 5.12 ). We know of no previous study of these constants; the closest work is [85 ] , which samples the unnormalised Gardner violation but defines no envelope and proves no bound.
Organisation
§ 2 reviews related work. § 3 fixes the notation, lists the inequalities and collects the tools. § 4 recalls central envelopes and noncentrality and proves the transfer principle. § 5 determines the violation constants item by item, § 6 proves Theorems A and C, § 7 treats the centre characterisations, and § 8 compares contrast and noncentrality, including direct sums. § 9 treats composite operators and the thresholds, § 10 the numerical evidence, and § 11 lists conjectures and open problems. Appendix § 12 shows the computed envelopes and random samples item by item.
Characterisations of the trace.
Gardner [125 ] proved that a linear functional on a C ∗ C^* C ∗ -algebra with ∣ φ ( x ) ∣ ≤ φ ( ∣ x ∣ ) |\varphi(x)|\le\varphi(|x|) ∣ φ ( x ) ∣ ≤ φ ( ∣ x ∣ ) for all x x x is a trace. Petz and Zemánek [153 ] showed that ∣ φ ( x k ) ∣ ≤ φ ( ∣ x ∣ k ) |\varphi(x^k)|\le\varphi(|x|^k) ∣ φ ( x k ) ∣ ≤ φ ( ∣ x ∣ k ) for all x x x and a single k ∈ N k\in\mathbb N k ∈ N already characterises positive traces, and characterised traces by subadditivity on the projection lattice [153 Theorem 4] . Stolyarov, Tikhonov and Sherstnev [160 ] characterised traces among normal semifinite weights by the subadditivity φ ( ∣ a 1 + a 2 ∣ ) ≤ φ ( ∣ a 1 ∣ ) + φ ( ∣ a 2 ∣ ) \varphi(|a_1+a_2|)\le\varphi(|a_1|)+\varphi(|a_2|) φ ( ∣ a 1 + a 2 ∣ ) ≤ φ ( ∣ a 1 ∣ ) + φ ( ∣ a 2 ∣ ) for self-adjoint a 1 , a 2 a_1,a_2 a 1 , a 2 , and Tikhonov [164 ] studied when the subadditivity φ ( f ( A + B ) ) ≤ φ ( f ( A ) ) + φ ( f ( B ) ) \varphi(f(A+B))\le\varphi(f(A))+\varphi(f(B)) φ ( f ( A + B )) ≤ φ ( f ( A )) + φ ( f ( B )) for positive A , B A,B A , B characterises tracial functionals; formally, for f = 1 ( 0 , ∞ ) f=\mathbf 1_{(0,\infty)} f = 1 ( 0 , ∞ ) and projections this inequality becomes φ ( P ∨ Q ) ≤ φ ( P ) + φ ( Q ) \varphi(P\vee Q)\le\varphi(P)+\varphi(Q) φ ( P ∨ Q ) ≤ φ ( P ) + φ ( Q ) . Bikchentaev and Tikhonov characterised the trace on M n M_n M n by Young’s inequality φ ( ∣ A B ∣ ) ≤ φ ( A p ) / p + φ ( B q ) / q \varphi(|AB|)\le\varphi(A^p)/p+\varphi(B^q)/q φ ( ∣ A B ∣ ) ≤ φ ( A p ) / p + φ ( B q ) / q for positive A , B A,B A , B [110 ] and by the monotonicity 0 ≤ A ≤ B ⇒ φ ( A p ) ≤ φ ( B p ) 0\le A\le
B\Rightarrow\varphi(A^p)\le\varphi(B^p) 0 ≤ A ≤ B ⇒ φ ( A p ) ≤ φ ( B p ) [111 ] ; Sano and Yatsu [155 ] treated monotonicity and convexity for general functions, Cho and Sano [117 ] Young’s inequality for arbitrary operators and general Young pairs, and Dinh Trung Hoa and Tikhonov the weighted forms of monotonicity [133 , 134 ] . Bikchentaev obtained a long series of characterisations by projection inequalities such as φ ( P Q P ) = φ ( Q P Q ) \varphi(PQP)=\varphi(QPQ) φ ( P QP ) = φ ( QP Q ) , φ ( P Q P ) ≤ φ ( Q ) \varphi(PQP)\le\varphi(Q) φ ( P QP ) ≤ φ ( Q ) and φ ( ∣ P − Q ∣ ) ≤ φ ( P + Q ) \varphi(|P-Q|)\le\varphi(P+Q) φ ( ∣ P − Q ∣ ) ≤ φ ( P + Q ) [95 , 96 , 97 , 100 ] , and by many other inequalities, see [101 ] and the survey [99 ] ; Bikchentaev and Abed [104 ] characterised traces by the anticommutator inequality φ ( A 2 + B 2 + K ( A B + B A ) ) ≥ 0 \varphi(A^2+B^2+K(AB+BA))\ge0 φ ( A 2 + B 2 + K ( A B + B A )) ≥ 0 for A , B ≥ 0 A,B\ge0 A , B ≥ 0 and all large K K K and by inequalities for two projections. Bikchentaev [98 ] characterised tracial functionals by the Peierls–Bogoliubov and the Araki–Lieb–Thirring inequalities, and Alhasan and Fawwaz [86 ] by φ ( A ) ≤ φ ( ∣ A + i B ∣ ) \varphi(A)\le\varphi(|A+iB|) φ ( A ) ≤ φ ( ∣ A + i B ∣ ) for A ≥ 0 A\ge0 A ≥ 0 and B = B ∗ B=B^* B = B ∗ . Further characterisations use the generalised Powers–Størmer inequality [132 ] , inequalities for determinants and power means [103 ] , the spectral geometric mean [105 ] , inequalities for ∣ φ ( A 2 ) ∣ |\varphi(A^2)| ∣ φ ( A 2 ) ∣ [109 ] and metrics on projections [102 ] . Recent contributions include [108 , 146 ] ; the canonical trace on M n M_n M n is characterised in [146 , 103 ] . Bikchentaev and Moslehian [107 ] survey these results; see also [91 , 106 ] .
Characterisations of central elements.
Novikov and Tikhonov [152 ] proved that a ≥ 0 a\ge0 a ≥ 0 is central if and only if p a p ≤ a pap\le a p a p ≤ a for all projections p p p , if and only if ∣ φ ( a ) ∣ ≤ ∣ φ ∣ ( a ) |\varphi(a)|\le|\varphi|(a) ∣ φ ( a ) ∣ ≤ ∣ φ ∣ ( a ) for all normal functionals φ \varphi φ , and gave several equivalent forms; Novikov [151 ] extended such results to unbounded operators affiliated with the algebra. Characterisations of central elements by local monotonicity or convexity of functions are due to Molnár [145 ] and Virosztek [167 ] . All these results are qualitative.
Reverse inequalities.
Our upper bounds are reverse inequalities of Kantorovich type. Kantorovich’s inequality [137 ] and its operator form due to Greub and Rheinboldt [126 ] bound ω ( ρ 2 ) ω ( 1 ) \omega(\rho^2)\omega(1) ω ( ρ 2 ) ω ( 1 ) by G 2 ω ( ρ ) 2 G^2\omega(\rho)^2 G 2 ω ( ρ ) 2 ; the Mond–Pečarić method [123 ] produces such constants systematically, Specht’s ratio [158 ] is the sharp constant of the reverse arithmetic–geometric mean inequality, and Bourin [112 , 113 ] and Fujii and Seo [122 ] proved reverse inequalities of Araki and Davis–Hansen–Pedersen type. The question here is different: not whether such a constant bounds the violation, but whether it is the smallest one.
Operator angles.
Krein [143 ] introduced the deviation of an operator and Gustafson [127 ] the equivalent angle ψ ( A ) \psi(A) ψ ( A ) , cos ψ ( A ) = inf x ≠ 0 R e ⟨ A x , x ⟩ / ( ∥ A x ∥ ∥ x ∥ ) \cos\psi(A)=\inf_{x\ne0}\mathrm{Re}\,\langle Ax,x\rangle/(\|Ax\|\|x\|) cos ψ ( A ) = inf x = 0 Re ⟨ A x , x ⟩ / ( ∥ A x ∥∥ x ∥ ) , the first antieigenvalue; for 0 < m ≤ A ≤ M 0<m\le A\le M 0 < m ≤ A ≤ M with m , M m,M m , M in the spectrum, cos ψ ( A ) = 2 M m / ( M + m ) \cos\psi(A)=2\sqrt{Mm}/(M+m) cos ψ ( A ) = 2 M m / ( M + m ) and sin ψ ( A ) = ( M − m ) / ( M + m ) \sin\psi(A)=(M-m)/(M+m) sin ψ ( A ) = ( M − m ) / ( M + m ) [129 , 128 ] . Thus the Michelson contrast of a positive operator is the sine of its angle, and the Kantorovich constant is G ( ρ ) = sec ψ ( ρ ) G(\rho)=\sec\psi(\rho) G ( ρ ) = sec ψ ( ρ ) . Gustafson used the angle to decide when products of positive operators are accretive [127 ] .
Jordan products.
The product A B AB A B of positive operators is accretive exactly when the Jordan product A B + B A = 2 R e ( A B ) AB+BA=2\mathrm{Re}\,(AB) A B + B A = 2 Re ( A B ) is positive. Bounds for the eigenvalues of A B + B A AB+BA A B + B A go back to Strang [161 ] and Nicholson [148 ] ; Alikakos and Bates [87 ] bounded all eigenvalues and showed that Strang’s lower bound for the smallest one is best possible [88 ] , and Conley, Pucci and Serrin [118 ] gave necessary and sufficient conditions for the positivity of the symmetrised product of two positive definite matrices. Moslehian and Najafi [147 Theorem 2.1] proved that A B + B A ≥ 0 AB+BA\ge0 A B + B A ≥ 0 if and only if f ( A + B ) ≤ f ( A ) + f ( B ) f(A+B)\le f(A)+f(B) f ( A + B ) ≤ f ( A ) + f ( B ) for every non-negative operator monotone function f f f on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) . This links Jordan products to the subadditivity inequalities that characterise traces [164 ] , and it is the background of § 9.1 .
Commutators, derivations and Hilbert’s metric.
Kittaneh [141 ] proved ∥ X Y − Y X ∥ ≤ ∥ X ∥ ∥ Y ∥ \|XY-YX\|\le\|X\|\|Y\| ∥ X Y − Y X ∥ ≤ ∥ X ∥∥ Y ∥ when X ≥ 0 X\ge0 X ≥ 0 , which gives ∥ A B − B A ∥ ≤ 1 2 ∥ A ∥ ∥ B ∥ \|AB-BA\|\le\frac12\|A\|\|B\| ∥ A B − B A ∥ ≤ 2 1 ∥ A ∥∥ B ∥ for positive A , B A,B A , B ; versions for unitarily invariant norms are in [140 , 94 ] . The norm of an inner derivation is twice the distance to the centre [159 , 135 , 124 , 168 ] , and the noncentrality is its multiplicative counterpart: the norm of the inner automorphism A d A \mathop{\mathrm{Ad}}_A Ad A is κ Z ( ∣ A ∣ ) \kappa_\mathcal{Z}(|A|) κ Z ( ∣ A ∣ ) , which follows from [116 , 157 , 165 ] (Remark 4.9 ). On the cone of positive definite matrices, Thompson’s metric [163 ] is d T ( A , B ) = ∥ log ( A − 1 / 2 B A − 1 / 2 ) ∥ d_T(A,B)=\|\log(A^{-1/2}BA^{-1/2})\| d T ( A , B ) = ∥ log ( A − 1/2 B A − 1/2 ) ∥ and Hilbert’s projective metric [115 , 144 ] is d H ( A , B ) = log ( λ max / λ min ) ( A − 1 / 2 B A − 1 / 2 ) d_H(A,B)=\log\bigl(\lambda_{\max}/\lambda_{\min}\bigr)(A^{-1/2}BA^{-1/2}) d H ( A , B ) = log ( λ m a x / λ m i n ) ( A − 1/2 B A − 1/2 ) . The logarithmic contrast of a positive definite A A A is
O L C ( A ) = artanh O M C ( A ) = 1 2 log λ max ( A ) λ min ( A ) = 1 2 d H ( A , 1 ) = min t > 0 d T ( A , t 1 ) , \mathop{\mathrm{OLC}}(A)=\operatorname{artanh}\mathop{\mathrm{OMC}}(A)=\tfrac12\log\frac{\lambda_{\max}(A)}{\lambda_{\min}(A)}=\tfrac12\,d_H(A,1)
=\min_{t>0}d_T(A,t1), OLC ( A ) = artanh OMC ( A ) = 2 1 log λ m i n ( A ) λ m a x ( A ) = 2 1 d H ( A , 1 ) = t > 0 min d T ( A , t 1 ) ,
the Thompson distance of A A A to the scalars; this is why it behaves additively under products (Proposition 9.1 ), and why its central version L N C \mathop{\mathrm{LNC}} LNC is the Thompson distance to the centre (Proposition 4.5 ).
Setting and tools
Von Neumann algebras and densities
M ⊂ B ( H ) \mathcal{M}\subset B(\mathcal{H}) M ⊂ B ( H ) is a von Neumann algebra with centre Z \mathcal{Z} Z ; M s a \mathcal{M}_{sa} M s a , M + \mathcal{M}_+ M + and M + + \mathcal{M}_{++} M ++ denote the self-adjoint, positive, and positive invertible elements. By the Gelfand representation Z ≅ C ( Ω ) \mathcal{Z}\cong C(\Omega) Z ≅ C ( Ω ) with Ω \Omega Ω extremally disconnected; we identify z ∈ Z z\in\mathcal{Z} z ∈ Z with its Gelfand transform. Central projections correspond to clopen subsets of Ω \Omega Ω , which form a base of the topology. For a projection e e e , c ( e ) c(e) c ( e ) is its central support, and s ( x ) s(x) s ( x ) is the support projection of x ∈ M + x\in\mathcal{M}_+ x ∈ M + . We use the comparison fact [136 Ch. 6] : if E , F E,F E , F are projections with c ( E ) c ( F ) ≠ 0 c(E)c(F)\ne0 c ( E ) c ( F ) = 0 , then there is a nonzero partial isometry v ∈ M v\in\mathcal{M} v ∈ M with v ∗ v ≤ F v^*v\le F v ∗ v ≤ F and v v ∗ ≤ E vv^*\le E v v ∗ ≤ E .
A von Neumann algebra is finite here if it carries a faithful normal tracial state τ \tau τ . Then ∥ X ∥ p = τ ( ∣ X ∣ p ) 1 / p \|X\|_p=\tau(|X|^p)^{1/p} ∥ X ∥ p = τ ( ∣ X ∣ p ) 1/ p , Hölder’s inequality ∣ τ ( X Y ) ∣ ≤ τ ( ∣ X Y ∣ ) ≤ ∥ X ∥ p ∥ Y ∥ q |\tau(XY)|\le\tau(|XY|)\le\|X\|_p\|Y\|_q ∣ τ ( X Y ) ∣ ≤ τ ( ∣ X Y ∣ ) ≤ ∥ X ∥ p ∥ Y ∥ q (1 / p + 1 / q = 1 1/p+1/q=1 1/ p + 1/ q = 1 ) and the duality ∥ X ∥ p = sup { ∣ τ ( X Y ) ∣ : ∥ Y ∥ q ≤ 1 } \|X\|_p=\sup\{|\tau(XY)|:\|Y\|_q\le1\} ∥ X ∥ p = sup { ∣ τ ( X Y ) ∣ : ∥ Y ∥ q ≤ 1 } hold [121 ] . Every X ∈ M X\in\mathcal{M} X ∈ M has a polar decomposition X = U ∣ X ∣ X=U|X| X = U ∣ X ∣ with U U U unitary, and ∣ X ∗ ∣ = U ∣ X ∣ U ∗ |X^*|=U|X|U^* ∣ X ∗ ∣ = U ∣ X ∣ U ∗ . On L 2 ( M , τ ) L^2(\mathcal{M},\tau) L 2 ( M , τ ) with ⟨ X , Y ⟩ = τ ( Y ∗ X ) \langle X,Y\rangle=\tau(Y^*X) ⟨ X , Y ⟩ = τ ( Y ∗ X ) the left and right multiplications L X h = X h L_X h=Xh L X h = X h , R X h = h X R_Xh=hX R X h = h X are bounded; for D ∈ M + + D\in\mathcal{M}_{++} D ∈ M ++ , L D L_D L D and R D R_D R D are commuting positive operators with spectra in σ ( D ) \sigma(D) σ ( D ) . We use the Araki–Lieb–Thirring inequality in Kosaki’s form [142 ] : for A , B ∈ M + A,B\in\mathcal{M}_+ A , B ∈ M + and r ≥ 1 r\ge1 r ≥ 1 ,
τ ( ( B 1 / 2 A B 1 / 2 ) r ) ≤ τ ( B r / 2 A r B r / 2 ) . \begin{equation}
\tau\bigl((B^{1/2}AB^{1/2})^r\bigr)\le\tau\bigl(B^{r/2}A^rB^{r/2}\bigr). \tag*{(2)}
\end{equation} τ ( ( B 1/2 A B 1/2 ) r ) ≤ τ ( B r /2 A r B r /2 ) . (2)
For ρ ∈ M + + \rho\in\mathcal{M}_{++} ρ ∈ M ++ put φ = φ ρ = τ ( ρ ⋅ ) \varphi=\varphi_\rho=\tau(\rho\,\cdot) φ = φ ρ = τ ( ρ ⋅ ) . Then φ \varphi φ is tracial if and only if ρ ∈ Z \rho\in\mathcal{Z} ρ ∈ Z : if τ ( ρ X Y ) = τ ( ρ Y X ) = τ ( X ρ Y ) \tau(\rho XY)=\tau(\rho YX)=\tau(X\rho Y) τ ( ρX Y ) = τ ( ρ Y X ) = τ ( X ρ Y ) for all X , Y X,Y X , Y , then τ ( ( ρ X − X ρ ) Y ) = 0 \tau((\rho X-X\rho)Y)=0 τ (( ρX − X ρ ) Y ) = 0 for all Y Y Y , so ρ X = X ρ \rho X=X\rho ρX = X ρ .
On M n = M n ( C ) M_n=M_n(\mathbb{C}) M n = M n ( C ) we use the usual trace T r \mathop{\mathrm{Tr}} Tr in place of τ \tau τ (all constants are invariant under τ ↦ s τ \tau\mapsto s\tau τ ↦ s τ ), ∣ X ∣ = ( X ∗ X ) 1 / 2 |X|=(X^*X)^{1/2} ∣ X ∣ = ( X ∗ X ) 1/2 , ∥ X ∥ 1 = T r ∣ X ∣ \|X\|_1=\mathop{\mathrm{Tr}}|X| ∥ X ∥ 1 = Tr ∣ X ∣ , R e X = 1 2 ( X + X ∗ ) \mathrm{Re}\,X=\frac12(X+X^*) Re X = 2 1 ( X + X ∗ ) , and φ = φ ρ = T r ( ρ ⋅ ) \varphi=\varphi_\rho=\mathop{\mathrm{Tr}}(\rho\,\cdot\,) φ = φ ρ = Tr ( ρ ⋅ ) with ρ ≥ 0 \rho\ge0 ρ ≥ 0 ; φ \varphi φ is a trace iff ρ ∈ C 1 \rho\in\mathbb{C}1 ρ ∈ C 1 . We write a ≥ b a\ge b a ≥ b for the extreme eigenvalues of ρ \rho ρ , e 1 , e 2 e_1,e_2 e 1 , e 2 for corresponding unit eigenvectors, and δ = O M C ( ρ ) \delta=\mathop{\mathrm{OMC}}(\rho) δ = OMC ( ρ ) ; unless stated otherwise ρ > 0 \rho>0 ρ > 0 , i.e. δ < 1 \delta<1 δ < 1 . Variables called P , Q , R P,Q,R P , Q , R are orthogonal projections, A , B ≥ 0 A,B\ge0 A , B ≥ 0 unless said otherwise, and X , Y X,Y X , Y arbitrary; P ∧ Q P\wedge Q P ∧ Q and P ∨ Q P\vee Q P ∨ Q are the projections onto r a n P ∩ r a n Q \mathrm{ran}\,P\cap\mathrm{ran}\,Q ran P ∩ ran Q and onto r a n P + r a n Q \mathrm{ran}\,P+\mathrm{ran}\,Q ran P + ran Q .
Table 2. Notation. In factors, in particular on M n M_n M n , κ Z = κ \kappa_\mathcal{Z}=\kappa κ Z = κ , δ Z = δ \delta_\mathcal{Z}=\delta δ Z = δ , ψ Z = ψ \psi_\mathcal{Z}=\psi ψ Z = ψ and G Z = G G_\mathcal{Z}=G G Z = G .
Contrast, angle and the Kantorovich constant
For a positive invertible operator T T T on a Hilbert space with m = min σ ( T ) m=\min\sigma(T) m = min σ ( T ) , M = max σ ( T ) M=\max\sigma(T) M = max σ ( T ) let
κ ( T ) = M m , O L C ( T ) = 1 2 ln κ ( T ) , O M C ( T ) = tanh O L C ( T ) = M − m M + m , G ( T ) = cosh O L C ( T ) = M + m 2 M m . \begin{gather*}
\kappa(T)=\frac Mm,\qquad \mathop{\mathrm{OLC}}(T)=\tfrac12\ln\kappa(T),\\
\mathop{\mathrm{OMC}}(T)=\tanh\mathop{\mathrm{OLC}}(T)=\frac{M-m}{M+m},\qquad G(T)=\cosh\mathop{\mathrm{OLC}}(T)=\frac{M+m}{2\sqrt{Mm}}.
\end{gather*} κ ( T ) = m M , OLC ( T ) = 2 1 ln κ ( T ) , OMC ( T ) = tanh OLC ( T ) = M + m M − m , G ( T ) = cosh OLC ( T ) = 2 M m M + m .
O M C \mathop{\mathrm{OMC}} OMC is the operator Michelson contrast of [85 ] and O L C \mathop{\mathrm{OLC}} OLC its logarithmic form [150 ] . In Gustafson’s operator trigonometry [127 ] the angle ψ ( T ) \psi(T) ψ ( T ) of T T T is defined by cos ψ ( T ) = inf h ≠ 0 ⟨ T h , h ⟩ / ( ∥ T h ∥ ∥ h ∥ ) \cos\psi(T)=\inf_{h\ne0}\langle Th,h\rangle/(\|Th\|\,\|h\|) cos ψ ( T ) = inf h = 0 ⟨ T h , h ⟩ / ( ∥ T h ∥ ∥ h ∥ ) , and cos ψ ( T ) = 2 M m / ( M + m ) \cos\psi(T)=2\sqrt{Mm}/(M+m) cos ψ ( T ) = 2 M m / ( M + m ) by the Krein inequality [127 , 143 ] . Hence
O M C ( T ) = sin ψ ( T ) , G ( T ) = cosh O L C ( T ) = sec ψ ( T ) . \mathop{\mathrm{OMC}}(T)=\sin\psi(T),\qquad G(T)=\cosh\mathop{\mathrm{OLC}}(T)=\sec\psi(T). OMC ( T ) = sin ψ ( T ) , G ( T ) = cosh OLC ( T ) = sec ψ ( T ) .
Lemma 3.1 (Kantorovich, Krein). Let 0 < m ≤ T ≤ M 0<m\le T\le M 0 < m ≤ T ≤ M and G = ( M + m ) / ( 2 M m ) G=(M+m)/(2\sqrt{Mm}) G = ( M + m ) / ( 2 M m ) .
ω ( T 2 ) ω ( 1 ) ≤ G 2 ω ( T ) 2 \omega(T^2)\,\omega(1)\le G^2\,\omega(T)^2 ω ( T 2 ) ω ( 1 ) ≤ G 2 ω ( T ) 2 for every positive functional ω \omega ω on a C ∗ C^* C ∗ -algebra containing T T T [126 ] .
∥ T h ∥ ∥ h ∥ ≤ G ⟨ T h , h ⟩ \|Th\|\,\|h\|\le G\,\langle Th,h\rangle ∥ T h ∥ ∥ h ∥ ≤ G ⟨ T h , h ⟩ for every vector h h h .
Proof. ( M − T ) ( T − m ) ≥ 0 (M-T)(T-m)\ge0 ( M − T ) ( T − m ) ≥ 0 gives ω ( T 2 ) ≤ ( M + m ) x − M m y \omega(T^2)\le(M+m)x-Mm\,y ω ( T 2 ) ≤ ( M + m ) x − M m y with x = ω ( T ) x=\omega(T) x = ω ( T ) , y = ω ( 1 ) y=\omega(1) y = ω ( 1 ) , and ( ( M + m ) x − M m y ) y ≤ G 2 x 2 ((M+m)x-Mmy)y\le G^2x^2 (( M + m ) x − M m y ) y ≤ G 2 x 2 because the difference is ( G x − M m y ) 2 (Gx-\sqrt{Mm}\,y)^2 ( G x − M m y ) 2 . Part k:b is k:a for the vector functional ω = ⟨ ⋅ h , h ⟩ \omega=\langle\cdot\,h,h\rangle ω = ⟨ ⋅ h , h ⟩ . ◻
With x = T 1 / 2 h x=T^{1/2}h x = T 1/2 h , Lemma 3.1 k:b is the Kantorovich inequality ⟨ T x , x ⟩ ⟨ T − 1 x , x ⟩ ≤ G ( T ) 2 ∥ x ∥ 4 \langle Tx,x\rangle\langle T^{-1}x,x\rangle\le G(T)^2\,\|x\|^4 ⟨ T x , x ⟩ ⟨ T − 1 x , x ⟩ ≤ G ( T ) 2 ∥ x ∥ 4 [137 ] ; equivalently, it is the case B = 1 B=1 B = 1 of the Greub–Rheinboldt form [126 ]
⟨ A 2 x , x ⟩ ⟨ B 2 x , x ⟩ ≤ ( M 1 M 2 + m 1 m 2 ) 2 4 m 1 m 2 M 1 M 2 ⟨ A B x , x ⟩ 2 \langle A^2x,x\rangle\langle B^2x,x\rangle\le\frac{(M_1M_2+m_1m_2)^2}{4m_1m_2M_1M_2}\,\langle ABx,x\rangle^2 ⟨ A 2 x , x ⟩ ⟨ B 2 x , x ⟩ ≤ 4 m 1 m 2 M 1 M 2 ( M 1 M 2 + m 1 m 2 ) 2 ⟨ A B x , x ⟩ 2
for commuting positive A , B A,B A , B with 0 < m 1 ≤ A ≤ M 1 0<m_1\le A\le M_1 0 < m 1 ≤ A ≤ M 1 and 0 < m 2 ≤ B ≤ M 2 0<m_2\le B\le M_2 0 < m 2 ≤ B ≤ M 2 . We call G ( T ) G(T) G ( T ) the Kantorovich constant of T T T ; its square G ( T ) 2 = ( M + m ) 2 / ( 4 M m ) G(T)^2=(M+m)^2/(4Mm) G ( T ) 2 = ( M + m ) 2 / ( 4 M m ) is the best constant in Kantorovich’s inequality, and the expression ⟨ T x , x ⟩ ⟨ T − 1 x , x ⟩ \langle Tx,x\rangle\langle T^{-1}x,x\rangle ⟨ T x , x ⟩ ⟨ T − 1 x , x ⟩ for a unit vector x x x will be called the Kantorovich expression.
Lemma 3.2 (Wielandt). Let δ = O M C ( T ) \delta=\mathop{\mathrm{OMC}}(T) δ = OMC ( T ) .
If R e ⟨ x , y ⟩ ≥ 0 \mathop{\mathrm{Re}}\langle x,y\rangle\ge0 Re ⟨ x , y ⟩ ≥ 0 , then − R e ⟨ T x , y ⟩ ≤ δ ⟨ T x , x ⟩ 1 / 2 ⟨ T y , y ⟩ 1 / 2 -\mathop{\mathrm{Re}}\langle Tx,y\rangle\le\delta\,\langle Tx,x\rangle^{1/2}\langle Ty,y\rangle^{1/2} − Re ⟨ T x , y ⟩ ≤ δ ⟨ T x , x ⟩ 1/2 ⟨ T y , y ⟩ 1/2 .
If x ⊥ y x\perp y x ⊥ y , then ∣ ⟨ T x , y ⟩ ∣ ≤ δ ⟨ T x , x ⟩ 1 / 2 ⟨ T y , y ⟩ 1 / 2 |\langle Tx,y\rangle|\le\delta\,\langle
Tx,x\rangle^{1/2}\langle Ty,y\rangle^{1/2} ∣ ⟨ T x , y ⟩ ∣ ≤ δ ⟨ T x , x ⟩ 1/2 ⟨ T y , y ⟩ 1/2 .
Proof. w:one Put u = T 1 / 2 x u=T^{1/2}x u = T 1/2 x , v = T 1 / 2 y v=T^{1/2}y v = T 1/2 y and c = 1 2 ( m − 1 + M − 1 ) c=\frac12(m^{-1}+M^{-1}) c = 2 1 ( m − 1 + M − 1 ) . Then T − 1 = c ( 1 + E ) T^{-1}=c(1+E) T − 1 = c ( 1 + E ) with E E E self-adjoint and ∥ E ∥ ≤ ( m − 1 − M − 1 ) / ( m − 1 + M − 1 ) = δ \|E\|\le(m^{-1}-M^{-1})/(m^{-1}+M^{-1})=\delta ∥ E ∥ ≤ ( m − 1 − M − 1 ) / ( m − 1 + M − 1 ) = δ . Hence 0 ≤ R e ⟨ x , y ⟩ = R e ⟨ T − 1 u , v ⟩ = c ( R e ⟨ u , v ⟩ + R e ⟨ E u , v ⟩ ) 0\le\mathop{\mathrm{Re}}\langle x,y\rangle=\mathop{\mathrm{Re}}\langle T^{-1}u,v\rangle=c(\mathop{\mathrm{Re}}\langle u,v\rangle+\mathop{\mathrm{Re}}\langle
Eu,v\rangle) 0 ≤ Re ⟨ x , y ⟩ = Re ⟨ T − 1 u , v ⟩ = c ( Re ⟨ u , v ⟩ + Re ⟨ E u , v ⟩) , so − R e ⟨ T x , y ⟩ = − R e ⟨ u , v ⟩ ≤ ∥ E ∥ ∥ u ∥ ∥ v ∥ -\mathop{\mathrm{Re}}\langle Tx,y\rangle=-\mathop{\mathrm{Re}}\langle u,v\rangle\le\|E\|\,\|u\|\,\|v\| − Re ⟨ T x , y ⟩ = − Re ⟨ u , v ⟩ ≤ ∥ E ∥ ∥ u ∥ ∥ v ∥ . w:two If x ⊥ y x\perp y x ⊥ y , then w:one applies to x x x and − e i θ y -e^{i\theta}y − e i θ y for every θ \theta θ . ◻
Part w:two is Wielandt’s inequality in the form of Bauer and Householder [92 ] ; part w:one is a one-sided version that we need for the Bikchentaev–Abed inequality.
Lemma 3.3. For every projection p p p on H \mathcal{H} H , p T p ≤ cosh 2 O L C ( T ) T pTp\le\cosh^2\mathop{\mathrm{OLC}}(T)\,T pT p ≤ cosh 2 OLC ( T ) T .
Proof. For k ∈ H k\in\mathcal{H} k ∈ H let x = p k x=pk x = p k , y = ( 1 − p ) k y=(1-p)k y = ( 1 − p ) k , X = ⟨ T x , x ⟩ X=\langle Tx,x\rangle X = ⟨ T x , x ⟩ , Y = ⟨ T y , y ⟩ Y=\langle Ty,y\rangle Y = ⟨ T y , y ⟩ and δ = O M C ( T ) \delta=\mathop{\mathrm{OMC}}(T) δ = OMC ( T ) . By Lemma 3.2 w:two , ⟨ T k , k ⟩ ≥ X + Y − 2 δ X Y ≥ ( 1 − δ 2 ) X \langle Tk,k\rangle\ge X+Y-2\delta\sqrt{XY}\ge(1-\delta^2)X ⟨ T k , k ⟩ ≥ X + Y − 2 δ X Y ≥ ( 1 − δ 2 ) X (minimise over Y \sqrt Y Y ), and ( 1 − δ 2 ) − 1 = cosh 2 O L C ( T ) (1-\delta^2)^{-1}=\cosh^2\mathop{\mathrm{OLC}}(T) ( 1 − δ 2 ) − 1 = cosh 2 OLC ( T ) . ◻
Lemma 3.3 is classical. Since p T p ≤ λ T pTp\le\lambda T pT p ≤ λ T if and only if ∥ T 1 / 2 p T − 1 / 2 ∥ 2 ≤ λ \|T^{1/2}pT^{-1/2}\|^2\le\lambda ∥ T 1/2 p T − 1/2 ∥ 2 ≤ λ , the best constant for a rank-one projection p = ∣ e ⟩ ⟨ e ∣ p=|e\rangle\langle e| p = ∣ e ⟩ ⟨ e ∣ is the Kantorovich expression ⟨ T e , e ⟩ ⟨ T − 1 e , e ⟩ \langle Te,e\rangle\langle T^{-1}e,e\rangle ⟨ T e , e ⟩ ⟨ T − 1 e , e ⟩ , and the lemma is then the Kantorovich inequality [137 ] . For general p p p it is the Wielandt (Bauer–Householder) bound [92 ] on the norm of the oblique projection T 1 / 2 p T − 1 / 2 T^{1/2}pT^{-1/2} T 1/2 p T − 1/2 , whose range T 1 / 2 p H T^{1/2}p\mathcal{H} T 1/2 p H and kernel T 1 / 2 ( 1 − p ) H T^{1/2}(1-p)\mathcal{H} T 1/2 ( 1 − p ) H make an angle with cosine at most δ \delta δ by Lemma 3.2 w:two . For matrices it is also the case A = p A=p A = p of Bourin’s inequality A Z A ≤ ( a + b ) 2 4 a b Z AZA\le\frac{(a+b)^2}{4ab}Z A Z A ≤ 4 ab ( a + b ) 2 Z for 0 ≤ A ≤ 1 0\le A\le1 0 ≤ A ≤ 1 and Z > 0 Z>0 Z > 0 with extremal eigenvalues a , b a,b a , b [113 Cor. 1.5] .
Thompson metric
For B , C ∈ M + + B,C\in\mathcal{M}_{++} B , C ∈ M ++ let M ( B / C ) = inf { t > 0 : B ≤ t C } M(B/C)=\inf\{t>0:B\le tC\} M ( B / C ) = inf { t > 0 : B ≤ tC } and d T ( B , C ) = ln max { M ( B / C ) , M ( C / B ) } d_T(B,C)=\ln\max\{M(B/C),M(C/B)\} d T ( B , C ) = ln max { M ( B / C ) , M ( C / B )} , the Thompson metric [163 ] .
Violation constants and envelopes
For each inequality L ≤ R \mathcal{L}\le\mathcal{R} L ≤ R of Table 3 the violation constant c M ( ρ ) c_\mathcal{M}(\rho) c M ( ρ ) is the best constant c c c in L ≤ c R \mathcal{L}\le c\,\mathcal{R} L ≤ c R on the algebra M \mathcal{M} M , i.e. the supremum of L / R \mathcal{L}/\mathcal{R} L / R over the admissible variables (Definition 4.11 makes this precise). On M n M_n M n we measure the violation also by the quotient ( L − R ) / N (\mathcal{L}-\mathcal{R})/\mathcal{N} ( L − R ) / N with the normalisation N \mathcal{N} N of Table 3 ; it is invariant under ρ ↦ t ρ \rho\mapsto t\rho ρ ↦ tρ , t > 0 t>0 t > 0 . For a fixed density V ( ρ ) V(\rho) V ( ρ ) denotes its supremum, and the envelope V n ( δ ) V_n(\delta) V n ( δ ) is the supremum of V ( ρ ) V(\rho) V ( ρ ) over all ρ \rho ρ in M n M_n M n with O M C ( ρ ) = δ \mathop{\mathrm{OMC}}(\rho)=\delta OMC ( ρ ) = δ . If N = R \mathcal{N}=\mathcal{R} N = R , then V ( ρ ) = c M n ( ρ ) − 1 V(\rho)=c_{M_n}(\rho)-1 V ( ρ ) = c M n ( ρ ) − 1 ; since κ Z = κ \kappa_\mathcal{Z}=\kappa κ Z = κ on M n M_n M n , every constant of the form c ( ρ ) = f ( κ Z ( ρ ) ) c(\rho)=f(\kappa_\mathcal{Z}(\rho)) c ( ρ ) = f ( κ Z ( ρ )) gives V n ( δ ) = f ( 1 + δ 1 − δ ) − 1 V_n(\delta)=f(\frac{1+\delta}{1-\delta})-1 V n ( δ ) = f ( 1 − δ 1 + δ ) − 1 , the same in every dimension n ≥ 2 n\ge2 n ≥ 2 . We call a value of a constant or an envelope attained if the supremum is a maximum and approached otherwise.
The inequalities
Table 3 lists the inequalities. Items 1–25 are those for which we compute matrix envelopes; items 26–38 are further characterisations of traces, several of them families that contain an item among 1–25; items 39–42 characterise central elements.
Table 3. The inequalities studied, with the normalisation N \mathcal{N} N of the violation ( L − R ) / N (\mathcal{L}-\mathcal{R})/\mathcal{N} ( L − R ) / N on
M n M_n M n and the source of the characterisation. Projections are P , Q , R P,Q,R P , Q , R ; positive operators A , B A,B A , B
unless stated otherwise. Remarks on the forms are in § 3.6 .
Several sources state families of inequalities; we use one member. In item 4, ∣ φ ( A k ) ∣ ≤ φ ( ∣ A ∣ k ) |\varphi(A^k)|\le\varphi(|A|^k) ∣ φ ( A k ) ∣ ≤ φ ( ∣ A ∣ k ) for all A A A and one fixed k ≥ 1 k\ge1 k ≥ 1 characterises the trace on M n M_n M n [153 Theorem 1] ; k = 1 k=1 k = 1 is Gardner’s inequality. Item 7 is the Araki–Lieb–Thirring inequality φ ( ( B 1 / 2 A B 1 / 2 ) r q ) ≤ φ ( ( B r / 2 A r B r / 2 ) q ) \varphi((B^{1/2}AB^{1/2})^{rq})\le\varphi((B^{r/2}A^rB^{r/2})^q) φ (( B 1/2 A B 1/2 ) r q ) ≤ φ (( B r /2 A r B r /2 ) q ) (item 38) with r = 2 r=2 r = 2 and outer exponent 1 1 1 [98 Lemma 2] ; we call the outer exponent q q q throughout, as in item 38, because p p p denotes the Young exponent of items 29–30. The published form of item 8 is φ ( ∣ A B ∣ ) ≤ 1 p φ ( A p ) + 1 q φ ( B q ) \varphi(|AB|)\le\frac1p\varphi(A^p)+
\frac1q\varphi(B^q) φ ( ∣ A B ∣ ) ≤ p 1 φ ( A p ) + q 1 φ ( B q ) for A , B ≥ 0 A,B\ge0 A , B ≥ 0 [110 ] ; we use p = q = 2 p=q=2 p = q = 2 and allow arbitrary A , B A,B A , B , in the form of Ando’s matrix Young inequality [89 ] , and Corollary 5.20 refers to this extension; it differs from item 29 (Remark 5.21 ). Item 9 is the member p = 1 p=1 p = 1 of the family φ ( P Q + Q P ) ≤ 2 φ ( ( Q P Q ) p ) \varphi(PQ+QP)\le2\varphi((QPQ)^p) φ ( P Q + QP ) ≤ 2 φ (( QP Q ) p ) , p ∈ ( 0 , 1 ] p\in(0,1] p ∈ ( 0 , 1 ] [104 ] . Items 10 and 25 are the cases r = 0 r=0 r = 0 and r = 1 r=1 r = 1 of φ ( A 2 r + 2 ) ≤ φ ( A r B 2 A r ) \varphi(A^{2r+2})\le\varphi(A^rB^2A^r) φ ( A 2 r + 2 ) ≤ φ ( A r B 2 A r ) for 0 ≤ A ≤ B 0\le A\le B 0 ≤ A ≤ B , which characterises traces for every fixed r ≥ 0 r\ge0 r ≥ 0 [134 Lemma 1, Theorems 4, 5] ; on M n M_n M n the case r = 0 r=0 r = 0 is in [111 , 155 ] and r > 0 r>0 r > 0 in [133 Theorem 6] . Items 11 and 24 characterise the canonical trace among the positive functionals with φ ( 1 ) = n \varphi(1)=n φ ( 1 ) = n [103 Theorem 3] ; item 24 is (vii) there with p = 2 p=2 p = 2 , q = 1 q=1 q = 1 and needs φ \varphi φ faithful. Item 12 is the case f ( t ) = t 1 / 2 f(t)=t^{1/2} f ( t ) = t 1/2 of Tikhonov’s characterisation by φ ( f ( A + B ) ) ≤ φ ( f ( A ) ) + φ ( f ( B ) ) \varphi(f(A+B))\le\varphi(f(A))+\varphi(f(B)) φ ( f ( A + B )) ≤ φ ( f ( A )) + φ ( f ( B )) for concave f f f with f ( 0 ) = 0 f(0)=0 f ( 0 ) = 0 [164 Corollary 1] . Item 13 is the generalised Powers–Størmer inequality φ ( A + B ) − φ ( ∣ A − B ∣ ) ≤ 2 φ ( f ( A ) 1 / 2 g ( B ) f ( A ) 1 / 2 ) \varphi(A+B)-\varphi(|A-B|)\le2\varphi(f(A)^{1/2}g(B)f(A)^{1/2}) φ ( A + B ) − φ ( ∣ A − B ∣ ) ≤ 2 φ ( f ( A ) 1/2 g ( B ) f ( A ) 1/2 ) , g ( t ) = t / f ( t ) g(t)=t/f(t) g ( t ) = t / f ( t ) , with f ( t ) = t 1 / 2 f(t)=t^{1/2} f ( t ) = t 1/2 [132 Lemma 3.1, Remark 4] ; with N = R \mathcal{N}=\mathcal{R} N = R its envelope would be infinite (Remark 5.42 ), so we normalise by φ ( A + B ) \varphi(A+B) φ ( A + B ) . Item 14 is Hölder’s inequality for projections with the exponent α = 3 \alpha=3 α = 3 ; for α = 2 \alpha=2 α = 2 it holds for every positive φ \varphi φ [97 remark after Theorem 2] . Item 16 is the Peierls–Bogoliubov inequality, stated in [98 ] for positive H , K H,K H , K ; both sides scale by e c e^c e c under A ↦ A + c A\mapsto A+c A ↦ A + c , B ↦ B + c B\mapsto B+c B ↦ B + c , so the envelope over self-adjoint arguments is the same. Items 17–19 are [96 Theorem 3.4] , announced in [95 ] , and items 14, 15, 20–22 are [97 Theorem 2] . The remaining items are taken as in Table 3 . In item 2 the four versions with arbitrary, self-adjoint and positive variables and with projections (item 18) have the same constant (Theorem 5.4 ); the self-adjoint version is the form in which [160 ] characterises traces. The equality φ ( ∣ X ∗ ∣ ) = φ ( ∣ X ∣ ) \varphi(|X^*|)=\varphi(|X|) φ ( ∣ X ∗ ∣ ) = φ ( ∣ X ∣ ) for all X X X characterises traces, since ∣ X ∗ ∣ = U ∣ X ∣ U ∗ |X^*|=U|X|U^* ∣ X ∗ ∣ = U ∣ X ∣ U ∗ (item 26). Item 28 is [104 Thm. 4.6] , in which the inequality is required for all K K K larger than some K 0 > − 1 K_0>-1 K 0 > − 1 ; Corollary 5.15 shows that arbitrarily large K K K suffice. In item 29 the Cho–Sano form φ ( ∣ A ∗ B ∣ ) \varphi(|A^*B|) φ ( ∣ A ∗ B ∣ ) gives the same constant (§ 5.6 ). Bikchentaev’s characterisation by φ ( e P Q P ) ≤ φ ( e Q ) \varphi(e^{PQP})\le\varphi(e^Q) φ ( e P QP ) ≤ φ ( e Q ) concerns projections [99 ] ; item 20 is this inequality with the roles of P P P and Q Q Q interchanged, and item 36 allows Q ≥ 0 Q\ge0 Q ≥ 0 (for self-adjoint Q Q Q it fails already for traces, § 5.12 ). Items 16 and 37 are the two forms of the Peierls–Bogoliubov inequality of [98 ] ; for non-tracial φ \varphi φ the number φ ( B e A ) \varphi(Be^A) φ ( B e A ) need not be real, and item 37 uses its real part. Items 39–42 are [152 Thm. 1 and Cor. 1] ; the convexity items (vii) and (viii) of [152 Remark 1] are not treated here.
Two projections
Lemma 3.4 (Two projections; Halmos [130 ] ). For projections P , Q P,Q P , Q on a finite-dimensional Hilbert space H H H , H H H is the orthogonal sum of subspaces of dimension 1 1 1 or 2 2 2 invariant under P P P and Q Q Q . On each two-dimensional one there is an orthonormal basis f , f ′ f,f' f , f ′ with P = f f ∗ P=ff^* P = f f ∗ and Q = q q ∗ Q=qq^* Q = q q ∗ , q = cos θ f + e i χ sin θ f ′ q=\cos\theta\,f+e^{i\chi}\sin\theta\,f' q = cos θ f + e i χ sin θ f ′ , 0 < θ < π / 2 0<\theta<\pi/2 0 < θ < π /2 .
Contrasts of the operator variables
The envelope V n ( δ ) V_n(\delta) V n ( δ ) lets the operator variables range over all operators. To see how the violation depends on the operators themselves, we also prescribe their contrast. For a positive variable A A A we use O M C ( A ) \mathop{\mathrm{OMC}}(A) OMC ( A ) , for an arbitrary or self-adjoint variable X X X the contrast O M C ( ∣ X ∣ ) \mathop{\mathrm{OMC}}(|X|) OMC ( ∣ X ∣ ) of its modulus (for the Peierls–Bogoliubov inequality, item 16, the variables are positive, as in [98 ] ), and we let every operator variable of an item have the same contrast δ ′ ∈ [ 0 , 1 ] \delta'\in[0,1] δ ′ ∈ [ 0 , 1 ] :
V n ( δ , δ ′ ) = sup { L − R N : O M C ( ρ ) = δ , every operator variable of contrast δ ′ } . V_n(\delta,\delta')=\sup\Bigl\{\frac{\mathcal{L}-\mathcal{R}}{\mathcal{N}}:\ \mathop{\mathrm{OMC}}(\rho)=\delta,\ \text{every operator variable of contrast }\delta'\Bigr\}. V n ( δ , δ ′ ) = sup { N L − R : OMC ( ρ ) = δ , every operator variable of contrast δ ′ } .
Projections other than 0 0 0 and 1 1 1 have contrast 1 1 1 , so for items 5, 9, 14, 15 and 17–23 only δ ′ = 1 \delta'=1 δ ′ = 1 occurs, and V n ( δ , 1 ) ≤ V n ( δ ) V_n(\delta,1)\le V_n(\delta) V n ( δ , 1 ) ≤ V n ( δ ) with equality whenever the supremum over all operators is approached by singular ones. A positive operator of contrast δ ′ \delta' δ ′ whose spectrum has only two points is, up to a positive factor,
A = ℓ ′ + ( 1 − ℓ ′ ) P , ℓ ′ = 1 − δ ′ 1 + δ ′ , P a projection , P ≠ 0 , 1 ; A=\ell'+(1-\ell')P,\qquad \ell'=\frac{1-\delta'}{1+\delta'},\qquad P\ \text{a projection},\ P\ne0,1; A = ℓ ′ + ( 1 − ℓ ′ ) P , ℓ ′ = 1 + δ ′ 1 − δ ′ , P a projection , P = 0 , 1 ;
we call such operators two-level. They are the affine images of projections, and as δ ′ → 1 \delta'\to1 δ ′ → 1 they become scaled projections.
Proposition 3.5. If every operator variable is a positive multiple of the identity (for arbitrary and self-adjoint variables: if every modulus is), then L ≤ R \mathcal{L}\le\mathcal{R} L ≤ R in every item 1–25 of Table 3 and for every positive φ \varphi φ . Hence V n ( δ , 0 ) ≤ 0 V_n(\delta,0)\le0 V n ( δ , 0 ) ≤ 0 for all items; and V n ( 0 , δ ′ ) ≤ 0 V_n(0,\delta')\le0 V n ( 0 , δ ′ ) ≤ 0 , since for δ = 0 \delta=0 δ = 0 the functional is a multiple of the trace (the canonical trace in items 11 and 24, where φ ( 1 ) = n \varphi(1)=n φ ( 1 ) = n ).
Proof. Item by item, with ∣ X ∣ = x |X|=x ∣ X ∣ = x , ∣ Y ∣ = y |Y|=y ∣ Y ∣ = y scalar and X = x U X=xU X = xU , Y = y V Y=yV Y = y V with unitaries (symmetries for self-adjoint X , Y X,Y X , Y ), and A = α A=\alpha A = α , B = β B=\beta B = β scalars: (1) ∣ φ ( U ) ∣ ≤ φ ( 1 ) |\varphi(U)|\le\varphi(1) ∣ φ ( U ) ∣ ≤ φ ( 1 ) ; (2) φ ( ∣ x U + y V ∣ ) ≤ ∥ x U + y V ∥ φ ( 1 ) ≤ ( x + y ) φ ( 1 ) \varphi(|xU+yV|)\le\|xU+yV\|\varphi(1)\le(x+y)\varphi(1) φ ( ∣ xU + y V ∣ ) ≤ ∥ xU + y V ∥ φ ( 1 ) ≤ ( x + y ) φ ( 1 ) ; (3) and (4) as (1); (6) and (7) are equalities; (8) ∣ A B ∗ ∣ = x y |AB^*|=xy ∣ A B ∗ ∣ = x y and x y ≤ 1 2 ( x 2 + y 2 ) xy\le\frac12(x^2+y^2) x y ≤ 2 1 ( x 2 + y 2 ) ; (10) and (25) follow from α ≤ β \alpha\le\beta α ≤ β ; (11) is an equality for φ ( 1 ) = n \varphi(1)=n φ ( 1 ) = n ; (12) α + β ≤ α + β \sqrt{\alpha+\beta}\le\sqrt\alpha+\sqrt\beta α + β ≤ α + β ; (13) α + β − ∣ α − β ∣ = 2 min ( α , β ) ≤ 2 α β \alpha+\beta-|\alpha-\beta|=2\min(\alpha,\beta)\le2\sqrt{\alpha\beta} α + β − ∣ α − β ∣ = 2 min ( α , β ) ≤ 2 α β ; (16) is an equality; (24) φ ( α 2 ) 1 / 2 = α n ≤ α n \varphi(\alpha^2)^{1/2}=\alpha\sqrt n\le\alpha n φ ( α 2 ) 1/2 = α n ≤ α n . ◻
So a violation needs both contrasts: the functional must be far from a trace and the operators must be far from scalars. The next lemma is the mechanism behind the two-level operators: whenever the violation depends convexly on the spectrum of a variable, the extremal spectra have only two points.
Lemma 3.6 (Vertex principle). Fix all variables except the eigenvalues λ ∈ [ ℓ ′ , 1 ] n \lambda\in[\ell',1]^n λ ∈ [ ℓ ′ , 1 ] n of one positive variable (its eigenbasis fixed), and let F ( λ ) F(\lambda) F ( λ ) and N ( λ ) > 0 \mathcal{N}(\lambda)>0 N ( λ ) > 0 be the resulting values of L − R \mathcal{L}-\mathcal{R} L − R and N \mathcal{N} N . If λ ↦ F ( λ ) − c N ( λ ) \lambda\mapsto F(\lambda)-c\mathcal{N}(\lambda) λ ↦ F ( λ ) − c N ( λ ) is convex for every c ∈ R c\in\mathbb{R} c ∈ R , then the supremum of F / N F/\mathcal{N} F / N over the spectra with max i λ i = 1 \max_i\lambda_i=1 max i λ i = 1 and min i λ i = ℓ ′ \min_i\lambda_i=\ell' min i λ i = ℓ ′ equals its supremum over the two-level spectra λ ∈ { ℓ ′ , 1 } n \lambda\in\{\ell',1\}^n λ ∈ { ℓ ′ , 1 } n (both values occurring).
Proof. The admissible set is the union of the faces { λ i = 1 , λ j = ℓ ′ } \{\lambda_i=1,\ \lambda_j=\ell'\} { λ i = 1 , λ j = ℓ ′ } of the cube, and the vertices of these faces are the two-level spectra. If F / N > c F/\mathcal{N}>c F / N > c at an admissible point, then F − c N > 0 F-c\mathcal{N}>0 F − c N > 0 there, and by convexity F − c N > 0 F-c\mathcal{N}>0 F − c N > 0 at a vertex of the same face, where therefore F / N > c F/\mathcal{N}>c F / N > c . ◻
The hypothesis holds for Gardner’s inequality (F − c N = ∣ φ ( U T ) ∣ − ( 1 + c ) φ ( T ) F-c\mathcal{N}=|\varphi(UT)|-(1+c)\varphi(T) F − c N = ∣ φ ( U T ) ∣ − ( 1 + c ) φ ( T ) for X = U T X=UT X = U T : the modulus of a linear function minus a linear one), for item 3, for item 24 (φ ( A 2 ) 1 / 2 = ∥ ρ 1 / 2 A ∥ 2 \varphi(A^2)^{1/2}=\|\rho^{1/2}A\|_2 φ ( A 2 ) 1/2 = ∥ ρ 1/2 A ∥ 2 is a norm of a linear function of λ \lambda λ ) and for the unnormalised Gardner violation of § 6 ; for these items two-level operators suffice. For the other items, whether they suffice is a question we examine numerically (§ 10.4 ).
Central envelopes, noncentrality and the transfer principle
Central envelopes and noncentrality
We collect, with short proofs, the facts on noncentrality that we use. The envelopes in Lemma 4.1 are the largest and smallest elements C u , C l C_u,C_l C u , C l of Halpern’s central spectrum [131 Prop. 3.11, 3.12] (taken with respect to the zero central ideal); we include short proofs of the elementary properties we need.
Lemma 4.1 (Central envelopes; cf. [131 ] ). Let B ∈ M s a B\in\mathcal{M}_{sa} B ∈ M s a . The set { z ∈ Z s a : z ≥ B } \{z\in\mathcal{Z}_{sa}:z\ge B\} { z ∈ Z s a : z ≥ B } has a least element ⌈ B ⌉ \lceil B\rceil ⌈ B ⌉ and { z ∈ Z s a : z ≤ B } \{z\in\mathcal{Z}_{sa}:z\le B\} { z ∈ Z s a : z ≤ B } has a greatest element ⌊ B ⌋ \lfloor B\rfloor ⌊ B ⌋ . Moreover:
⌊ B ⌋ ≤ B ≤ ⌈ B ⌉ \lfloor B\rfloor\le B\le\lceil B\rceil ⌊ B ⌋ ≤ B ≤ ⌈ B ⌉ , ⌊ B ⌋ = − ⌈ − B ⌉ \lfloor B\rfloor=-\lceil -B\rceil ⌊ B ⌋ = − ⌈ − B ⌉ , ⌈ B + z ⌉ = ⌈ B ⌉ + z \lceil B+z\rceil=\lceil B\rceil+z ⌈ B + z ⌉ = ⌈ B ⌉ + z for z ∈ Z s a z\in\mathcal{Z}_{sa} z ∈ Z s a , ⌈ z B ⌉ = z ⌈ B ⌉ \lceil zB\rceil=z\lceil B\rceil ⌈ z B ⌉ = z ⌈ B ⌉ for z ∈ Z + + z\in\mathcal{Z}_{++} z ∈ Z ++ , and ⌈ U B U ∗ ⌉ = ⌈ B ⌉ \lceil UBU^*\rceil=\lceil B\rceil ⌈ U B U ∗ ⌉ = ⌈ B ⌉ for unitaries U ∈ M U\in\mathcal{M} U ∈ M ;
if B ∈ M + + B\in\mathcal{M}_{++} B ∈ M ++ , then ⌈ B ⌉ , ⌊ B ⌋ ∈ Z + + \lceil B\rceil,\lfloor B\rfloor\in\mathcal{Z}_{++} ⌈ B ⌉ , ⌊ B ⌋ ∈ Z ++ , ⌈ B s ⌉ = ⌈ B ⌉ s \lceil B^s\rceil=\lceil B\rceil^s ⌈ B s ⌉ = ⌈ B ⌉ s and ⌊ B s ⌋ = ⌊ B ⌋ s \lfloor B^s\rfloor=\lfloor B\rfloor^s ⌊ B s ⌋ = ⌊ B ⌋ s for s > 0 s>0 s > 0 , and ⌊ B ⌋ = ⌈ B − 1 ⌉ − 1 \lfloor B\rfloor=\lceil B^{-1}\rceil^{-1} ⌊ B ⌋ = ⌈ B − 1 ⌉ − 1 ;
if M \mathcal{M} M is a factor, then ⌈ B ⌉ = max σ ( B ) \lceil B\rceil=\max\sigma(B) ⌈ B ⌉ = max σ ( B ) and ⌊ B ⌋ = min σ ( B ) \lfloor B\rfloor=\min\sigma(B) ⌊ B ⌋ = min σ ( B ) ; if M \mathcal{M} M is abelian, then ⌈ B ⌉ = ⌊ B ⌋ = B \lceil B\rceil=\lfloor B\rfloor=B ⌈ B ⌉ = ⌊ B ⌋ = B ; for a projection e e e , ⌈ e ⌉ = c ( e ) \lceil e\rceil=c(e) ⌈ e ⌉ = c ( e ) .
Proof. The set S = { z ∈ Z s a : z ≥ B } S=\{z\in\mathcal{Z}_{sa}:z\ge B\} S = { z ∈ Z s a : z ≥ B } is nonempty, bounded below and directed downwards: for z 1 , z 2 ∈ S z_1,z_2\in S z 1 , z 2 ∈ S and the central projection q = 1 ( − ∞ , 0 ] ( z 1 − z 2 ) q=\mathbf 1_{(-\infty,0]}(z_1-z_2) q = 1 ( − ∞ , 0 ] ( z 1 − z 2 ) , the element z 1 q + z 2 ( 1 − q ) z_1q+z_2(1-q) z 1 q + z 2 ( 1 − q ) lies in S S S , since q q q commutes with B B B . A decreasing net in Z s a \mathcal{Z}_{sa} Z s a converges strongly to its infimum, which is central and still dominates B B B . The statements in env:a follow from the definition; for instance w ≥ z B w\ge zB w ≥ z B iff z − 1 w ≥ B z^{-1}w\ge B z − 1 w ≥ B for z ∈ Z + + z\in\mathcal{Z}_{++} z ∈ Z ++ . For env:b : B ≥ ∥ B − 1 ∥ − 1 B\ge\|B^{-1}\|^{-1} B ≥ ∥ B − 1 ∥ − 1 gives ⌊ B ⌋ ≥ ∥ B − 1 ∥ − 1 \lfloor B\rfloor\ge\|B^{-1}\|^{-1} ⌊ B ⌋ ≥ ∥ B − 1 ∥ − 1 ; for central w > 0 w>0 w > 0 , w w w and B B B commute, so w ≥ B w\ge B w ≥ B iff w s ≥ B s w^s\ge B^s w s ≥ B s , and w ≤ B w\le B w ≤ B iff w − 1 ≥ B − 1 w^{-1}\ge B^{-1} w − 1 ≥ B − 1 . For env:c , only the last claim needs comment: if z ∈ Z z\in\mathcal{Z} z ∈ Z , z ≥ e z\ge e z ≥ e , and q = 1 ( − ∞ , 1 − t ) ( z ) q=\mathbf 1_{(-\infty,1-t)}(z) q = 1 ( − ∞ , 1 − t ) ( z ) with t > 0 t>0 t > 0 , then e q ≤ ( 1 − t ) q eq\le(1-t)q e q ≤ ( 1 − t ) q forces the projection e q eq e q to vanish, so q ≤ 1 − c ( e ) q\le1-c(e) q ≤ 1 − c ( e ) and z ≥ ( 1 − t ) c ( e ) z\ge(1-t)c(e) z ≥ ( 1 − t ) c ( e ) (note z ≥ e ≥ 0 z\ge e\ge0 z ≥ e ≥ 0 ); hence z ≥ c ( e ) z\ge c(e) z ≥ c ( e ) , and since c ( e ) ≥ e c(e)\ge e c ( e ) ≥ e is central, ⌈ e ⌉ = c ( e ) \lceil e\rceil=c(e) ⌈ e ⌉ = c ( e ) . ◻
Lemma 4.2 (Local attainment). Let B ∈ M s a B\in\mathcal{M}_{sa} B ∈ M s a , ω 0 ∈ Ω \omega_0\in\Omega ω 0 ∈ Ω , ε > 0 \varepsilon>0 ε > 0 , α = ⌈ B ⌉ ( ω 0 ) \alpha=\lceil B\rceil(\omega_0) α = ⌈ B ⌉ ( ω 0 ) and β = ⌊ B ⌋ ( ω 0 ) \beta=\lfloor B\rfloor(\omega_0) β = ⌊ B ⌋ ( ω 0 ) . Let P P P be a central projection whose clopen set contains ω 0 \omega_0 ω 0 and on which ⌈ B ⌉ \lceil B\rceil ⌈ B ⌉ and ⌊ B ⌋ \lfloor B\rfloor ⌊ B ⌋ oscillate by less than ε \varepsilon ε (such P P P exists). Then the spectral projections E = 1 [ α − 2 ε , ∞ ) ( B ) P E=\mathbf 1_{[\alpha-2\varepsilon,\infty)}(B)P E = 1 [ α − 2 ε , ∞ ) ( B ) P and F = 1 ( − ∞ , β + 2 ε ] ( B ) P F=\mathbf 1_{(-\infty,\beta+2\varepsilon]}(B)P F = 1 ( − ∞ , β + 2 ε ] ( B ) P satisfy c ( E ) = c ( F ) = P c(E)=c(F)=P c ( E ) = c ( F ) = P .
Proof. E E E commutes with Z \mathcal{Z} Z , so c ( E ) ≤ P c(E)\le P c ( E ) ≤ P . If Q ≤ P Q\le P Q ≤ P is a nonzero central projection with E Q = 0 EQ=0 E Q = 0 , then B Q ≤ ( α − 2 ε ) Q BQ\le(\alpha-2\varepsilon)Q B Q ≤ ( α − 2 ε ) Q , so w = ⌈ B ⌉ ( 1 − Q ) + ( α − 2 ε ) Q w=\lceil B\rceil(1-Q)+(\alpha-2\varepsilon)Q w = ⌈ B ⌉ ( 1 − Q ) + ( α − 2 ε ) Q is central and w ≥ B w\ge B w ≥ B , hence w ≥ ⌈ B ⌉ w\ge\lceil B\rceil w ≥ ⌈ B ⌉ ; on Q Q Q this gives α − 2 ε ≥ ⌈ B ⌉ > α − ε \alpha-2\varepsilon\ge\lceil B\rceil>\alpha-\varepsilon α − 2 ε ≥ ⌈ B ⌉ > α − ε , a contradiction. Apply this to − B -B − B for F F F . P P P exists by continuity of Gelfand transforms. ◻
Corollary 4.3. For B ∈ M s a B\in\mathcal{M}_{sa} B ∈ M s a , max σ ( B ) = max Ω ⌈ B ⌉ \max\sigma(B)=\max_\Omega\lceil B\rceil max σ ( B ) = max Ω ⌈ B ⌉ and min σ ( B ) = min Ω ⌊ B ⌋ \min\sigma(B)=\min_\Omega\lfloor B\rfloor min σ ( B ) = min Ω ⌊ B ⌋ .
Proof. B ≤ ⌈ B ⌉ ≤ max ⌈ B ⌉ B\le\lceil B\rceil\le\max\lceil B\rceil B ≤ ⌈ B ⌉ ≤ max ⌈ B ⌉ ; conversely, at a maximum point of ⌈ B ⌉ \lceil B\rceil ⌈ B ⌉ the projection E E E of Lemma 4.2 is nonzero for every ε \varepsilon ε . Apply this to − B -B − B as well. ◻
Definition 4.4. For B ∈ M + + B\in\mathcal{M}_{++} B ∈ M ++ let u = ⌈ B ⌉ u=\lceil B\rceil u = ⌈ B ⌉ , l = ⌊ B ⌋ l=\lfloor B\rfloor l = ⌊ B ⌋ and
κ Z ( B ) = ∥ u l − 1 ∥ = max Ω u l , L N C ( B ) = 1 2 ln κ Z ( B ) , N C ( B ) = tanh L N C ( B ) = κ Z ( B ) − 1 κ Z ( B ) + 1 . \kappa_\mathcal{Z}(B)=\|ul^{-1}\|=\max_\Omega\frac ul,\qquad \mathop{\mathrm{LNC}}(B)=\tfrac12\ln\kappa_\mathcal{Z}(B),\qquad
\mathop{\mathrm{NC}}(B)=\tanh\mathop{\mathrm{LNC}}(B)=\frac{\kappa_\mathcal{Z}(B)-1}{\kappa_\mathcal{Z}(B)+1}. κ Z ( B ) = ∥ u l − 1 ∥ = Ω max l u , LNC ( B ) = 2 1 ln κ Z ( B ) , NC ( B ) = tanh LNC ( B ) = κ Z ( B ) + 1 κ Z ( B ) − 1 .
By Corollary 4.3 , κ ( B ) = max Ω u / min Ω l \kappa(B)=\max_\Omega u/\min_\Omega l κ ( B ) = max Ω u / min Ω l , whereas κ Z ( B ) = max Ω ( u / l ) \kappa_\mathcal{Z}(B)=\max_\Omega(u/l) κ Z ( B ) = max Ω ( u / l ) .
Proposition 4.5. Let B ∈ M + + B\in\mathcal{M}_{++} B ∈ M ++ , u = ⌈ B ⌉ u=\lceil B\rceil u = ⌈ B ⌉ , l = ⌊ B ⌋ l=\lfloor B\rfloor l = ⌊ B ⌋ and z 0 = ( u l ) − 1 / 2 z_0=(ul)^{-1/2} z 0 = ( u l ) − 1/2 .
κ Z ( B ) = min { κ ( z B ) : z ∈ Z + + } \kappa_\mathcal{Z}(B)=\min\{\kappa(zB):z\in\mathcal{Z}_{++}\} κ Z ( B ) = min { κ ( z B ) : z ∈ Z ++ } , attained at z = z 0 z=z_0 z = z 0 .
L N C ( B ) = min { d T ( B , w ) : w ∈ Z + + } \mathop{\mathrm{LNC}}(B)=\min\{d_T(B,w):w\in\mathcal{Z}_{++}\} LNC ( B ) = min { d T ( B , w ) : w ∈ Z ++ } , attained at w = ( u l ) 1 / 2 w=(ul)^{1/2} w = ( u l ) 1/2 .
N C ( B ) = min { ∥ 1 − z B ∥ : z ∈ Z + } \mathop{\mathrm{NC}}(B)=\min\{\|1-zB\|:z\in\mathcal{Z}_+\} NC ( B ) = min { ∥1 − z B ∥ : z ∈ Z + } , attained at z = 2 ( u + l ) − 1 z=2(u+l)^{-1} z = 2 ( u + l ) − 1 .
κ Z ( B ) ≥ 1 \kappa_\mathcal{Z}(B)\ge1 κ Z ( B ) ≥ 1 with equality iff B ∈ Z B\in\mathcal{Z} B ∈ Z ; κ Z ( z B ) = κ Z ( B − 1 ) = κ Z ( U B U ∗ ) = κ Z ( B ) \kappa_\mathcal{Z}(zB)=\kappa_\mathcal{Z}(B^{-1})=\kappa_\mathcal{Z}(UBU^*)=\kappa_\mathcal{Z}(B) κ Z ( z B ) = κ Z ( B − 1 ) = κ Z ( U B U ∗ ) = κ Z ( B ) and κ Z ( B s ) = κ Z ( B ) s \kappa_\mathcal{Z}(B^s)=\kappa_\mathcal{Z}(B)^s κ Z ( B s ) = κ Z ( B ) s for z ∈ Z + + z\in\mathcal{Z}_{++} z ∈ Z ++ , unitaries U ∈ M U\in\mathcal{M} U ∈ M and s > 0 s>0 s > 0 ; κ Z ( B ) ≤ κ ( B ) \kappa_\mathcal{Z}(B)\le\kappa(B) κ Z ( B ) ≤ κ ( B ) , with equality if M \mathcal{M} M is a factor; κ Z ≡ 1 \kappa_\mathcal{Z}\equiv1 κ Z ≡ 1 if M \mathcal{M} M is abelian; and κ Z ( ⨁ i B i ) = sup i κ Z ( B i ) \kappa_\mathcal{Z}(\bigoplus_iB_i)=\sup_i\kappa_\mathcal{Z}(B_i) κ Z ( ⨁ i B i ) = sup i κ Z ( B i ) in ⨁ i M i \bigoplus_i\mathcal{M}_i ⨁ i M i .
Proof. ch:resc By Lemma 4.1 and Corollary 4.3 , ∥ z B ∥ = max z u \|zB\|=\max zu ∥ z B ∥ = max z u and ∥ ( z B ) − 1 ∥ = max ( z l ) − 1 \|(zB)^{-1}\|=\max(zl)^{-1} ∥ ( z B ) − 1 ∥ = max ( z l ) − 1 , and max ( z u ) ⋅ max ( z l ) − 1 ≥ max ( u / l ) \max(zu)\cdot\max(zl)^{-1}\ge\max(u/l) max ( z u ) ⋅ max ( z l ) − 1 ≥ max ( u / l ) , with equality for z = z 0 z=z_0 z = z 0 , where z 0 u = ( z 0 l ) − 1 = ( u / l ) 1 / 2 z_0u=(z_0l)^{-1}=(u/l)^{1/2} z 0 u = ( z 0 l ) − 1 = ( u / l ) 1/2 . ch:thomp For w ∈ Z + + w\in\mathcal{Z}_{++} w ∈ Z ++ , B ≤ t w B\le tw B ≤ tw iff u ≤ t w u\le tw u ≤ tw , and w ≤ s B w\le sB w ≤ s B iff w ≤ s l w\le sl w ≤ s l , so M ( B / w ) = ∥ u w − 1 ∥ M(B/w)=\|uw^{-1}\| M ( B / w ) = ∥ u w − 1 ∥ and M ( w / B ) = ∥ w l − 1 ∥ M(w/B)=\|wl^{-1}\| M ( w / B ) = ∥ w l − 1 ∥ ; the larger of the two is at least ∥ u l − 1 ∥ 1 / 2 \|ul^{-1}\|^{1/2} ∥ u l − 1 ∥ 1/2 , with equality for w = ( u l ) 1 / 2 w=(ul)^{1/2} w = ( u l ) 1/2 . ch:mich For z ∈ Z + z\in\mathcal{Z}_+ z ∈ Z + , z l ≤ z B ≤ z u zl\le zB\le zu z l ≤ z B ≤ z u gives ∥ 1 − z B ∥ ≤ max Ω max { ∣ 1 − z u ∣ , ∣ 1 − z l ∣ } \|1-zB\|\le\max_\Omega\max\{|1-zu|,|1-zl|\} ∥1 − z B ∥ ≤ max Ω max { ∣1 − z u ∣ , ∣1 − z l ∣ } . Conversely, if ∣ 1 − z u ∣ |1-zu| ∣1 − z u ∣ is maximal at ω 0 \omega_0 ω 0 and P , E P,E P , E are as in Lemma 4.2 with z z z also oscillating by less than ε \varepsilon ε on P P P , then for a unit vector ξ ∈ E H \xi\in E\mathcal{H} ξ ∈ E H we have ⟨ B ξ , ξ ⟩ ∈ [ α − 2 ε , α + ε ] \langle B\xi,\xi\rangle\in[\alpha-2\varepsilon,\alpha+\varepsilon] ⟨ B ξ , ξ ⟩ ∈ [ α − 2 ε , α + ε ] and ⟨ z B ξ , ξ ⟩ = z ( ω 0 ) ⟨ B ξ , ξ ⟩ + O ( ε ) \langle zB\xi,\xi\rangle=z(\omega_0)\langle B\xi,\xi\rangle+O(\varepsilon) ⟨ z B ξ , ξ ⟩ = z ( ω 0 ) ⟨ B ξ , ξ ⟩ + O ( ε ) , so ∥ 1 − z B ∥ ≥ ∣ 1 − z ( ω 0 ) u ( ω 0 ) ∣ − O ( ε ) \|1-zB\|\ge|1-z(\omega_0)u(\omega_0)|-O(\varepsilon) ∥1 − z B ∥ ≥ ∣1 − z ( ω 0 ) u ( ω 0 ) ∣ − O ( ε ) ; similarly for l l l . Hence ∥ 1 − z B ∥ = max Ω max { ∣ 1 − z u ∣ , ∣ 1 − z l ∣ } \|1-zB\|=\max_\Omega\max\{|1-zu|,|1-zl|\} ∥1 − z B ∥ = max Ω max { ∣1 − z u ∣ , ∣1 − z l ∣ } , and pointwise max { ∣ 1 − t u ∣ , ∣ 1 − t l ∣ } ≥ ( u − l ) / ( u + l ) \max\{|1-tu|,|1-tl|\}\ge(u-l)/(u+l) max { ∣1 − t u ∣ , ∣1 − tl ∣ } ≥ ( u − l ) / ( u + l ) with equality at t = 2 / ( u + l ) t=2/(u+l) t = 2/ ( u + l ) . ch:prop If u = l u=l u = l , then B = u ∈ Z B=u\in\mathcal{Z} B = u ∈ Z . The invariances follow from Lemma 4.1 (for B − 1 B^{-1} B − 1 : ⌈ B − 1 ⌉ = l − 1 \lceil B^{-1}\rceil=l^{-1} ⌈ B − 1 ⌉ = l − 1 , ⌊ B − 1 ⌋ = u − 1 \lfloor B^{-1}\rfloor=u^{-1} ⌊ B − 1 ⌋ = u − 1 ). The rest follows from max ( u / l ) ≤ max u / min l \max(u/l)\le\max u/\min l max ( u / l ) ≤ max u / min l , Lemma 4.1 env:c , and the fact that envelopes are computed summandwise. ◻
The transfer principle
Violation constants
We fix the class of quantities to which the transfer principle applies. An admissible expression in noncommuting variables X 1 , … , X r X_1,\dots,X_r X 1 , … , X r is built from the variables by sums, scalar multiples, products, adjoints, and Borel functions g : R → C g:\mathbb{R}\to\mathbb{C} g : R → C with g ( 0 ) = 0 g(0)=0 g ( 0 ) = 0 applied to self-adjoint expressions. Examples are ∣ X 1 ∣ p |X_1|^p ∣ X 1 ∣ p , X 1 k X_1^k X 1 k , ∣ X 1 X 2 ∣ |X_1X_2| ∣ X 1 X 2 ∣ , s ( X 1 ) s(X_1) s ( X 1 ) for X 1 ≥ 0 X_1\ge0 X 1 ≥ 0 , and X 1 ∨ X 2 = s ( X 1 + X 2 ) X_1\vee X_2=s(X_1+X_2) X 1 ∨ X 2 = s ( X 1 + X 2 ) for projections.
Definition 4.11. A violation constant is given by a test class and two functionals. The test class assigns to every finite von Neumann algebra M \mathcal{M} M the set T ( M ) ⊂ M r \mathcal{T}(\mathcal{M})\subset\mathcal{M}^r T ( M ) ⊂ M r of tuples satisfying finitely many conditions of the form X i ∈ M s a X_i\in\mathcal{M}_{sa} X i ∈ M s a , X i ∈ M + X_i\in\mathcal{M}_+ X i ∈ M + , X i X_i X i is a projection, X i ≤ X j X_i\le X_j X i ≤ X j . The functionals are N φ ( X ) = Ψ N ( φ ( F 1 ( X ) ) , … , φ ( F s ( X ) ) ) N_\varphi(X)=\Psi_N(\varphi(F_1(X)),\dots,\varphi(F_s(X))) N φ ( X ) = Ψ N ( φ ( F 1 ( X )) , … , φ ( F s ( X ))) and D φ ( X ) = Ψ D ( φ ( F 1 ( X ) ) , … , φ ( F s ( X ) ) ) D_\varphi(X)=\Psi_D(\varphi(F_1(X)),\dots,\varphi(F_s(X))) D φ ( X ) = Ψ D ( φ ( F 1 ( X )) , … , φ ( F s ( X ))) , where the F j F_j F j are admissible expressions and Ψ N : C s → R \Psi_N:\mathbb{C}^s\to\mathbb{R} Ψ N : C s → R , Ψ D : C s → [ 0 , ∞ ) \Psi_D:\mathbb{C}^s\to[0,\infty) Ψ D : C s → [ 0 , ∞ ) are continuous and positively homogeneous of the same degree; for instance, for the constant γ \gamma γ of Theorem (?) one takes Ψ N ( z ) = − R e z 1 \Psi_N(z)=-\mathop{\mathrm{Re}}z_1 Ψ N ( z ) = − Re z 1 and Ψ D ( z ) = ( ∣ z 2 ∣ ∣ z 3 ∣ ) 1 / 2 \Psi_D(z)=(|z_2||z_3|)^{1/2} Ψ D ( z ) = ( ∣ z 2 ∣∣ z 3 ∣ ) 1/2 with F 1 = X 1 X 2 F_1=X_1X_2 F 1 = X 1 X 2 , F 2 = X 1 2 F_2=X_1^2 F 2 = X 1 2 , F 3 = X 2 2 F_3=X_2^2 F 3 = X 2 2 . For a faithful normal positive functional φ \varphi φ on M \mathcal{M} M ,
c M ( φ ) = sup { N φ ( X ) / D φ ( X ) : X ∈ T ( M ) , D φ ( X ) > 0 } , c_\mathcal{M}(\varphi)=\sup\{N_\varphi(X)/D_\varphi(X):X\in\mathcal{T}(\mathcal{M}),\ D_\varphi(X)>0\}, c M ( φ ) = sup { N φ ( X ) / D φ ( X ) : X ∈ T ( M ) , D φ ( X ) > 0 } , and c M ( ρ ) = c M ( φ ρ ) c_\mathcal{M}(\rho)=c_\mathcal{M}(\varphi_\rho) c M ( ρ ) = c M ( φ ρ ) for ρ ∈ M + + \rho\in\mathcal{M}_{++} ρ ∈ M ++ . By homogeneity c M ( s φ ) = c M ( φ ) c_\mathcal{M}(s\varphi)=c_\mathcal{M}(\varphi) c M ( s φ ) = c M ( φ ) for s > 0 s>0 s > 0 , so the normalisation of τ \tau τ is irrelevant; on M n ( C ) M_n(\mathbb{C}) M n ( C ) we use T r \mathop{\mathrm{Tr}} Tr .
The constant m exp m_{\exp} m e x p of § 5.7 and the exponential constants of § 5.12 fall outside this definition, since exp ( 0 ) ≠ 0 \exp(0)\ne0 exp ( 0 ) = 0 , and so does the spectral geometric mean (item 6), whose variables are invertible; they are treated directly. Condition (R) below and Lemma 4.12 apply verbatim to these suprema, and the lower bounds of Theorem 4.15 t:low are obtained there by a direct modification of its proof.
The constant satisfies the rescaling condition (R) if for every z ∈ Z + + z\in\mathcal{Z}_{++} z ∈ Z ++ there is a bijection σ z \sigma_z σ z of T ( M ) \mathcal{T}(\mathcal{M}) T ( M ) with N z ρ ( X ) = N ρ ( σ z X ) N_{z\rho}(X)=N_\rho(\sigma_zX) N z ρ ( X ) = N ρ ( σ z X ) and D z ρ ( X ) = D ρ ( σ z X ) D_{z\rho}(X)=D_\rho(\sigma_zX) D z ρ ( X ) = D ρ ( σ z X ) for all ρ \rho ρ and X X X .
Lemma 4.12 (Central rescaling). If (R) holds, then c M ( z ρ ) = c M ( ρ ) c_\mathcal{M}(z\rho)=c_\mathcal{M}(\rho) c M ( z ρ ) = c M ( ρ ) for all z ∈ Z + + z\in\mathcal{Z}_{++} z ∈ Z ++ . In particular c M ( ρ ) = c M ( z 0 ρ ) c_\mathcal{M}(\rho)=c_\mathcal{M}(z_0\rho) c M ( ρ ) = c M ( z 0 ρ ) , where z 0 = ( ⌈ ρ ⌉ ⌊ ρ ⌋ ) − 1 / 2 z_0=(\lceil \rho\rceil\lfloor \rho\rfloor)^{-1/2} z 0 = (⌈ ρ ⌉ ⌊ ρ ⌋ ) − 1/2 and κ ( z 0 ρ ) = κ Z ( ρ ) \kappa(z_0\rho)=\kappa_\mathcal{Z}(\rho) κ ( z 0 ρ ) = κ Z ( ρ ) .
Proof. The first claim is immediate from (R); the second is Proposition 4.5 ch:resc . ◻
The corner construction
Lemma 4.13. Let π : M n ( C ) → M \pi:M_n(\mathbb{C})\to\mathcal{M} π : M n ( C ) → M be an injective ∗ * ∗ -homomorphism and e = π ( 1 ) e=\pi(1) e = π ( 1 ) . Then g ( π ( x ) ) = π ( g ( x ) ) g(\pi(x))=\pi(g(x)) g ( π ( x )) = π ( g ( x )) for every self-adjoint x ∈ M n ( C ) x\in M_n(\mathbb{C}) x ∈ M n ( C ) and every Borel function g g g with g ( 0 ) = 0 g(0)=0 g ( 0 ) = 0 ; consequently F ( π ( X 1 ) , … , π ( X r ) ) = π ( F ( X 1 , … , X r ) ) F(\pi(X_1),\dots,\pi(X_r))=\pi(F(X_1,\dots,X_r)) F ( π ( X 1 ) , … , π ( X r )) = π ( F ( X 1 , … , X r )) for every admissible expression F F F . Moreover π ( x ) ≥ 0 \pi(x)\ge0 π ( x ) ≥ 0 iff x ≥ 0 x\ge0 x ≥ 0 , π ( x ) \pi(x) π ( x ) is a projection iff x x x is, and φ ∘ π \varphi\circ\pi φ ∘ π is a faithful positive functional on M n ( C ) M_n(\mathbb{C}) M n ( C ) for every faithful positive φ \varphi φ on M \mathcal{M} M .
Proof. Write x = ∑ λ λ q λ x=\sum_\lambda\lambda q_\lambda x = ∑ λ λ q λ with spectral projections q λ q_\lambda q λ . The π ( q λ ) \pi(q_\lambda) π ( q λ ) are orthogonal projections with sum e e e , so π ( x ) = ∑ λ π ( q λ ) + 0 ⋅ ( 1 − e ) \pi(x)=\sum\lambda\pi(q_\lambda)+0\cdot(1-e) π ( x ) = ∑ λπ ( q λ ) + 0 ⋅ ( 1 − e ) and g ( π ( x ) ) = ∑ g ( λ ) π ( q λ ) = π ( g ( x ) ) g(\pi(x))=\sum g(\lambda)\pi(q_\lambda)=\pi(g(x)) g ( π ( x )) = ∑ g ( λ ) π ( q λ ) = π ( g ( x )) . The remaining claims follow from this and the injectivity of π \pi π . ◻
Lemma 4.14 (Corner construction). Let a ∈ M + + ∖ Z a\in\mathcal{M}_{++}\setminus\mathcal{Z} a ∈ M ++ ∖ Z , let ω 0 ∈ Ω \omega_0\in\Omega ω 0 ∈ Ω be a point where ⌈ a ⌉ / ⌊ a ⌋ \lceil a\rceil/\lfloor a\rfloor ⌈ a ⌉ / ⌊ a ⌋ attains its maximum κ Z ( a ) \kappa_\mathcal{Z}(a) κ Z ( a ) , α = ⌈ a ⌉ ( ω 0 ) \alpha=\lceil a\rceil(\omega_0) α = ⌈ a ⌉ ( ω 0 ) , β = ⌊ a ⌋ ( ω 0 ) \beta=\lfloor a\rfloor(\omega_0) β = ⌊ a ⌋ ( ω 0 ) , so that α / β = κ Z ( a ) \alpha/\beta=\kappa_\mathcal{Z}(a) α / β = κ Z ( a ) , and let 0 < ε < ( α − β ) / 4 0<\varepsilon<(\alpha-\beta)/4 0 < ε < ( α − β ) /4 , I 1 = [ α − 2 ε , α + ε ] I_1=[\alpha-2\varepsilon,\alpha+\varepsilon] I 1 = [ α − 2 ε , α + ε ] , I 2 = [ β − ε , β + 2 ε ] I_2=[\beta-\varepsilon,\beta+2\varepsilon] I 2 = [ β − ε , β + 2 ε ] .
There is a nonzero partial isometry v ∈ M v\in\mathcal{M} v ∈ M with f 1 : = v v ∗ ≤ 1 I 1 ( a ) f_1:=vv^*\le\mathbf 1_{I_1}(a) f 1 := v v ∗ ≤ 1 I 1 ( a ) and f 2 : = v ∗ v ≤ 1 I 2 ( a ) f_2:=v^*v\le\mathbf 1_{I_2}(a) f 2 := v ∗ v ≤ 1 I 2 ( a ) . Hence f 1 f 2 = 0 f_1f_2=0 f 1 f 2 = 0 , f 1 g ( a ) f 2 = 0 f_1g(a)f_2=0 f 1 g ( a ) f 2 = 0 for every bounded Borel g g g , and ( inf I i g ) f i ≤ f i g ( a ) f i ≤ ( sup I i g ) f i (\inf_{I_i}g)f_i\le f_ig(a)f_i\le(\sup_{I_i}g)f_i ( inf I i g ) f i ≤ f i g ( a ) f i ≤ ( sup I i g ) f i for real g g g .
e 11 = f 1 e_{11}=f_1 e 11 = f 1 , e 12 = v e_{12}=v e 12 = v , e 21 = v ∗ e_{21}=v^* e 21 = v ∗ , e 22 = f 2 e_{22}=f_2 e 22 = f 2 are matrix units, and π ( w ) = ∑ i , j w i j e i j \pi(w)=\sum_{i,j}w_{ij}e_{ij} π ( w ) = ∑ i , j w ij e ij is an injective ∗ * ∗ -homomorphism M 2 ( C ) → M M_2(\mathbb{C})\to\mathcal{M} M 2 ( C ) → M . For a unit vector ξ ∈ f 2 H \xi\in f_2\mathcal{H} ξ ∈ f 2 H the vectors η 1 = v ξ \eta_1=v\xi η 1 = v ξ , η 2 = ξ \eta_2=\xi η 2 = ξ are orthonormal, π ( w ) η j = ∑ i w i j η i \pi(w)\eta_j=\sum_iw_{ij}\eta_i π ( w ) η j = ∑ i w ij η i , ⟨ g ( a ) η 1 , η 2 ⟩ = 0 \langle g(a)\eta_1,\eta_2\rangle=0 ⟨ g ( a ) η 1 , η 2 ⟩ = 0 and ⟨ g ( a ) η i , η i ⟩ ∈ [ inf I i g , sup I i g ] \langle g(a)\eta_i,\eta_i\rangle\in[\inf_{I_i}g,\sup_{I_i}g] ⟨ g ( a ) η i , η i ⟩ ∈ [ inf I i g , sup I i g ] .
If M \mathcal{M} M is finite, then τ ( f 1 ) = τ ( f 2 ) = : t > 0 \tau(f_1)=\tau(f_2)=:t>0 τ ( f 1 ) = τ ( f 2 ) =: t > 0 and τ ( g ( a ) π ( w ) ) = t ( r 1 w 11 + r 2 w 22 ) \tau(g(a)\pi(w))=t\,(r_1w_{11}+r_2w_{22}) τ ( g ( a ) π ( w )) = t ( r 1 w 11 + r 2 w 22 ) with r i = τ ( g ( a ) f i ) / t ∈ [ inf I i g , sup I i g ] r_i=\tau(g(a)f_i)/t\in[\inf_{I_i}g,\sup_{I_i}g] r i = τ ( g ( a ) f i ) / t ∈ [ inf I i g , sup I i g ] .
If M \mathcal{M} M is a II1 _1 1 factor, λ 1 , … , λ n ∈ σ ( a ) \lambda_1,\dots,\lambda_n\in\sigma(a) λ 1 , … , λ n ∈ σ ( a ) and ε > 0 \varepsilon>0 ε > 0 , then there are matrix units { e i j } i , j = 1 n ⊂ M \{e_{ij}\}_{i,j=1}^n\subset\mathcal{M} { e ij } i , j = 1 n ⊂ M with e i i ≤ 1 [ λ i − ε , λ i + ε ] ( a ) e_{ii}\le\mathbf 1_{[\lambda_i-\varepsilon,\lambda_i+\varepsilon]}(a) e ii ≤ 1 [ λ i − ε , λ i + ε ] ( a ) , τ ( e i i ) = t > 0 \tau(e_{ii})=t>0 τ ( e ii ) = t > 0 and ∣ τ ( a e i j ) − δ i j λ i t ∣ ≤ ε t |\tau(ae_{ij})-\delta_{ij}\lambda_it|\le\varepsilon t ∣ τ ( a e ij ) − δ ij λ i t ∣ ≤ εt .
Proof. c:a Choose P P P as in Lemma 4.2 for B = a B=a B = a . Since a P ≤ ⌈ a ⌉ P ≤ ( α + ε ) P aP\le\lceil a\rceil P\le(\alpha+\varepsilon)P a P ≤ ⌈ a ⌉ P ≤ ( α + ε ) P and a P ≥ ⌊ a ⌋ P ≥ ( β − ε ) P aP\ge\lfloor a\rfloor P\ge(\beta-\varepsilon)P a P ≥ ⌊ a ⌋ P ≥ ( β − ε ) P , the projections E , F E,F E , F of that lemma satisfy E ≤ 1 I 1 ( a ) E\le\mathbf 1_{I_1}(a) E ≤ 1 I 1 ( a ) and F ≤ 1 I 2 ( a ) F\le\mathbf 1_{I_2}(a) F ≤ 1 I 2 ( a ) ; they are orthogonal because I 1 ∩ I 2 = ∅ I_1\cap I_2=\emptyset I 1 ∩ I 2 = ∅ , and c ( E ) c ( F ) = P ≠ 0 c(E)c(F)=P\ne0 c ( E ) c ( F ) = P = 0 . The comparison fact gives v v v . Since E , F E,F E , F commute with a a a , f 1 g ( a ) f 2 = f 1 g ( a ) E F f 2 = 0 f_1g(a)f_2=f_1g(a)EFf_2=0 f 1 g ( a ) f 2 = f 1 g ( a ) E F f 2 = 0 , and f i g ( a ) f i = f i ( g ( a ) 1 I i ( a ) ) f i f_ig(a)f_i=f_i(g(a)\mathbf 1_{I_i}(a))f_i f i g ( a ) f i = f i ( g ( a ) 1 I i ( a )) f i gives the bounds. c:b is a direct computation: v ∗ v ξ = ξ v^*v\xi=\xi v ∗ v ξ = ξ , e i j = e i j f j e_{ij}=e_{ij}f_j e ij = e ij f j , and ⟨ g ( a ) η 1 , η 2 ⟩ = ⟨ f 2 g ( a ) f 1 η 1 , η 2 ⟩ = 0 \langle
g(a)\eta_1,\eta_2\rangle=\langle f_2g(a)f_1\eta_1,\eta_2\rangle=0 ⟨ g ( a ) η 1 , η 2 ⟩ = ⟨ f 2 g ( a ) f 1 η 1 , η 2 ⟩ = 0 . c:c τ ( v v ∗ ) = τ ( v ∗ v ) \tau(vv^*)=\tau(v^*v) τ ( v v ∗ ) = τ ( v ∗ v ) ; τ ( g ( a ) e 12 ) = τ ( g ( a ) f 1 v f 2 ) = τ ( f 2 g ( a ) f 1 v ) = 0 \tau(g(a)e_{12})=\tau(g(a)f_1vf_2)=\tau(f_2g(a)f_1v)=0 τ ( g ( a ) e 12 ) = τ ( g ( a ) f 1 v f 2 ) = τ ( f 2 g ( a ) f 1 v ) = 0 , and likewise for e 21 e_{21} e 21 . c:d Shrinking ε \varepsilon ε we may assume that the windows of distinct values among the λ i \lambda_i λ i are disjoint. For each distinct value μ \mu μ of multiplicity n μ n_\mu n μ the spectral projection E μ = 1 [ μ − ε , μ + ε ] ( a ) E_\mu=\mathbf 1_{[\mu-\varepsilon,\mu+\varepsilon]}(a) E μ = 1 [ μ − ε , μ + ε ] ( a ) is nonzero. Let t = min μ τ ( E μ ) / n μ t=\min_\mu\tau(E_\mu)/n_\mu t = min μ τ ( E μ ) / n μ . In a II1 _1 1 factor we can choose orthogonal subprojections of the E μ E_\mu E μ of trace t t t , n μ n_\mu n μ of them under E μ E_\mu E μ ; projections of equal trace are equivalent, which gives matrix units with e i i ≤ E λ i e_{ii}\le E_{\lambda_i} e ii ≤ E λ i . Then τ ( a e i j ) = τ ( a E λ i e i j ) = λ i τ ( e i j ) + τ ( ( a − λ i ) E λ i e i j ) \tau(ae_{ij})=\tau(aE_{\lambda_i}e_{ij})=\lambda_i\tau(e_{ij})+\tau((a-\lambda_i)E_{\lambda_i}e_{ij}) τ ( a e ij ) = τ ( a E λ i e ij ) = λ i τ ( e ij ) + τ (( a − λ i ) E λ i e ij ) , τ ( e i j ) = δ i j t \tau(e_{ij})=\delta_{ij}t τ ( e ij ) = δ ij t and ∣ τ ( Y e i j ) ∣ ≤ ∥ Y ∥ τ ( ∣ e i j ∣ ) = ∥ Y ∥ t |\tau(Ye_{ij})|\le\|Y\|\tau(|e_{ij}|)=\|Y\|t ∣ τ ( Y e ij ) ∣ ≤ ∥ Y ∥ τ ( ∣ e ij ∣ ) = ∥ Y ∥ t . ◻
The transfer theorem
Theorem 4.15. Let c c c be a violation constant.
For every finite M \mathcal{M} M and ρ ∈ M + + ∖ Z \rho\in\mathcal{M}_{++}\setminus\mathcal{Z} ρ ∈ M ++ ∖ Z , c M ( ρ ) ≥ c M 2 ( d i a g ( κ Z ( ρ ) , 1 ) ) c_\mathcal{M}(\rho)\ge c_{M_2}(\mathop{\mathrm{diag}}(\kappa_\mathcal{Z}(\rho),1)) c M ( ρ ) ≥ c M 2 ( diag ( κ Z ( ρ ) , 1 )) . If M \mathcal{M} M is a II1 _1 1 factor, then c M ( ρ ) ≥ c M n ( d i a g ( λ 1 , … , λ n ) ) c_\mathcal{M}(\rho)\ge c_{M_n}(\mathop{\mathrm{diag}}(\lambda_1,\dots,\lambda_n)) c M ( ρ ) ≥ c M n ( diag ( λ 1 , … , λ n )) for all λ 1 , … , λ n ∈ σ ( ρ ) \lambda_1,\dots,\lambda_n\in\sigma(\rho) λ 1 , … , λ n ∈ σ ( ρ ) .
Suppose (R) holds and f : [ 1 , ∞ ) → [ 0 , ∞ ] f:[1,\infty)\to[0,\infty] f : [ 1 , ∞ ) → [ 0 , ∞ ] satisfies c M ( ρ ) ≤ f ( κ ( ρ ) ) c_\mathcal{M}(\rho)\le f(\kappa(\rho)) c M ( ρ ) ≤ f ( κ ( ρ )) for all finite M \mathcal{M} M and all ρ ∈ M + + \rho\in\mathcal{M}_{++} ρ ∈ M ++ . Then c M ( ρ ) ≤ f ( κ Z ( ρ ) ) c_\mathcal{M}(\rho)\le f(\kappa_\mathcal{Z}(\rho)) c M ( ρ ) ≤ f ( κ Z ( ρ )) .
If, in addition, c M 2 ( d i a g ( κ , 1 ) ) ≥ f ( κ ) c_{M_2}(\mathop{\mathrm{diag}}(\kappa,1))\ge f(\kappa) c M 2 ( diag ( κ , 1 )) ≥ f ( κ ) for every κ > 1 \kappa>1 κ > 1 , and c M ( ρ ) ≥ f ( 1 ) c_\mathcal{M}(\rho)\ge f(1) c M ( ρ ) ≥ f ( 1 ) for central ρ \rho ρ , then c M ( ρ ) = f ( κ Z ( ρ ) ) c_\mathcal{M}(\rho)=f(\kappa_\mathcal{Z}(\rho)) c M ( ρ ) = f ( κ Z ( ρ )) for every finite M \mathcal{M} M and every ρ ∈ M + + \rho\in\mathcal{M}_{++} ρ ∈ M ++ . In particular c = f ( κ ) c=f(\kappa) c = f ( κ ) in every finite factor.
Proof. t:low Let π \pi π be the homomorphism of Lemma 4.14 c:b for a = ρ a=\rho a = ρ . By Lemma 4.14 c:c with g ( t ) = t g(t)=t g ( t ) = t , φ ∘ π = t T r ( R ε ⋅ ) \varphi\circ\pi=t\mathop{\mathrm{Tr}}(R_\varepsilon\,\cdot) φ ∘ π = t Tr ( R ε ⋅ ) with R ε = d i a g ( r 1 , r 2 ) → Λ = d i a g ( α , β ) R_\varepsilon=\mathop{\mathrm{diag}}(r_1,r_2)\to\Lambda=\mathop{\mathrm{diag}}(\alpha,\beta) R ε = diag ( r 1 , r 2 ) → Λ = diag ( α , β ) as ε → 0 \varepsilon\to0 ε → 0 . By Lemma 4.13 , π ( T ( M 2 ) ) ⊂ T ( M ) \pi(\mathcal{T}(M_2))\subset\mathcal{T}(\mathcal{M}) π ( T ( M 2 )) ⊂ T ( M ) , N φ ( π X ) = N φ ∘ π ( X ) N_\varphi(\pi X)=N_{\varphi\circ\pi}(X) N φ ( π X ) = N φ ∘ π ( X ) and D φ ( π X ) = D φ ∘ π ( X ) D_\varphi(\pi X)=D_{\varphi\circ\pi}(X) D φ ( π X ) = D φ ∘ π ( X ) . Hence, for every X ∈ T ( M 2 ) X\in\mathcal{T}(M_2) X ∈ T ( M 2 ) with D Λ ( X ) > 0 D_{\Lambda}(X)>0 D Λ ( X ) > 0 and all small ε \varepsilon ε , c M ( ρ ) ≥ N R ε ( X ) / D R ε ( X ) c_\mathcal{M}(\rho)\ge N_{R_\varepsilon}(X)/D_{R_\varepsilon}(X) c M ( ρ ) ≥ N R ε ( X ) / D R ε ( X ) , and the right side tends to N Λ ( X ) / D Λ ( X ) N_\Lambda(X)/D_\Lambda(X) N Λ ( X ) / D Λ ( X ) because N N N and D D D depend continuously on the functional. Thus c M ( ρ ) ≥ c M 2 ( Λ ) = c M 2 ( d i a g ( κ Z ( ρ ) , 1 ) ) c_\mathcal{M}(\rho)\ge c_{M_2}(\Lambda)=c_{M_2}(\mathop{\mathrm{diag}}(\kappa_\mathcal{Z}(\rho),1)) c M ( ρ ) ≥ c M 2 ( Λ ) = c M 2 ( diag ( κ Z ( ρ ) , 1 )) by homogeneity. In a II1 _1 1 factor use Lemma 4.14 c:d : φ ∘ π = t T r ( R ε ⋅ ) \varphi\circ\pi=t\mathop{\mathrm{Tr}}(R_\varepsilon\,\cdot) φ ∘ π = t Tr ( R ε ⋅ ) with ( R ε ) j i = τ ( ρ e i j ) / t → δ i j λ i (R_\varepsilon)_{ji}=\tau(\rho e_{ij})/t\to\delta_{ij}\lambda_i ( R ε ) j i = τ ( ρ e ij ) / t → δ ij λ i . t:up By Lemma 4.12 , c M ( ρ ) = c M ( z 0 ρ ) ≤ f ( κ ( z 0 ρ ) ) = f ( κ Z ( ρ ) ) c_\mathcal{M}(\rho)=c_\mathcal{M}(z_0\rho)\le f(\kappa(z_0\rho))=f(\kappa_\mathcal{Z}(\rho)) c M ( ρ ) = c M ( z 0 ρ ) ≤ f ( κ ( z 0 ρ )) = f ( κ Z ( ρ )) . t:eq combines t:low and t:up . ◻
In the applications all conditions are checked as follows: the constant is a violation constant in the sense of Definition 4.11 by inspection; (R) is verified by exhibiting σ z \sigma_z σ z ; the upper bound f ( κ ) f(\kappa) f ( κ ) is proved in an arbitrary finite von Neumann algebra; and an example in M 2 ( C ) M_2(\mathbb{C}) M 2 ( C ) gives the lower bound. For central ρ \rho ρ the value 1 = f ( 1 ) 1=f(1) 1 = f ( 1 ) is attained at a trivial tuple (e.g. X = 1 X=1 X = 1 ).
The catalogue of violation constants
This section determines the violation constants item by item. Where a constant is known in finite von Neumann algebras we prove it there, with one general upper bound and one 2 × 2 2\times2 2 × 2 example transferred by Theorem 4.15 ; the matrix envelopes follow by § 3.4 . With the operator variables free, the matrix envelopes are as follows, where a ≥ b > 0 a\ge b>0 a ≥ b > 0 are the extreme eigenvalues of ρ \rho ρ and
κ = a b = 1 + δ 1 − δ , G = a + b 2 a b = 1 1 − δ 2 = sec ψ , ψ = arcsin δ . \kappa=\frac ab=\frac{1+\delta}{1-\delta},\qquad G=\frac{a+b}{2\sqrt{ab}}=\frac1{\sqrt{1-\delta^2}}=\sec\psi,\qquad
\psi=\arcsin\delta . κ = b a = 1 − δ 1 + δ , G = 2 ab a + b = 1 − δ 2 1 = sec ψ , ψ = arcsin δ .
Theorem B 1. Let n ≥ 2 n\ge2 n ≥ 2 and δ = O M C ( ρ ) \delta=\mathop{\mathrm{OMC}}(\rho) δ = OMC ( ρ ) for densities ρ \rho ρ in M n M_n M n .
V n = G − 1 V_n=G-1 V n = G − 1 for Gardner’s inequality, the triangle inequalities for φ ( ∣ ⋅ ∣ ) \varphi(|\cdot|) φ ( ∣ ⋅ ∣ ) on operators and on projections, φ ( A + A ∗ ) ≤ 2 φ ( ∣ A ∗ ∣ ) \varphi(A+A^*)\le2\varphi(|A^*|) φ ( A + A ∗ ) ≤ 2 φ ( ∣ A ∗ ∣ ) , ∣ φ ( A 2 ) ∣ ≤ φ ( A ∗ A ) |\varphi(A^2)|\le\varphi(A^*A) ∣ φ ( A 2 ) ∣ ≤ φ ( A ∗ A ) , φ ( ∣ P − Q ∣ ) ≤ φ ( P + Q ) \varphi(|P-Q|)\le\varphi(P+Q) φ ( ∣ P − Q ∣ ) ≤ φ ( P + Q ) and φ ( R e P Q ) ≤ φ ( ∣ P Q ∣ ) \varphi(\mathrm{Re}\,PQ)\le\varphi(|PQ|) φ ( Re P Q ) ≤ φ ( ∣ P Q ∣ ) (items 1–5, 18, 22), in every dimension (Theorems 5.1 , 5.4 , (?) , Corollaries 5.3 , 5.6 , Proposition 5.33 ).
V n = G 2 − 1 V_n=G^2-1 V n = G 2 − 1 for φ ( P Q P ) ≤ φ ( Q ) \varphi(PQP)\le\varphi(Q) φ ( P QP ) ≤ φ ( Q ) ; V n = 1 2 tan 2 ( ψ / 2 ) V_n=\frac12\tan^2(\psi/2) V n = 2 1 tan 2 ( ψ /2 ) for φ ( R e P Q ) ≥ 0 \varphi(\mathrm{Re}\,PQ)\ge0 φ ( Re P Q ) ≥ 0 ; V n = δ / ( 1 − δ ) V_n=\delta/(1-\delta) V n = δ / ( 1 − δ ) for φ ( P ∨ Q ) ≤ φ ( P ) + φ ( Q ) \varphi(P\vee Q)\le\varphi(P)+\varphi(Q) φ ( P ∨ Q ) ≤ φ ( P ) + φ ( Q ) ; V n = δ V_n=\delta V n = δ for φ ( P + Q ) ≤ φ ( 1 + P ∧ Q ) \varphi(P+Q)\le\varphi(1+P\wedge Q) φ ( P + Q ) ≤ φ ( 1 + P ∧ Q ) (n n n even); V n V_n V n is the maximum of an explicit function of one variable, ( e − 1 ) 2 e ( e + 1 ) δ 2 + O ( δ 3 ) \frac{(e-1)^2}{e(e+1)}\delta^2+O(\delta^3) e ( e + 1 ) ( e − 1 ) 2 δ 2 + O ( δ 3 ) , for φ ( e Q P Q ) ≤ φ ( e P ) \varphi(e^{QPQ})\le\varphi(e^P) φ ( e QP Q ) ≤ φ ( e P ) (n n n even); and V n = ( G − 1 ) 2 / ( 4 G ) = δ 4 / 16 + O ( δ 6 ) V_n=(G-1)^2/(4G)=\delta^4/16+O(\delta^6) V n = ( G − 1 ) 2 / ( 4 G ) = δ 4 /16 + O ( δ 6 ) for the monotonicity 0 ≤ A ≤ B ⇒ φ ( A 2 ) ≤ φ ( B 2 ) 0\le A\le B\Rightarrow\varphi(A^2)\le\varphi(B^2) 0 ≤ A ≤ B ⇒ φ ( A 2 ) ≤ φ ( B 2 ) (Theorems 5.32 , 5.34 , 5.27 and Proposition 5.16 ).
V n = κ − 1 V_n=\kappa-1 V n = κ − 1 for the extension of Young’s inequality to arbitrary operators; V 2 = s / ( 1 + s ) V_2=s/(1+s) V 2 = s / ( 1 + s ) , s = sin ( ψ / 2 ) s=\sin(\psi/2) s = sin ( ψ /2 ) , for φ ( P Q + Q P ) ≤ 2 φ ( Q P Q ) \varphi(PQ+QP)\le2\varphi(QPQ) φ ( P Q + QP ) ≤ 2 φ ( QP Q ) ; for the inequalities that characterise the canonical trace, V 2 = 1 − 1 − δ 2 V_2=1-\sqrt{1-\delta^2} V 2 = 1 − 1 − δ 2 for the inequality ( det A ) 1 / n ≤ 1 n φ ( A ) (\det A)^{1/n}\le\frac1n\varphi(A) ( det A ) 1/ n ≤ n 1 φ ( A ) , increasing with n n n to 1 − 1 / S ( κ ) 1-1/S(\kappa) 1 − 1/ S ( κ ) with the Specht ratio S S S , and V n = ( ℓ + n − 1 ) / ( n ℓ ) − 1 V_n=\sqrt{(\ell+n-1)/(n\ell)}-1 V n = ( ℓ + n − 1 ) / ( n ℓ ) − 1 , ℓ = 1 / κ \ell=1/\kappa ℓ = 1/ κ , for φ ( A 2 ) 1 / 2 ≤ φ ( A ) \varphi(A^2)^{1/2}\le\varphi(A) φ ( A 2 ) 1/2 ≤ φ ( A ) (Corollary 5.20 , Propositions 5.17 , 5.36 and 5.37 ).
V n ≥ G − 1 V_n\ge G-1 V n ≥ G − 1 for the spectral geometric mean inequality, V n ≥ G 2 − 1 V_n\ge G^2-1 V n ≥ G 2 − 1 for the Araki–Lieb–Thirring inequality with ( r , q ) = ( 2 , 1 ) (r,q)=(2,1) ( r , q ) = ( 2 , 1 ) , and V n ≥ δ 8 / 4096 ( 1 + O ( δ 2 ) ) V_n\ge\delta^8/4096\,(1+O(\delta^2)) V n ≥ δ 8 /4096 ( 1 + O ( δ 2 )) for the weighted monotonicity with r = 1 r=1 r = 1 (Propositions 5.43 , (?) and 5.31 ); and V n ≤ κ − 1 V_n\le\kappa-1 V n ≤ κ − 1 for items 1–8, 10 and 25: for items 1–5 by (a), since G ≤ κ G\le\kappa G ≤ κ , for item 8 by (c) (Corollary 5.20 ), for item 10 by (b) (Theorem 5.27 ), and for items 6, 7 and 25 by Propositions 5.43 , 5.51 and 5.31 .
Figure 2. The ladder of orders: the computed envelopes V 2 ( δ ) V_2(\delta) V 2 ( δ ) of items 1–25 (certified lower bounds for the suprema, § 10 ) on logarithmic axes, coloured by the order of vanishing at the trace, with guide lines of slopes 1 1 1 , 2 2 2 , 4 4 4 and 8 8 8 . The horizontal axis is the contrast δ = O M C ( ρ ) \delta=\mathop{\mathrm{OMC}}(\rho) δ = OMC ( ρ ) of the density, written Δ ( W ) \Delta(W) Δ ( W ) in the figure. Item 25 is shown for δ ≥ 0.4 \delta\ge0.4 δ ≥ 0.4 (the search resolves it for δ ≥ 0.3 \delta\ge0.3 δ ≥ 0.3 ).
The proved envelopes thus vanish at the trace to different orders, δ \delta δ , δ 2 \delta^2 δ 2 and δ 4 \delta^4 δ 4 ; for item 25 the lower bound is of order δ 8 \delta^8 δ 8 , while only the trivial upper bound κ − 1 = 2 δ / ( 1 − δ ) \kappa-1=2\delta/(1-\delta) κ − 1 = 2 δ / ( 1 − δ ) is known (Proposition 5.31 ), and the computed envelope is consistent with g 25 g_{25} g 25 for δ ≥ 0.7 \delta\ge0.7 δ ≥ 0.7 (Figure 2 and Table 4 ). Two inequalities that are equivalent for traces may fail at different rates. For instance, φ ( P ∨ Q ) + φ ( P ∧ Q ) = φ ( P ) + φ ( Q ) \varphi(P\vee Q)+\varphi(P\wedge Q)=\varphi(P)+\varphi(Q) φ ( P ∨ Q ) + φ ( P ∧ Q ) = φ ( P ) + φ ( Q ) for every trace, yet the lattice subadditivity fails as δ / ( 1 − δ ) \delta/(1-\delta) δ / ( 1 − δ ) , while φ ( ∣ P − Q ∣ ) ≤ φ ( P + Q ) \varphi(|P-Q|)\le\varphi(P+Q) φ ( ∣ P − Q ∣ ) ≤ φ ( P + Q ) , with the same right-hand side, fails only as G − 1 ∼ δ 2 / 2 G-1\sim\delta^2/2 G − 1 ∼ δ 2 /2 . The orders refer to the normalisations N \mathcal{N} N of Table 3 ; a normalisation that vanishes at the extremisers can change them (Remark 5.42 ).
Throughout this section, except where M n M_n M n is named, M \mathcal{M} M is a von Neumann algebra with a faithful normal tracial state τ \tau τ , ρ ∈ M + + \rho\in\mathcal{M}_{++} ρ ∈ M ++ , φ = τ ( ρ ⋅ ) \varphi=\tau(\rho\,\cdot) φ = τ ( ρ ⋅ ) , κ = κ ( ρ ) \kappa=\kappa(\rho) κ = κ ( ρ ) , κ Z = κ Z ( ρ ) \kappa_\mathcal{Z}=\kappa_\mathcal{Z}(\rho) κ Z = κ Z ( ρ ) , u = ⌈ ρ ⌉ u=\lceil \rho\rceil u = ⌈ ρ ⌉ , l = ⌊ ρ ⌋ l=\lfloor \rho\rfloor l = ⌊ ρ ⌋ and G Z = cosh L N C ( ρ ) G_\mathcal{Z}=\cosh\mathop{\mathrm{LNC}}(\rho) G Z = cosh LNC ( ρ ) . All suprema are over tuples for which the denominator is nonzero. We repeatedly use that, for Y ∈ M + Y\in\mathcal{M}_+ Y ∈ M + ,
φ ( Y ) ≤ τ ( u Y ) and τ ( l Y ) ≤ φ ( Y ) , and u ≤ κ Z l , \begin{equation}
\varphi(Y)\le\tau(uY)\quad\text{and}\quad\tau(lY)\le\varphi(Y),\qquad\text{and}\quad u\le\kappa_\mathcal{Z}l, \tag*{(3)}
\end{equation} φ ( Y ) ≤ τ ( u Y ) and τ ( l Y ) ≤ φ ( Y ) , and u ≤ κ Z l , (3)
since τ ( ρ Y ) = τ ( Y 1 / 2 ρ Y 1 / 2 ) \tau(\rho Y)=\tau(Y^{1/2}\rho Y^{1/2}) τ ( ρ Y ) = τ ( Y 1/2 ρ Y 1/2 ) , l ≤ ρ ≤ u l\le\rho\le u l ≤ ρ ≤ u , and u / l ≤ κ Z u/l\le\kappa_\mathcal{Z} u / l ≤ κ Z on Ω \Omega Ω .
Gardner’s inequality
Theorem 5.1 (Item 1). sup X ≠ 0 ∣ φ ( X ) ∣ φ ( ∣ X ∣ ) = G Z = 1 + κ Z 2 κ Z \displaystyle\sup_{X\ne0}\frac{|\varphi(X)|}{\varphi(|X|)}=G_\mathcal{Z}=\frac{1+\kappa_\mathcal{Z}}{2\sqrt{\kappa_\mathcal{Z}}} X = 0 sup φ ( ∣ X ∣ ) ∣ φ ( X ) ∣ = G Z = 2 κ Z 1 + κ Z . In M n ( C ) M_n(\mathbb{C}) M n ( C ) the supremum is G ( ρ ) = cosh O L C ( ρ ) = sec ψ ( ρ ) G(\rho)=\cosh\mathop{\mathrm{OLC}}(\rho)=\sec\psi(\rho) G ( ρ ) = cosh OLC ( ρ ) = sec ψ ( ρ ) and is attained; for M n ( C ) M_n(\mathbb{C}) M n ( C ) cf. Bourin [113 Cor. 1.4, Rem. 1.9] and Remark 5.2 .
Proof. Upper bound with κ \kappa κ . Let X = U ∣ X ∣ X=U|X| X = U ∣ X ∣ and h = ∣ X ∣ 1 / 2 h=|X|^{1/2} h = ∣ X ∣ 1/2 . Then φ ( X ) = τ ( h ρ U h ) = ⟨ U h , ρ h ⟩ L 2 \varphi(X)=\tau(h\rho Uh)=\langle Uh,\rho h\rangle_{L^2} φ ( X ) = τ ( h ρ U h ) = ⟨ U h , ρ h ⟩ L 2 , so by Lemma 3.1 for L ρ L_\rho L ρ , whose spectrum lies in σ ( ρ ) \sigma(\rho) σ ( ρ ) , ∣ φ ( X ) ∣ ≤ ∥ h ∥ ∥ L ρ h ∥ ≤ cosh O L C ( ρ ) ⟨ L ρ h , h ⟩ = cosh O L C ( ρ ) φ ( ∣ X ∣ ) |\varphi(X)|\le\|h\|\,\|L_\rho h\|\le\cosh\mathop{\mathrm{OLC}}(\rho)\langle L_\rho h,h\rangle=\cosh\mathop{\mathrm{OLC}}(\rho)\varphi(|X|) ∣ φ ( X ) ∣ ≤ ∥ h ∥ ∥ L ρ h ∥ ≤ cosh OLC ( ρ ) ⟨ L ρ h , h ⟩ = cosh OLC ( ρ ) φ ( ∣ X ∣ ) . (R). σ z ( X ) = z X \sigma_z(X)=zX σ z ( X ) = z X : τ ( z ρ X ) = τ ( ρ z X ) \tau(z\rho X)=\tau(\rho zX) τ ( z ρX ) = τ ( ρ z X ) and ∣ z X ∣ = z ∣ X ∣ |zX|=z|X| ∣ z X ∣ = z ∣ X ∣ . Example. In M 2 M_2 M 2 with ρ = d i a g ( κ , 1 ) \rho=\mathop{\mathrm{diag}}(\kappa,1) ρ = diag ( κ , 1 ) let b = ( κ + 1 ) − 1 / 2 ( 1 , κ ) b=(\kappa+1)^{-1/2}(1,\sqrt\kappa) b = ( κ + 1 ) − 1/2 ( 1 , κ ) , a = ρ b / ∥ ρ b ∥ a=\rho b/\|\rho b\| a = ρ b /∥ ρ b ∥ and X = ∣ a ⟩ ⟨ b ∣ X=|a\rangle\langle b| X = ∣ a ⟩ ⟨ b ∣ . Then ∣ X ∣ = ∣ b ⟩ ⟨ b ∣ |X|=|b\rangle\langle b| ∣ X ∣ = ∣ b ⟩ ⟨ b ∣ , φ ( ∣ X ∣ ) = ⟨ ρ b , b ⟩ = 2 κ / ( κ + 1 ) \varphi(|X|)=\langle\rho b,b\rangle=2\kappa/(\kappa+1) φ ( ∣ X ∣ ) = ⟨ ρ b , b ⟩ = 2 κ / ( κ + 1 ) and φ ( X ) = ⟨ a , ρ b ⟩ = ∥ ρ b ∥ = κ \varphi(X)=\langle a,\rho b\rangle=\|\rho b\|=\sqrt\kappa φ ( X ) = ⟨ a , ρ b ⟩ = ∥ ρ b ∥ = κ , so the ratio is ( 1 + κ ) / ( 2 κ ) (1+\kappa)/(2\sqrt\kappa) ( 1 + κ ) / ( 2 κ ) . Theorem 4.15 and Remark 4.16 complete the proof. ◻
By Theorem 5.1 , the constant on M n M_n M n depends on ρ \rho ρ only through its extreme eigenvalues, not on the dimension or on the rest of the spectrum. Item 3 is a consequence.
Corollary 5.3 (Item 3). For every A ∈ M A\in\mathcal{M} A ∈ M , φ ( A + A ∗ ) ≤ 2 G Z φ ( ∣ A ∗ ∣ ) \varphi(A+A^*)\le2G_\mathcal{Z}\,\varphi(|A^*|) φ ( A + A ∗ ) ≤ 2 G Z φ ( ∣ A ∗ ∣ ) , and the constant is optimal. In M n ( C ) M_n(\mathbb{C}) M n ( C ) it is G G G and attained, so V n = G − 1 V_n=G-1 V n = G − 1 for item 3.
Proof. Since ρ ≥ 0 \rho\ge0 ρ ≥ 0 , φ ( A ∗ ) = τ ( A ρ ) ‾ = φ ( A ) ‾ \varphi(A^*)=\overline{\tau(A\rho)}=\overline{\varphi(A)} φ ( A ∗ ) = τ ( A ρ ) = φ ( A ) , so φ ( A + A ∗ ) = 2 R e φ ( A ) ≤ 2 ∣ φ ( A ∗ ) ∣ ≤ 2 G Z φ ( ∣ A ∗ ∣ ) \varphi(A+A^*)=2\mathop{\mathrm{Re}}\varphi(A)\le2|\varphi(A^*)|\le2G_\mathcal{Z}\varphi(|A^*|) φ ( A + A ∗ ) = 2 Re φ ( A ) ≤ 2∣ φ ( A ∗ ) ∣ ≤ 2 G Z φ ( ∣ A ∗ ∣ ) by Theorem 5.1 for X = A ∗ X=A^* X = A ∗ . For the lower bound let X X X be the example in the proof of Theorem 5.1 , for which φ ( X ) = κ > 0 \varphi(X)=\sqrt\kappa>0 φ ( X ) = κ > 0 , and A = X ∗ A=X^* A = X ∗ . Then φ ( A + A ∗ ) = 2 φ ( X ) \varphi(A+A^*)=2\varphi(X) φ ( A + A ∗ ) = 2 φ ( X ) and φ ( ∣ A ∗ ∣ ) = φ ( ∣ X ∣ ) \varphi(|A^*|)=\varphi(|X|) φ ( ∣ A ∗ ∣ ) = φ ( ∣ X ∣ ) , so the ratio is ( 1 + κ ) / ( 2 κ ) (1+\kappa)/(2\sqrt\kappa) ( 1 + κ ) / ( 2 κ ) . The constant is a violation constant (one arbitrary variable, F 1 = X 1 + X 1 ∗ F_1=X_1+X_1^* F 1 = X 1 + X 1 ∗ , F 2 = ∣ X 1 ∗ ∣ F_2=|X_1^*| F 2 = ∣ X 1 ∗ ∣ ), so Theorem 4.15 t:low and Remark 4.16 complete the proof. ◻
Subadditivity of the modulus
Consider
c g e n = sup A , B ∈ M φ ( ∣ A + B ∣ ) φ ( ∣ A ∣ ) + φ ( ∣ B ∣ ) , c s a = sup A , B ∈ M s a φ ( ∣ A + B ∣ ) φ ( ∣ A ∣ ) + φ ( ∣ B ∣ ) , c p d = sup A , B ∈ M + φ ( ∣ A − B ∣ ) φ ( A + B ) , c p r o j = sup P , Q φ ( ∣ P − Q ∣ ) φ ( P + Q ) , \begin{gather*}
c_{\rm gen}=\sup_{A,B\in\mathcal{M}}\frac{\varphi(|A+B|)}{\varphi(|A|)+\varphi(|B|)},\qquad
c_{\rm sa}=\sup_{A,B\in\mathcal{M}_{sa}}\frac{\varphi(|A+B|)}{\varphi(|A|)+\varphi(|B|)},\\
c_{\rm pd}=\sup_{A,B\in\mathcal{M}_+}\frac{\varphi(|A-B|)}{\varphi(A+B)},\qquad
c_{\rm proj}=\sup_{P,Q}\frac{\varphi(|P-Q|)}{\varphi(P+Q)},
\end{gather*} c gen = A , B ∈ M sup φ ( ∣ A ∣ ) + φ ( ∣ B ∣ ) φ ( ∣ A + B ∣ ) , c sa = A , B ∈ M s a sup φ ( ∣ A ∣ ) + φ ( ∣ B ∣ ) φ ( ∣ A + B ∣ ) , c pd = A , B ∈ M + sup φ ( A + B ) φ ( ∣ A − B ∣ ) , c proj = P , Q sup φ ( P + Q ) φ ( ∣ P − Q ∣ ) ,
the last over projections. Clearly c p r o j ≤ c p d ≤ c s a ≤ c g e n c_{\rm proj}\le c_{\rm pd}\le c_{\rm sa}\le c_{\rm gen} c proj ≤ c pd ≤ c sa ≤ c gen (for c p d ≤ c s a c_{\rm pd}\le c_{\rm sa} c pd ≤ c sa take a 1 = A a_1=A a 1 = A , a 2 = − B a_2=-B a 2 = − B ). The self-adjoint version is the form in which [160 ] characterises traces; c p r o j c_{\rm proj} c proj corresponds to Bikchentaev’s inequality φ ( ∣ P − Q ∣ ) ≤ φ ( P + Q ) \varphi(|P-Q|)\le\varphi(P+Q) φ ( ∣ P − Q ∣ ) ≤ φ ( P + Q ) from the same series [95 , 96 , 97 , 100 ] .
Theorem 5.4 (Items 2 and 18). c g e n = c s a = c p d = c p r o j = G Z c_{\rm gen}=c_{\rm sa}=c_{\rm pd}=c_{\rm proj}=G_\mathcal{Z} c gen = c sa = c pd = c proj = G Z . In M n ( C ) M_n(\mathbb{C}) M n ( C ) all four suprema are attained, by a pair of rank-one projections.
Proof. Let A + B = U ∣ A + B ∣ A+B=U|A+B| A + B = U ∣ A + B ∣ . Then ∣ A + B ∣ = U ∗ A + U ∗ B |A+B|=U^*A+U^*B ∣ A + B ∣ = U ∗ A + U ∗ B , so by Theorem 5.1 φ ( ∣ A + B ∣ ) ≤ ∣ φ ( U ∗ A ) ∣ + ∣ φ ( U ∗ B ) ∣ ≤ G Z ( φ ( ∣ U ∗ A ∣ ) + φ ( ∣ U ∗ B ∣ ) ) \varphi(|A+B|)\le|\varphi(U^*A)|+|\varphi(U^*B)|\le G_\mathcal{Z}(\varphi(|U^*A|)+\varphi(|U^*B|)) φ ( ∣ A + B ∣ ) ≤ ∣ φ ( U ∗ A ) ∣ + ∣ φ ( U ∗ B ) ∣ ≤ G Z ( φ ( ∣ U ∗ A ∣ ) + φ ( ∣ U ∗ B ∣ )) , and ∣ U ∗ A ∣ = ∣ A ∣ |U^*A|=|A| ∣ U ∗ A ∣ = ∣ A ∣ because ∣ U ∗ A ∣ 2 = A ∗ U U ∗ A = ∣ A ∣ 2 |U^*A|^2=A^*UU^*A=|A|^2 ∣ U ∗ A ∣ 2 = A ∗ U U ∗ A = ∣ A ∣ 2 (U U U is unitary). Hence c g e n ≤ G Z c_{\rm gen}\le G_\mathcal{Z} c gen ≤ G Z . For the lower bound take, in M 2 M_2 M 2 with ρ = d i a g ( κ , 1 ) \rho=\mathop{\mathrm{diag}}(\kappa,1) ρ = diag ( κ , 1 ) , s = ( κ + 1 ) − 1 / 2 s=(\kappa+1)^{-1/2} s = ( κ + 1 ) − 1/2 , c = ( κ / ( κ + 1 ) ) 1 / 2 c=(\kappa/(\kappa+1))^{1/2} c = ( κ / ( κ + 1 ) ) 1/2 , b ± = ( s , ± c ) b_\pm=(s,\pm c) b ± = ( s , ± c ) and P ± = ∣ b ± ⟩ ⟨ b ± ∣ P_\pm=|b_\pm\rangle\langle b_\pm| P ± = ∣ b ± ⟩ ⟨ b ± ∣ . Then P + − P − = 2 s c σ P_+-P_-=2sc\,\sigma P + − P − = 2 sc σ with σ = ( 0 1 1 0 ) \sigma=\bigl(\begin{smallmatrix}0&1\\1&0\end{smallmatrix}\bigr) σ = ( 0 1 1 0 ) , ∣ P + − P − ∣ = 2 s c ⋅ 1 |P_+-P_-|=2sc\cdot1 ∣ P + − P − ∣ = 2 sc ⋅ 1 , φ ( ∣ P + − P − ∣ ) = 2 s c ( κ + 1 ) = 2 κ \varphi(|P_+-P_-|)=2sc(\kappa+1)=2\sqrt\kappa φ ( ∣ P + − P − ∣ ) = 2 sc ( κ + 1 ) = 2 κ and φ ( P ± ) = 2 κ / ( κ + 1 ) \varphi(P_\pm)=2\kappa/(\kappa+1) φ ( P ± ) = 2 κ / ( κ + 1 ) , so the ratio is ( 1 + κ ) / ( 2 κ ) (1+\kappa)/(2\sqrt\kappa) ( 1 + κ ) / ( 2 κ ) . Theorem 4.15 t:low applied to c p r o j c_{\rm
proj} c proj gives c p r o j ≥ G Z c_{\rm proj}\ge G_\mathcal{Z} c proj ≥ G Z . ◻
Corollary 5.5. φ ( ∣ A + B ∣ ) ≤ ( 1 − N C ( ρ ) 2 ) − 1 / 2 ( φ ( ∣ A ∣ ) + φ ( ∣ B ∣ ) ) \varphi(|A+B|)\le(1-\mathop{\mathrm{NC}}(\rho)^2)^{-1/2}(\varphi(|A|)+\varphi(|B|)) φ ( ∣ A + B ∣ ) ≤ ( 1 − NC ( ρ ) 2 ) − 1/2 ( φ ( ∣ A ∣ ) + φ ( ∣ B ∣ )) for all A , B ∈ M A,B\in\mathcal{M} A , B ∈ M , and the constant is optimal already for pairs ( P , − Q ) (P,-Q) ( P , − Q ) of projections.
Corollary 5.6 (Items 5 and 18). For projections P , Q , R ∈ M P,Q,R\in\mathcal{M} P , Q , R ∈ M , φ ( ∣ P − R ∣ ) ≤ G Z ( φ ( ∣ P − Q ∣ ) + φ ( ∣ Q − R ∣ ) ) , φ ( ∣ P − Q ∣ ) ≤ G Z φ ( P + Q ) , \varphi(|P-R|)\le G_\mathcal{Z}\bigl(\varphi(|P-Q|)+\varphi(|Q-R|)\bigr),\qquad \varphi(|P-Q|)\le G_\mathcal{Z}\,\varphi(P+Q), φ ( ∣ P − R ∣ ) ≤ G Z ( φ ( ∣ P − Q ∣ ) + φ ( ∣ Q − R ∣ ) ) , φ ( ∣ P − Q ∣ ) ≤ G Z φ ( P + Q ) , and both constants are optimal; in M n ( C ) M_n(\mathbb{C}) M n ( C ) they are attained (by rank-one P , Q P,Q P , Q , and for the first inequality with Q = 0 Q=0 Q = 0 and rank-one P , R P,R P , R ). Hence V n = G − 1 V_n=G-1 V n = G − 1 for items 5 and 18.
Proof. The first inequality is c g e n ≤ G Z c_{\rm gen}\le G_\mathcal{Z} c gen ≤ G Z for A = P − Q A=P-Q A = P − Q , B = Q − R B=Q-R B = Q − R ; the second is c p r o j = G Z c_{\rm proj}=G_\mathcal{Z} c proj = G Z . For Q = 0 Q=0 Q = 0 the first inequality becomes the second with ( P , R ) (P,R) ( P , R ) in place of ( P , Q ) (P,Q) ( P , Q ) , so its constant is at least c p r o j c_{\rm proj} c proj . Attainment follows from Theorem 5.4 . ◻
The modulus of the adjoint
The equality φ ( ∣ X ∗ ∣ ) = φ ( ∣ X ∣ ) \varphi(|X^*|)=\varphi(|X|) φ ( ∣ X ∗ ∣ ) = φ ( ∣ X ∣ ) for all X X X characterises traces, since ∣ X ∗ ∣ = U ∣ X ∣ U ∗ |X^*|=U|X|U^* ∣ X ∗ ∣ = U ∣ X ∣ U ∗ .
Theorem (Item 26). For every partial isometry U U U and Y ∈ M + Y\in\mathcal{M}_+ Y ∈ M + with U ∗ U ≥ s ( Y ) U^*U\ge s(Y) U ∗ U ≥ s ( Y ) , φ ( U Y U ∗ ) ≤ κ Z φ ( Y ) \varphi(UYU^*)\le\kappa_\mathcal{Z}\varphi(Y) φ ( U Y U ∗ ) ≤ κ Z φ ( Y ) . Consequently, for every q > 0 q>0 q > 0 , sup X ≠ 0 φ ( ∣ X ∗ ∣ q ) φ ( ∣ X ∣ q ) = κ Z , inf X ≠ 0 φ ( ∣ X ∗ ∣ q ) φ ( ∣ X ∣ q ) = κ Z − 1 . \sup_{X\ne0}\frac{\varphi(|X^*|^q)}{\varphi(|X|^q)}=\kappa_\mathcal{Z},\qquad
\inf_{X\ne0}\frac{\varphi(|X^*|^q)}{\varphi(|X|^q)}=\kappa_\mathcal{Z}^{-1}. sup X = 0 φ ( ∣ X ∣ q ) φ ( ∣ X ∗ ∣ q ) = κ Z , inf X = 0 φ ( ∣ X ∣ q ) φ ( ∣ X ∗ ∣ q ) = κ Z − 1 . In M n ( C ) M_n(\mathbb{C}) M n ( C ) both are attained at X = ∣ e max ⟩ ⟨ e min ∣ X=|e_{\max}\rangle\langle e_{\min}| X = ∣ e m a x ⟩ ⟨ e m i n ∣ and its adjoint.
Proof. By (3) and centrality of u u u , φ ( U Y U ∗ ) ≤ τ ( u U Y U ∗ ) = τ ( u Y 1 / 2 U ∗ U Y 1 / 2 ) = τ ( u Y ) ≤ κ Z τ ( l Y ) ≤ κ Z φ ( Y ) \varphi(UYU^*)\le\tau(uUYU^*)=\tau(uY^{1/2}U^*UY^{1/2})=\tau(uY)\le\kappa_\mathcal{Z}\tau(lY)\le\kappa_\mathcal{Z}\varphi(Y) φ ( U Y U ∗ ) ≤ τ ( uU Y U ∗ ) = τ ( u Y 1/2 U ∗ U Y 1/2 ) = τ ( u Y ) ≤ κ Z τ ( l Y ) ≤ κ Z φ ( Y ) . With X = U ∣ X ∣ X=U|X| X = U ∣ X ∣ we have ∣ X ∗ ∣ q = U ∣ X ∣ q U ∗ |X^*|^q=U|X|^qU^* ∣ X ∗ ∣ q = U ∣ X ∣ q U ∗ , which gives ≤ \le ≤ ; the infimum follows by X ↔ X ∗ X\leftrightarrow X^* X ↔ X ∗ . For the lower bound, X = e 12 X=e_{12} X = e 12 in M 2 M_2 M 2 with ρ = d i a g ( κ , 1 ) \rho=\mathop{\mathrm{diag}}(\kappa,1) ρ = diag ( κ , 1 ) has ∣ X ∣ q = e 22 |X|^q=e_{22} ∣ X ∣ q = e 22 , ∣ X ∗ ∣ q = e 11 |X^*|^q=e_{11} ∣ X ∗ ∣ q = e 11 and ratio κ \kappa κ ; apply Theorem 4.15 t:low . ◻
The Petz–Zemánek inequality
For k ≥ 2 k\ge2 k ≥ 2 let c k ( ρ ) = sup X ≠ 0 ∣ φ ( X k ) ∣ / φ ( ∣ X ∣ k ) c_k(\rho)=\sup_{X\ne0}|\varphi(X^k)|/\varphi(|X|^k) c k ( ρ ) = sup X = 0 ∣ φ ( X k ) ∣/ φ ( ∣ X ∣ k ) . For k = 2 k=2 k = 2 the inequality ∣ φ ( X 2 ) ∣ ≤ φ ( ∣ X ∣ 2 ) |\varphi(X^2)|\le\varphi(|X|^2) ∣ φ ( X 2 ) ∣ ≤ φ ( ∣ X ∣ 2 ) is also one of the characterisations in [109 ] .
Theorem (Item 4). c 2 ( ρ ) = G Z c_2(\rho)=G_\mathcal{Z} c 2 ( ρ ) = G Z ; moreover ∣ R e φ ( X 2 ) ∣ ≤ ∣ φ ( X 2 ) ∣ ≤ G Z φ ( ∣ X ∣ 2 ) |\mathop{\mathrm{Re}}\varphi(X^2)|\le|\varphi(X^2)|\le G_\mathcal{Z}\varphi(|X|^2) ∣ Re φ ( X 2 ) ∣ ≤ ∣ φ ( X 2 ) ∣ ≤ G Z φ ( ∣ X ∣ 2 ) , and both constants are optimal. In M n ( C ) M_n(\mathbb{C}) M n ( C ) the supremum is attained.
Proof. Upper bound. On L 2 ( M , τ ) L^2(\mathcal{M},\tau) L 2 ( M , τ ) let T h = ρ 1 / 2 h ρ − 1 / 2 Th=\rho^{1/2}h\rho^{-1/2} T h = ρ 1/2 h ρ − 1/2 , a positive invertible operator with σ ( T ) ⊂ [ κ − 1 / 2 , κ 1 / 2 ] \sigma(T)\subset[\kappa^{-1/2},\kappa^{1/2}] σ ( T ) ⊂ [ κ − 1/2 , κ 1/2 ] , and J h = h ∗ Jh=h^* J h = h ∗ . Then J T J = T − 1 JTJ=T^{-1} J T J = T − 1 , hence J T 1 / 2 J = T − 1 / 2 JT^{1/2}J=T^{-1/2} J T 1/2 J = T − 1/2 . For h = X ρ 1 / 2 h=X\rho^{1/2} h = X ρ 1/2 we have ∥ h ∥ 2 = φ ( ∣ X ∣ 2 ) \|h\|^2=\varphi(|X|^2) ∥ h ∥ 2 = φ ( ∣ X ∣ 2 ) and T h = ρ 1 / 2 X Th=\rho^{1/2}X T h = ρ 1/2 X , so, using τ ( x y ) = ⟨ x , J y ⟩ \tau(xy)=\langle x,Jy\rangle τ ( x y ) = ⟨ x , J y ⟩ ,
φ ( X 2 ) = τ ( ρ 1 / 2 X ⋅ X ρ 1 / 2 ) = ⟨ T h , J h ⟩ = ⟨ T 1 / 2 h , J T − 1 / 2 h ⟩ , \varphi(X^2)=\tau(\rho^{1/2}X\cdot X\rho^{1/2})=\langle Th,Jh\rangle=\langle
T^{1/2}h,JT^{-1/2}h\rangle, φ ( X 2 ) = τ ( ρ 1/2 X ⋅ X ρ 1/2 ) = ⟨ T h , J h ⟩ = ⟨ T 1/2 h , J T − 1/2 h ⟩ , and ∣ φ ( X 2 ) ∣ ≤ ∥ T 1 / 2 h ∥ ∥ T − 1 / 2 h ∥ ≤ 1 2 ⟨ ( T + T − 1 ) h , h ⟩ ≤ cosh O L C ( ρ ) ∥ h ∥ 2 |\varphi(X^2)|\le\|T^{1/2}h\|\,\|T^{-1/2}h\|\le\frac12\langle(T+T^{-1})h,h\rangle\le\cosh\mathop{\mathrm{OLC}}(\rho)\|h\|^2 ∣ φ ( X 2 ) ∣ ≤ ∥ T 1/2 h ∥ ∥ T − 1/2 h ∥ ≤ 2 1 ⟨( T + T − 1 ) h , h ⟩ ≤ cosh OLC ( ρ ) ∥ h ∥ 2 . (R). σ z ( X ) = z 1 / 2 X \sigma_z(X)=z^{1/2}X σ z ( X ) = z 1/2 X . Example. For ρ = d i a g ( κ , 1 ) \rho=\mathop{\mathrm{diag}}(\kappa,1) ρ = diag ( κ , 1 ) and X = ρ 1 / 2 σ X=\rho^{1/2}\sigma X = ρ 1/2 σ with σ \sigma σ the flip, X 2 = κ ⋅ 1 X^2=\sqrt\kappa\cdot1 X 2 = κ ⋅ 1 and ∣ X ∣ 2 = σ ρ σ = d i a g ( 1 , κ ) |X|^2=\sigma\rho\sigma=\mathop{\mathrm{diag}}(1,\kappa) ∣ X ∣ 2 = σ ρ σ = diag ( 1 , κ ) , so the ratio is ( κ + 1 ) κ / ( 2 κ ) (\kappa+1)\sqrt\kappa/(2\kappa) ( κ + 1 ) κ / ( 2 κ ) . Theorem 4.15 and Remark 4.16 apply. Since X 2 X^2 X 2 is self-adjoint in the example, the constant for ∣ R e φ ( X 2 ) ∣ |\mathop{\mathrm{Re}}\varphi(X^2)| ∣ Re φ ( X 2 ) ∣ is the same. ◻
Proposition (Item 27). Let k ≥ 3 k\ge3 k ≥ 3 .
c k ( z ρ ) = c k ( ρ ) c_k(z\rho)=c_k(\rho) c k ( z ρ ) = c k ( ρ ) for z ∈ Z + + z\in\mathcal{Z}_{++} z ∈ Z ++ , and c k ( ρ ) ≤ κ Z 1 − 1 / k c_k(\rho)\le\kappa_\mathcal{Z}^{1-1/k} c k ( ρ ) ≤ κ Z 1 − 1/ k .
If ρ ∉ Z \rho\notin\mathcal{Z} ρ ∈ / Z , then c k ( ρ ) ≥ c k M 2 ( d i a g ( κ Z , 1 ) ) > 1 c_k(\rho)\ge c_k^{M_2}(\mathop{\mathrm{diag}}(\kappa_\mathcal{Z},1))>1 c k ( ρ ) ≥ c k M 2 ( diag ( κ Z , 1 )) > 1 ; moreover c k ( ρ ) ≥ G Z c_k(\rho)\ge G_\mathcal{Z} c k ( ρ ) ≥ G Z for every k k k (Proposition (?) ).
In M n ( C ) M_n(\mathbb{C}) M n ( C ) , c k ( ρ ) ≥ 1 k ∑ i λ i / ( ∏ i λ i ) 1 / k c_k(\rho)\ge\frac1k\sum_i\lambda_i/(\prod_i\lambda_i)^{1/k} c k ( ρ ) ≥ k 1 ∑ i λ i / ( ∏ i λ i ) 1/ k for the eigenvalues λ 1 , … , λ k \lambda_1,\dots,\lambda_k λ 1 , … , λ k of any k ≤ n k\le n k ≤ n orthonormal eigenvectors of ρ \rho ρ . In a II1 _1 1 factor, c k ( ρ ) ≥ max 1 ≤ j < k ( j κ + k − j ) / ( k κ j / k ) c_k(\rho)\ge\max_{1\le j<k}(j\kappa+k-j)/(k\kappa^{j/k}) c k ( ρ ) ≥ max 1 ≤ j < k ( j κ + k − j ) / ( k κ j / k ) ; in particular 1 k κ 1 − 1 / k ≤ c k ( ρ ) ≤ κ 1 − 1 / k \frac1k\kappa^{1-1/k}\le c_k(\rho)\le\kappa^{1-1/k} k 1 κ 1 − 1/ k ≤ c k ( ρ ) ≤ κ 1 − 1/ k there.
Proof. pz:inv σ z ( X ) = z 1 / k X \sigma_z(X)=z^{1/k}X σ z ( X ) = z 1/ k X gives (R). Let ϵ = ρ 1 / k \epsilon=\rho^{1/k} ϵ = ρ 1/ k and Z = X ϵ Z=X\epsilon Z = X ϵ . Moving the last factor ϵ − 1 \epsilon^{-1} ϵ − 1 of ( Z ϵ − 1 ) k (Z\epsilon^{-1})^k ( Z ϵ − 1 ) k to the front, φ ( X k ) = τ ( ϵ k − 1 Z ϵ − 1 Z ⋯ ϵ − 1 Z ) \varphi(X^k)=\tau(\epsilon^{k-1}Z\epsilon^{-1}Z\cdots\epsilon^{-1}Z) φ ( X k ) = τ ( ϵ k − 1 Z ϵ − 1 Z ⋯ ϵ − 1 Z ) with k k k factors Z Z Z and k − 1 k-1 k − 1 factors ϵ − 1 \epsilon^{-1} ϵ − 1 . Hölder’s inequality (Z Z Z in L k L^k L k , the rest in L ∞ L^\infty L ∞ ) gives ∣ φ ( X k ) ∣ ≤ ∥ ρ ∥ ( k − 1 ) / k ∥ ρ − 1 ∥ ( k − 1 ) / k ∥ Z ∥ k k |\varphi(X^k)|\le\|\rho\|^{(k-1)/k}\|\rho^{-1}\|^{(k-1)/k}\|Z\|_k^k ∣ φ ( X k ) ∣ ≤ ∥ ρ ∥ ( k − 1 ) / k ∥ ρ − 1 ∥ ( k − 1 ) / k ∥ Z ∥ k k , and by (2) with r = k / 2 r=k/2 r = k /2 , ∥ Z ∥ k k = τ ( ( ϵ X ∗ X ϵ ) k / 2 ) ≤ τ ( ρ 1 / 2 ∣ X ∣ k ρ 1 / 2 ) = φ ( ∣ X ∣ k ) \|Z\|_k^k=\tau((\epsilon X^*X\epsilon)^{k/2})\le\tau(\rho^{1/2}|X|^k\rho^{1/2})=\varphi(|X|^k) ∥ Z ∥ k k = τ (( ϵ X ∗ X ϵ ) k /2 ) ≤ τ ( ρ 1/2 ∣ X ∣ k ρ 1/2 ) = φ ( ∣ X ∣ k ) . So c k ≤ κ 1 − 1 / k c_k\le\kappa^{1-1/k} c k ≤ κ 1 − 1/ k ; apply Theorem 4.15 t:up . pz:low The first inequality is Theorem 4.15 t:low , and c k M 2 ( d i a g ( κ , 1 ) ) > 1 c_k^{M_2}(\mathop{\mathrm{diag}}(\kappa,1))>1 c k M 2 ( diag ( κ , 1 )) > 1 for κ > 1 \kappa>1 κ > 1 by the theorem of Petz and Zemánek [153 ] applied in M 2 ( C ) M_2(\mathbb{C}) M 2 ( C ) . The bound c k ≥ G Z c_k\ge G_\mathcal{Z} c k ≥ G Z is Proposition (?) below. pz:cyc With W W W the cyclic shift e i ↦ e i + 1 e_i\mapsto e_{i+1} e i ↦ e i + 1 (indices mod k k k ) on the span of the eigenvectors and X = W ρ − 1 / k X=W\rho^{-1/k} X = W ρ − 1/ k there, X k = ( ∏ λ i ) − 1 / k ⋅ 1 X^k=(\prod\lambda_i)^{-1/k}\cdot1 X k = ( ∏ λ i ) − 1/ k ⋅ 1 and ∣ X ∣ k = ρ − 1 |X|^k=\rho^{-1} ∣ X ∣ k = ρ − 1 on the span, which gives the ratio. In a II1 _1 1 factor apply Theorem 4.15 t:low with j j j copies of max σ ( ρ ) \max\sigma(\rho) max σ ( ρ ) and k − j k-j k − j copies of min σ ( ρ ) \min\sigma(\rho) min σ ( ρ ) . ◻
Proposition.
For every k ≥ 2 k\ge2 k ≥ 2 , c k ( ρ ) ≥ G Z c_k(\rho)\ge G_\mathcal{Z} c k ( ρ ) ≥ G Z . In M 2 ( C ) M_2(\mathbb{C}) M 2 ( C ) with ρ = d i a g ( κ , 1 ) \rho=\mathop{\mathrm{diag}}(\kappa,1) ρ = diag ( κ , 1 ) the ratio G Z G_\mathcal{Z} G Z is attained for every k k k .
lim inf k → ∞ c k ( ρ ) ≥ c ∞ ( ρ ) : = sup A ∈ M ∣ φ ( e A ) ∣ / φ ( e R e A ) \liminf_{k\to\infty}c_k(\rho)\ge c_\infty(\rho):=\sup_{A\in\mathcal{M}}|\varphi(e^A)|/\varphi(e^{\mathop{\mathrm{Re}}A}) lim inf k → ∞ c k ( ρ ) ≥ c ∞ ( ρ ) := sup A ∈ M ∣ φ ( e A ) ∣/ φ ( e Re A ) .
Proof. pzG:a Let ρ = d i a g ( κ , 1 ) \rho=\mathop{\mathrm{diag}}(\kappa,1) ρ = diag ( κ , 1 ) with κ > 1 \kappa>1 κ > 1 , θ = π / k \theta=\pi/k θ = π / k , s = κ − 1 / ( 2 k ) s=\kappa^{-1/(2k)} s = κ − 1/ ( 2 k ) and c = 2 cos θ / ( s + s − 1 ) c=2\cos\theta/(s+s^{-1}) c = 2 cos θ / ( s + s − 1 ) , so that 0 ≤ c < 1 0\le c<1 0 ≤ c < 1 . Put
V = ( c − 1 − c 2 1 − c 2 c ) , D = d i a g ( s , s − 1 ) , X = V D . V=\begin{pmatrix}c&-\sqrt{1-c^2}\\ \sqrt{1-c^2}&c\end{pmatrix},\qquad D=\mathop{\mathrm{diag}}(s,s^{-1}),\qquad X=VD. V = ( c 1 − c 2 − 1 − c 2 c ) , D = diag ( s , s − 1 ) , X = V D . Since V V V is unitary, X ∗ X = D 2 X^*X=D^2 X ∗ X = D 2 , so ∣ X ∣ k = D k |X|^k=D^k ∣ X ∣ k = D k and φ ( ∣ X ∣ k ) = κ s k + s − k = 2 κ \varphi(|X|^k)=\kappa s^k+s^{-k}=2\sqrt\kappa φ ( ∣ X ∣ k ) = κ s k + s − k = 2 κ . Moreover det X = 1 \det X=1 det X = 1 and tr X = c ( s + s − 1 ) = 2 cos θ \operatorname{tr}X=c(s+s^{-1})=2\cos\theta tr X = c ( s + s − 1 ) = 2 cos θ , so X X X has the distinct eigenvalues e ± i θ e^{\pm i\theta} e ± i θ . Hence X X X is diagonalisable and X k = − 1 X^k=-1 X k = − 1 , so ∣ φ ( X k ) ∣ = κ + 1 |\varphi(X^k)|=\kappa+1 ∣ φ ( X k ) ∣ = κ + 1 and the ratio is ( κ + 1 ) / ( 2 κ ) (\kappa+1)/(2\sqrt\kappa) ( κ + 1 ) / ( 2 κ ) . For ρ ∉ Z \rho\notin\mathcal{Z} ρ ∈ / Z the claim follows from Theorem 4.15 t:low ; for ρ ∈ Z \rho\in\mathcal{Z} ρ ∈ Z both sides equal 1 1 1 .
pzG:b For X k = 1 + A / k X_k=1+A/k X k = 1 + A / k we have X k k → e A X_k^k\to e^A X k k → e A and ∣ X k ∣ k = exp ( k 2 ln ( 1 + A + A ∗ k + A ∗ A k 2 ) ) → e R e A |X_k|^k=\exp\bigl(\frac k2\ln(1+\frac{A+A^*}k+\frac{A^*A}{k^2})\bigr)\to e^{\mathop{\mathrm{Re}}A} ∣ X k ∣ k = exp ( 2 k ln ( 1 + k A + A ∗ + k 2 A ∗ A ) ) → e Re A in norm, so lim inf k c k ( ρ ) ≥ ∣ φ ( e A ) ∣ / φ ( e R e A ) \liminf_kc_k(\rho)\ge|\varphi(e^A)|/\varphi(e^{\mathop{\mathrm{Re}}A}) lim inf k c k ( ρ ) ≥ ∣ φ ( e A ) ∣/ φ ( e Re A ) for every A A A . ◻
For k = 2 k=2 k = 2 the matrix in pzG:a is, up to a scalar and a diagonal unitary, the example in the proof of Theorem (?) .
For k = 2 k=2 k = 2 the bounds in pz:inv and pz:low meet. For k ≥ 3 k\ge3 k ≥ 3 they do not, and the constant depends on more than the extreme points of the spectrum: Proposition 5.11 proves this for k = 3 k=3 k = 3 and Proposition 5.12 for k ≥ 4 k\ge4 k ≥ 4 .
Proposition 5.11. Let μ = ( 2 / 27 ) 1 / 2 ≈ 0.2722 \mu=(2/27)^{1/2}\approx0.2722 μ = ( 2/27 ) 1/2 ≈ 0.2722 , κ > 1 \kappa>1 κ > 1 and θ = κ 1 / 3 \theta=\kappa^{1/3} θ = κ 1/3 .
In M 2 ( C ) M_2(\mathbb{C}) M 2 ( C ) , c 3 ( d i a g ( κ , 1 ) ) ≤ μ θ 2 + 2 θ + 2 + 2 θ − 1 + θ − 2 c_3(\mathop{\mathrm{diag}}(\kappa,1))\le\mu\theta^2+2\theta+2+2\theta^{-1}+\theta^{-2} c 3 ( diag ( κ , 1 )) ≤ μ θ 2 + 2 θ + 2 + 2 θ − 1 + θ − 2 .
In M 3 ( C ) M_3(\mathbb{C}) M 3 ( C ) , c 3 ( d i a g ( κ , κ , 1 ) ) ≤ μ θ 2 + 8 θ + 9 + 4 θ − 1 + 4 θ − 2 c_3(\mathop{\mathrm{diag}}(\kappa,\kappa,1))\le\mu\theta^2+8\theta+9+4\theta^{-1}+4\theta^{-2} c 3 ( diag ( κ , κ , 1 )) ≤ μ θ 2 + 8 θ + 9 + 4 θ − 1 + 4 θ − 2 .
In M 3 ( C ) M_3(\mathbb{C}) M 3 ( C ) , c 3 ( d i a g ( κ , 1 , 1 ) ) ≥ ( κ + 2 ) / ( 3 κ 1 / 3 ) = 1 3 θ 2 + 2 3 θ c_3(\mathop{\mathrm{diag}}(\kappa,1,1))\ge(\kappa+2)/(3\kappa^{1/3})=\frac13\theta^2+\frac2{3\theta} c 3 ( diag ( κ , 1 , 1 )) ≥ ( κ + 2 ) / ( 3 κ 1/3 ) = 3 1 θ 2 + 3 θ 2 .
The bound in pz3:c exceeds the bound in pz3:a for κ ≥ 4 ⋅ 10 4 \kappa\ge4\cdot10^4 κ ≥ 4 ⋅ 1 0 4 and the bound in pz3:b for κ ≥ 2.5 ⋅ 10 6 \kappa\ge2.5\cdot10^6 κ ≥ 2.5 ⋅ 1 0 6 . Hence c 3 c_3 c 3 is not a function of the condition number (which equals κ Z \kappa_\mathcal{Z} κ Z here) even on the single algebra M 3 ( C ) M_3(\mathbb{C}) M 3 ( C ) : the densities d i a g ( κ , κ , 1 ) \mathop{\mathrm{diag}}(\kappa,\kappa,1) diag ( κ , κ , 1 ) and d i a g ( κ , 1 , 1 ) \mathop{\mathrm{diag}}(\kappa,1,1) diag ( κ , 1 , 1 ) have the same condition number but different constants.
Proof. First, every W ∈ M 2 ( C ) W\in M_2(\mathbb{C}) W ∈ M 2 ( C ) with singular values s 1 ≥ s 2 s_1\ge s_2 s 1 ≥ s 2 satisfies
∣ W 12 W 22 W 21 ∣ ≤ ( 1 3 ( ∣ W 12 ∣ 2 + ∣ W 22 ∣ 2 + ∣ W 21 ∣ 2 ) ) 3 / 2 ≤ ( s 1 2 + s 2 2 3 ) 3 / 2 ≤ μ ( s 1 3 + s 2 3 ) , \begin{equation}
|W_{12}W_{22}W_{21}|\le\Bigl(\tfrac13(|W_{12}|^2+|W_{22}|^2+|W_{21}|^2)\Bigr)^{3/2}
\le\Bigl(\tfrac{s_1^2+s_2^2}3\Bigr)^{3/2}\le\mu\,(s_1^3+s_2^3), \tag*{(4)}
\end{equation} ∣ W 12 W 22 W 21 ∣ ≤ ( 3 1 ( ∣ W 12 ∣ 2 + ∣ W 22 ∣ 2 + ∣ W 21 ∣ 2 ) ) 3/2 ≤ ( 3 s 1 2 + s 2 2 ) 3/2 ≤ μ ( s 1 3 + s 2 3 ) , (4) by the AM–GM inequality, ∑ ∣ W i j ∣ 2 = s 1 2 + s 2 2 \sum|W_{ij}|^2=s_1^2+s_2^2 ∑ ∣ W ij ∣ 2 = s 1 2 + s 2 2 , and the power-mean inequality ( s 1 2 + s 2 2 ) 3 / 2 ≤ 2 ( s 1 3 + s 2 3 ) (s_1^2+s_2^2)^{3/2}\le\sqrt2(s_1^3+s_2^3) ( s 1 2 + s 2 2 ) 3/2 ≤ 2 ( s 1 3 + s 2 3 ) .
Let ρ \rho ρ be diagonal, ϵ = ρ 1 / 3 \epsilon=\rho^{1/3} ϵ = ρ 1/3 and Z = X ϵ Z=X\epsilon Z = X ϵ . As in Proposition (?) , φ ( X 3 ) = ∑ i , j , l ϵ i 2 ϵ j − 1 ϵ l − 1 Z i j Z j l Z l i \varphi(X^3)=\sum_{i,j,l}\epsilon_i^2\epsilon_j^{-1}\epsilon_l^{-1}Z_{ij}Z_{jl}Z_{li} φ ( X 3 ) = ∑ i , j , l ϵ i 2 ϵ j − 1 ϵ l − 1 Z ij Z j l Z l i and φ ( ∣ X ∣ 3 ) ≥ T r ∣ Z ∣ 3 \varphi(|X|^3)\ge\mathop{\mathrm{Tr}}|Z|^3 φ ( ∣ X ∣ 3 ) ≥ Tr ∣ Z ∣ 3 ; moreover ∣ Z i j Z j l Z l i ∣ ≤ ∥ Z ∥ 3 ≤ T r ∣ Z ∣ 3 |Z_{ij}Z_{jl}Z_{li}|\le\|Z\|^3\le\mathop{\mathrm{Tr}}|Z|^3 ∣ Z ij Z j l Z l i ∣ ≤ ∥ Z ∥ 3 ≤ Tr ∣ Z ∣ 3 .
pz3:a Here ϵ = d i a g ( θ , 1 ) \epsilon=\mathop{\mathrm{diag}}(\theta,1) ϵ = diag ( θ , 1 ) . The weight ϵ i 2 / ( ϵ j ϵ l ) \epsilon_i^2/(\epsilon_j\epsilon_l) ϵ i 2 / ( ϵ j ϵ l ) is θ 2 \theta^2 θ 2 for ( i , j , l ) = ( 1 , 2 , 2 ) (i,j,l)=(1,2,2) ( i , j , l ) = ( 1 , 2 , 2 ) , and the other seven weights are 1 , θ , θ , θ − 2 , θ − 1 , θ − 1 , 1 1,\theta,\theta,\theta^{-2},\theta^{-1},\theta^{-1},1 1 , θ , θ , θ − 2 , θ − 1 , θ − 1 , 1 . The term with weight θ 2 \theta^2 θ 2 is bounded by (4) with W = Z W=Z W = Z , the others by T r ∣ Z ∣ 3 \mathop{\mathrm{Tr}}|Z|^3 Tr ∣ Z ∣ 3 .
pz3:b Here ϵ = d i a g ( θ , θ , 1 ) \epsilon=\mathop{\mathrm{diag}}(\theta,\theta,1) ϵ = diag ( θ , θ , 1 ) . The weight θ 2 \theta^2 θ 2 occurs for i ∈ { 1 , 2 } i\in\{1,2\} i ∈ { 1 , 2 } , j = l = 3 j=l=3 j = l = 3 ; the weight θ \theta θ for 8 8 8 triples, 1 1 1 for 9 9 9 , θ − 1 \theta^{-1} θ − 1 for 4 4 4 and θ − 2 \theta^{-2} θ − 2 for 4 4 4 . The θ 2 \theta^2 θ 2 -part is T = Z 33 ( Z 13 Z 31 + Z 23 Z 32 ) T=Z_{33}(Z_{13}Z_{31}+Z_{23}Z_{32}) T = Z 33 ( Z 13 Z 31 + Z 23 Z 32 ) , so ∣ T ∣ ≤ ∣ Z 33 ∣ β γ |T|\le|Z_{33}|\beta\gamma ∣ T ∣ ≤ ∣ Z 33 ∣ β γ with β = ( ∣ Z 13 ∣ 2 + ∣ Z 23 ∣ 2 ) 1 / 2 \beta=(|Z_{13}|^2+|Z_{23}|^2)^{1/2} β = ( ∣ Z 13 ∣ 2 + ∣ Z 23 ∣ 2 ) 1/2 , γ = ( ∣ Z 31 ∣ 2 + ∣ Z 32 ∣ 2 ) 1 / 2 \gamma=(|Z_{31}|^2+|Z_{32}|^2)^{1/2} γ = ( ∣ Z 31 ∣ 2 + ∣ Z 32 ∣ 2 ) 1/2 . If β γ ≠ 0 \beta\gamma\ne0 β γ = 0 , let u , u ′ ∈ span { e 1 , e 2 } u,u'\in\operatorname{span}\{e_1,e_2\} u , u ′ ∈ span { e 1 , e 2 } be the unit vectors with coordinates Z ˉ 3 k / γ \bar Z_{3k}/\gamma Z ˉ 3 k / γ and Z k 3 / β Z_{k3}/\beta Z k 3 / β (k = 1 , 2 k=1,2 k = 1 , 2 ), and let W W W be the matrix of the compression of Z Z Z from span { u , e 3 } \operatorname{span}\{u,e_3\} span { u , e 3 } to span { u ′ , e 3 } \operatorname{span}\{u',e_3\} span { u ′ , e 3 } . Then W 12 = β W_{12}=\beta W 12 = β , W 21 = γ W_{21}=\gamma W 21 = γ , W 22 = Z 33 W_{22}=Z_{33} W 22 = Z 33 , and the singular values of W W W do not exceed those of Z Z Z . By (4) , ∣ T ∣ ≤ μ T r ∣ Z ∣ 3 |T|\le\mu\mathop{\mathrm{Tr}}|Z|^3 ∣ T ∣ ≤ μ Tr ∣ Z ∣ 3 .
pz3:c is Proposition (?) pz:cyc .
The difference between pz3:c and pz3:a is h 1 ( θ ) = ( 1 3 − μ ) θ 2 − 2 θ − 2 − 4 3 θ − 1 θ 2 h_1(\theta)=(\frac13-\mu)\theta^2-2\theta-2-\frac4{3\theta}-\frac1{\theta^2} h 1 ( θ ) = ( 3 1 − μ ) θ 2 − 2 θ − 2 − 3 θ 4 − θ 2 1 , which is increasing for θ > ( 1 3 − μ ) − 1 ≈ 16.35 \theta>(\frac13-\mu)^{-1}\approx16.35 θ > ( 3 1 − μ ) − 1 ≈ 16.35 and positive at θ = ( 4 ⋅ 10 4 ) 1 / 3 ≈ 34.20 \theta=(4\cdot10^4)^{1/3}\approx34.20 θ = ( 4 ⋅ 1 0 4 ) 1/3 ≈ 34.20 . The difference between pz3:c and pz3:b is h 2 ( θ ) = ( 1 3 − μ ) θ 2 − 8 θ − 9 − 10 3 θ − 4 θ 2 h_2(\theta)=(\frac13-\mu)\theta^2-8\theta-9-\frac{10}{3\theta}-\frac4{\theta^2} h 2 ( θ ) = ( 3 1 − μ ) θ 2 − 8 θ − 9 − 3 θ 10 − θ 2 4 , which is increasing for θ > 4 ( 1 3 − μ ) − 1 ≈ 65.4 \theta>4(\frac13-\mu)^{-1}\approx65.4 θ > 4 ( 3 1 − μ ) − 1 ≈ 65.4 and positive at θ = ( 2.5 ⋅ 10 6 ) 1 / 3 ≈ 135.7 \theta=(2.5\cdot10^6)^{1/3}\approx135.7 θ = ( 2.5 ⋅ 1 0 6 ) 1/3 ≈ 135.7 . ◻
The thresholds 4 ⋅ 10 4 4\cdot10^4 4 ⋅ 1 0 4 and 2.5 ⋅ 10 6 2.5\cdot10^6 2.5 ⋅ 1 0 6 are conservative: they come from bounding all lower-order terms crudely by T r ∣ Z ∣ 3 \mathop{\mathrm{Tr}}|Z|^3 Tr ∣ Z ∣ 3 . Numerically the constants already differ at κ = 10 \kappa=10 κ = 10 (Remark 5.13 ).
For k ≥ 4 k\ge4 k ≥ 4 the same comparison works, with a different 2 × 2 2\times2 2 × 2 estimate in place of (4) .
Proposition 5.12. Let k ≥ 4 k\ge4 k ≥ 4 , n ≥ k n\ge k n ≥ k , κ > 1 \kappa>1 κ > 1 and θ = κ 1 / k \theta=\kappa^{1/k} θ = κ 1/ k . Put ν 4 = 1 / ( 2 ⋅ 3 3 / 4 ) ≈ 0.2193 \nu_4=1/(2\cdot3^{3/4})\approx0.2193 ν 4 = 1/ ( 2 ⋅ 3 3/4 ) ≈ 0.2193 and ν k = 2 k ( 1 − 2 k ) ( k − 2 ) / 2 \nu_k=\frac2k(1-\frac2k)^{(k-2)/2} ν k = k 2 ( 1 − k 2 ) ( k − 2 ) /2 for k ≥ 5 k\ge5 k ≥ 5 . Then ν k < 1 k \nu_k<\frac1k ν k < k 1 , and in M n ( C ) M_n(\mathbb{C}) M n ( C ) :
c k ( d i a g ( κ , … , κ , 1 ) ) ≤ ν k θ k − 1 + n k θ k − 2 c_k(\mathop{\mathrm{diag}}(\kappa,\dots,\kappa,1))\le\nu_k\theta^{k-1}+n^k\theta^{k-2} c k ( diag ( κ , … , κ , 1 )) ≤ ν k θ k − 1 + n k θ k − 2 ;
c k ( d i a g ( κ , 1 , … , 1 ) ) ≥ ( κ + k − 1 ) / ( k κ 1 / k ) > 1 k θ k − 1 c_k(\mathop{\mathrm{diag}}(\kappa,1,\dots,1))\ge(\kappa+k-1)/(k\kappa^{1/k})>\frac1k\theta^{k-1} c k ( diag ( κ , 1 , … , 1 )) ≥ ( κ + k − 1 ) / ( k κ 1/ k ) > k 1 θ k − 1 .
Hence c k ( d i a g ( κ , 1 , … , 1 ) ) > c k ( d i a g ( κ , … , κ , 1 ) ) c_k(\mathop{\mathrm{diag}}(\kappa,1,\dots,1))>c_k(\mathop{\mathrm{diag}}(\kappa,\dots,\kappa,1)) c k ( diag ( κ , 1 , … , 1 )) > c k ( diag ( κ , … , κ , 1 )) for κ ≥ ( n k / ( 1 k − ν k ) ) k \kappa\ge\bigl(n^k/(\frac1k-\nu_k)\bigr)^k κ ≥ ( n k / ( k 1 − ν k ) ) k . Both densities have condition number κ \kappa κ (= κ Z =\kappa_\mathcal{Z} = κ Z here), so c k c_k c k is not a function of the condition number on M n ( C ) M_n(\mathbb{C}) M n ( C ) for any n ≥ k n\ge k n ≥ k .
Proof. The constants. 4 ν 4 = 2 ⋅ 3 − 3 / 4 < 1 4\nu_4=2\cdot3^{-3/4}<1 4 ν 4 = 2 ⋅ 3 − 3/4 < 1 . For k ≥ 5 k\ge5 k ≥ 5 , k ν k = 2 e g ( k / 2 ) k\nu_k=2e^{g(k/2)} k ν k = 2 e g ( k /2 ) with g ( m ) = ( m − 1 ) ln ( 1 − 1 m ) g(m)=(m-1)\ln(1-\frac1m) g ( m ) = ( m − 1 ) ln ( 1 − m 1 ) ; since g ′ ( m ) = ln ( 1 − 1 m ) + 1 m < 0 g'(m)=\ln(1-\frac1m)+\frac1m<0 g ′ ( m ) = ln ( 1 − m 1 ) + m 1 < 0 and g ( 2 ) = − ln 2 g(2)=-\ln2 g ( 2 ) = − ln 2 , we get k ν k < 1 k\nu_k<1 k ν k < 1 .
A 2 × 2 2\times2 2 × 2 estimate. We show that every W ∈ M 2 ( C ) W\in M_2(\mathbb{C}) W ∈ M 2 ( C ) satisfies
∣ W 22 ∣ k − 2 ∣ W 12 ∣ ∣ W 21 ∣ ≤ ν k T r ∣ W ∣ k . \begin{equation}
|W_{22}|^{k-2}|W_{12}|\,|W_{21}|\le\nu_k\mathop{\mathrm{Tr}}|W|^k. \tag*{(5)}
\end{equation} ∣ W 22 ∣ k − 2 ∣ W 12 ∣ ∣ W 21 ∣ ≤ ν k Tr ∣ W ∣ k . (5) Let x , a , b , c x,a,b,c x , a , b , c be the moduli of W 11 , W 22 , W 12 , W 21 W_{11},W_{22},W_{12},W_{21} W 11 , W 22 , W 12 , W 21 . Let k ≥ 5 k\ge5 k ≥ 5 . Replacing W W W by its transpose, which has the same singular values, we may assume c ≤ b c\le b c ≤ b . Then ∥ W ∥ 2 ≥ ∥ W e 2 ∥ 2 = a 2 + b 2 \|W\|^2\ge\|We_2\|^2=a^2+b^2 ∥ W ∥ 2 ≥ ∥ W e 2 ∥ 2 = a 2 + b 2 , so T r ∣ W ∣ k ≥ ( a 2 + b 2 ) k / 2 \mathop{\mathrm{Tr}}|W|^k\ge(a^2+b^2)^{k/2} Tr ∣ W ∣ k ≥ ( a 2 + b 2 ) k /2 , and a k − 2 b c ≤ a k − 2 b 2 = ( a 2 + b 2 ) k / 2 ( 1 − s ) ( k − 2 ) / 2 s a^{k-2}bc\le a^{k-2}b^2=(a^2+b^2)^{k/2}(1-s)^{(k-2)/2}s a k − 2 b c ≤ a k − 2 b 2 = ( a 2 + b 2 ) k /2 ( 1 − s ) ( k − 2 ) /2 s with s = b 2 / ( a 2 + b 2 ) s=b^2/(a^2+b^2) s = b 2 / ( a 2 + b 2 ) ; the maximum over s ∈ [ 0 , 1 ] s\in[0,1] s ∈ [ 0 , 1 ] is attained at s = 2 / k s=2/k s = 2/ k and equals ν k \nu_k ν k . For k = 4 k=4 k = 4 this argument gives (5) only with the constant 1 / 4 = 1 / k 1/4=1/k 1/4 = 1/ k in place of ν 4 \nu_4 ν 4 , whereas the final comparison of pz4:a with pz4:b needs a constant < 1 / k <1/k < 1/ k ; so we argue differently. Let k = 4 k=4 k = 4 . With singular values s 1 , s 2 s_1,s_2 s 1 , s 2 , T r ∣ W ∣ 4 = ( s 1 2 + s 2 2 ) 2 − 2 s 1 2 s 2 2 = ( ∑ ∣ W i j ∣ 2 ) 2 − 2 ∣ det W ∣ 2 \mathop{\mathrm{Tr}}|W|^4=(s_1^2+s_2^2)^2-2s_1^2s_2^2=(\sum|W_{ij}|^2)^2-2|\det W|^2 Tr ∣ W ∣ 4 = ( s 1 2 + s 2 2 ) 2 − 2 s 1 2 s 2 2 = ( ∑ ∣ W ij ∣ 2 ) 2 − 2∣ det W ∣ 2 , and ∣ det W ∣ ≤ x a + b c ≤ P 1 / 2 Q 1 / 2 |\det W|\le xa+bc\le P^{1/2}Q^{1/2} ∣ det W ∣ ≤ x a + b c ≤ P 1/2 Q 1/2 with P = x 2 + b 2 P=x^2+b^2 P = x 2 + b 2 , Q = a 2 + c 2 Q=a^2+c^2 Q = a 2 + c 2 . Hence T r ∣ W ∣ 4 ≥ ( P + Q ) 2 − 2 P Q = P 2 + Q 2 ≥ b 4 + ( a 2 + c 2 ) 2 \mathop{\mathrm{Tr}}|W|^4\ge(P+Q)^2-2PQ=P^2+Q^2\ge b^4+(a^2+c^2)^2 Tr ∣ W ∣ 4 ≥ ( P + Q ) 2 − 2 P Q = P 2 + Q 2 ≥ b 4 + ( a 2 + c 2 ) 2 . For fixed a , c a,c a , c and K = ( a 2 + c 2 ) 2 K=(a^2+c^2)^2 K = ( a 2 + c 2 ) 2 , max b > 0 b / ( b 4 + K ) = 3 3 / 4 4 K − 3 / 4 \max_{b>0}b/(b^4+K)=\frac{3^{3/4}}4K^{-3/4} max b > 0 b / ( b 4 + K ) = 4 3 3/4 K − 3/4 (at b 4 = K / 3 b^4=K/3 b 4 = K /3 ), and a 2 c ( a 2 + c 2 ) − 3 / 2 ≤ 2 ⋅ 3 − 3 / 2 a^2c\,(a^2+c^2)^{-3/2}\le2\cdot3^{-3/2} a 2 c ( a 2 + c 2 ) − 3/2 ≤ 2 ⋅ 3 − 3/2 (at c 2 = a 2 / 2 c^2=a^2/2 c 2 = a 2 /2 ). The product of the two constants is ν 4 \nu_4 ν 4 .
pz4:a Let ρ = d i a g ( κ , … , κ , 1 ) \rho=\mathop{\mathrm{diag}}(\kappa,\dots,\kappa,1) ρ = diag ( κ , … , κ , 1 ) , ϵ = ρ 1 / k = d i a g ( θ , … , θ , 1 ) \epsilon=\rho^{1/k}=\mathop{\mathrm{diag}}(\theta,\dots,\theta,1) ϵ = ρ 1/ k = diag ( θ , … , θ , 1 ) and Z = X ϵ Z=X\epsilon Z = X ϵ . As in the proof of Proposition (?) pz:inv ,
φ ( X k ) = ∑ i 1 , … , i k ϵ i 1 k − 1 ϵ i 2 − 1 ⋯ ϵ i k − 1 Z i 1 i 2 Z i 2 i 3 ⋯ Z i k i 1 , φ ( ∣ X ∣ k ) ≥ T r ∣ Z ∣ k , \varphi(X^k)=\sum_{i_1,\dots,i_k}\epsilon_{i_1}^{k-1}\epsilon_{i_2}^{-1}\cdots\epsilon_{i_k}^{-1}
Z_{i_1i_2}Z_{i_2i_3}\cdots Z_{i_ki_1},\qquad \varphi(|X|^k)\ge\mathop{\mathrm{Tr}}|Z|^k, φ ( X k ) = i 1 , … , i k ∑ ϵ i 1 k − 1 ϵ i 2 − 1 ⋯ ϵ i k − 1 Z i 1 i 2 Z i 2 i 3 ⋯ Z i k i 1 , φ ( ∣ X ∣ k ) ≥ Tr ∣ Z ∣ k , and every product of k k k entries of Z Z Z is bounded by ∥ Z ∥ k ≤ T r ∣ Z ∣ k \|Z\|^k\le\mathop{\mathrm{Tr}}|Z|^k ∥ Z ∥ k ≤ Tr ∣ Z ∣ k . The weight is θ k − 1 \theta^{k-1} θ k − 1 if i 1 < n i_1<n i 1 < n and i 2 = ⋯ = i k = n i_2=\dots=i_k=n i 2 = ⋯ = i k = n , and at most θ k − 2 \theta^{k-2} θ k − 2 for each of the fewer than n k n^k n k other index tuples. The θ k − 1 \theta^{k-1} θ k − 1 -part is T = Z n n k − 2 ∑ i < n Z i n Z n i T=Z_{nn}^{k-2}\sum_{i<n}Z_{in}Z_{ni} T = Z nn k − 2 ∑ i < n Z in Z ni , so ∣ T ∣ ≤ ∣ Z n n ∣ k − 2 β γ |T|\le|Z_{nn}|^{k-2}\beta\gamma ∣ T ∣ ≤ ∣ Z nn ∣ k − 2 β γ with β = ( ∑ i < n ∣ Z i n ∣ 2 ) 1 / 2 \beta=(\sum_{i<n}|Z_{in}|^2)^{1/2} β = ( ∑ i < n ∣ Z in ∣ 2 ) 1/2 and γ = ( ∑ i < n ∣ Z n i ∣ 2 ) 1 / 2 \gamma=(\sum_{i<n}|Z_{ni}|^2)^{1/2} γ = ( ∑ i < n ∣ Z ni ∣ 2 ) 1/2 . If β γ ≠ 0 \beta\gamma\ne0 β γ = 0 , the compression W W W of Z Z Z constructed as in the proof of Proposition 5.11 pz3:b , with span { e 1 , … , e n − 1 } \operatorname{span}\{e_1,\dots,e_{n-1}\} span { e 1 , … , e n − 1 } and e n e_n e n in place of span { e 1 , e 2 } \operatorname{span}\{e_1,e_2\} span { e 1 , e 2 } and e 3 e_3 e 3 , has W 12 = β W_{12}=\beta W 12 = β , W 21 = γ W_{21}=\gamma W 21 = γ , W 22 = Z n n W_{22}=Z_{nn} W 22 = Z nn , and singular values not exceeding those of Z Z Z . Hence T r ∣ W ∣ k ≤ T r ∣ Z ∣ k \mathop{\mathrm{Tr}}|W|^k\le\mathop{\mathrm{Tr}}|Z|^k Tr ∣ W ∣ k ≤ Tr ∣ Z ∣ k , and (5) gives ∣ T ∣ ≤ ν k T r ∣ Z ∣ k |T|\le\nu_k\mathop{\mathrm{Tr}}|Z|^k ∣ T ∣ ≤ ν k Tr ∣ Z ∣ k . Altogether ∣ φ ( X k ) ∣ ≤ ( ν k θ k − 1 + n k θ k − 2 ) φ ( ∣ X ∣ k ) |\varphi(X^k)|\le(\nu_k\theta^{k-1}+n^k\theta^{k-2})\varphi(|X|^k) ∣ φ ( X k ) ∣ ≤ ( ν k θ k − 1 + n k θ k − 2 ) φ ( ∣ X ∣ k ) .
pz4:b is Proposition (?) pz:cyc for κ , 1 , … , 1 \kappa,1,\dots,1 κ , 1 , … , 1 , and θ k + k − 1 > θ k \theta^k+k-1>\theta^k θ k + k − 1 > θ k .
Finally, if θ ≥ n k / ( 1 k − ν k ) \theta\ge n^k/(\frac1k-\nu_k) θ ≥ n k / ( k 1 − ν k ) , then ( 1 k − ν k ) θ k − 1 ≥ n k θ k − 2 (\frac1k-\nu_k)\theta^{k-1}\ge n^k\theta^{k-2} ( k 1 − ν k ) θ k − 1 ≥ n k θ k − 2 , and the claim follows from pz4:a and pz4:b . ◻
As for k = 3 k=3 k = 3 , the threshold is conservative (for k = n = 4 k=n=4 k = n = 4 it is about 5 ⋅ 10 15 5\cdot10^{15} 5 ⋅ 1 0 15 ), while numerically the constants differ already at κ = 10 \kappa=10 κ = 10 .
The Bikchentaev–Abed inequality
Let
γ ( ρ ) = sup A , B ∈ M + ∖ 0 − R e φ ( A B ) φ ( A 2 ) φ ( B 2 ) , K ( ρ ) = { K ∈ R : φ ( A 2 + B 2 + K ( A B + B A ) ) ≥ 0 for all A , B ∈ M + } . \begin{gather*}
\gamma(\rho)=\sup_{A,B\in\mathcal{M}_+\setminus0}\frac{-\mathop{\mathrm{Re}}\varphi(AB)}{\sqrt{\varphi(A^2)\varphi(B^2)}},\\
\mathcal K(\rho)=\{K\in\mathbb{R}:\varphi(A^2+B^2+K(AB+BA))\ge0\ \text{ for all }A,B\in\mathcal{M}_+\}.
\end{gather*} γ ( ρ ) = A , B ∈ M + ∖ 0 sup φ ( A 2 ) φ ( B 2 ) − Re φ ( A B ) , K ( ρ ) = { K ∈ R : φ ( A 2 + B 2 + K ( A B + B A )) ≥ 0 for all A , B ∈ M + } .
Theorem (Item 28). Let M ≠ C 1 \mathcal{M}\ne\mathbb{C}1 M = C 1 . Then γ ( ρ ) = tanh 2 L N C ( ρ ) 2 = ( κ Z − 1 κ Z + 1 ) 2 = G Z − 1 G Z + 1 \gamma(\rho)=\tanh^2\frac{\mathop{\mathrm{LNC}}(\rho)}2=\Bigl(\frac{\sqrt{\kappa_\mathcal{Z}}-1}{\sqrt{\kappa_\mathcal{Z}}+1}\Bigr)^2=\frac{G_\mathcal{Z}-1}{G_\mathcal{Z}+1} γ ( ρ ) = tanh 2 2 LNC ( ρ ) = ( κ Z + 1 κ Z − 1 ) 2 = G Z + 1 G Z − 1 , and K ( ρ ) = [ − 1 , K ∗ ( ρ ) ] \mathcal K(\rho)=[-1,K^*(\rho)] K ( ρ ) = [ − 1 , K ∗ ( ρ )] with K ∗ ( ρ ) = 1 γ ( ρ ) = coth 2 L N C ( ρ ) 2 = G Z + 1 G Z − 1 ( K ∗ = + ∞ iff ρ ∈ Z ) . K^*(\rho)=\frac1{\gamma(\rho)}=\coth^2\frac{\mathop{\mathrm{LNC}}(\rho)}2=\frac{G_\mathcal{Z}+1}{G_\mathcal{Z}-1}\qquad(K^*=+\infty\text{ iff }\rho\in\mathcal{Z}). K ∗ ( ρ ) = γ ( ρ ) 1 = coth 2 2 LNC ( ρ ) = G Z − 1 G Z + 1 ( K ∗ = + ∞ iff ρ ∈ Z ) . Equivalently, R e φ ( A B ) ≥ − tanh 2 ( L N C ( ρ ) / 2 ) φ ( A 2 ) φ ( B 2 ) \mathop{\mathrm{Re}}\varphi(AB)\ge-\tanh^2(\mathop{\mathrm{LNC}}(\rho)/2)\sqrt{\varphi(A^2)\varphi(B^2)} Re φ ( A B ) ≥ − tanh 2 ( LNC ( ρ ) /2 ) φ ( A 2 ) φ ( B 2 ) for all A , B ∈ M + A,B\in\mathcal{M}_+ A , B ∈ M + , with the optimal constant. In M n ( C ) M_n(\mathbb{C}) M n ( C ) the supremum is attained.
Proof. Reduction. φ ( A B + B A ) = 2 R e φ ( A B ) \varphi(AB+BA)=2\mathop{\mathrm{Re}}\varphi(AB) φ ( A B + B A ) = 2 Re φ ( A B ) , since τ ( ρ A B ) ‾ = τ ( B A ρ ) \overline{\tau(\rho AB)}=\tau(BA\rho) τ ( ρ A B ) = τ ( B A ρ ) . Replacing ( A , B ) (A,B) ( A , B ) by ( s A , B / s ) (sA,B/s) ( s A , B / s ) and minimising s 2 φ ( A 2 ) + s − 2 φ ( B 2 ) s^2\varphi(A^2)+s^{-2}\varphi(B^2) s 2 φ ( A 2 ) + s − 2 φ ( B 2 ) shows that K ∈ K ( ρ ) K\in\mathcal K(\rho) K ∈ K ( ρ ) iff − K R e φ ( A B ) ≤ φ ( A 2 ) φ ( B 2 ) -K\mathop{\mathrm{Re}}\varphi(AB)\le\sqrt{\varphi(A^2)\varphi(B^2)} − K Re φ ( A B ) ≤ φ ( A 2 ) φ ( B 2 ) for all A , B ≥ 0 A,B\ge0 A , B ≥ 0 . For K ∈ [ − 1 , 0 ] K\in[-1,0] K ∈ [ − 1 , 0 ] this holds by the Cauchy–Schwarz inequality for ( X , Y ) ↦ φ ( Y ∗ X ) (X,Y)\mapsto\varphi(Y^*X) ( X , Y ) ↦ φ ( Y ∗ X ) ; for K < − 1 K<-1 K < − 1 it fails at A = B A=B A = B ; for K > 0 K>0 K > 0 it holds iff K γ ≤ 1 K\gamma\le1 K γ ≤ 1 . Since γ ≥ 0 \gamma\ge0 γ ≥ 0 (take nonzero A , B ≥ 0 A,B\ge0 A , B ≥ 0 with A B = 0 AB=0 A B = 0 , which exist as M ≠ C 1 \mathcal{M}\ne\mathbb{C}1 M = C 1 ), K = [ − 1 , 1 / γ ] \mathcal K=[-1,1/\gamma] K = [ − 1 , 1/ γ ] .
Upper bound with κ \kappa κ . Let D = ρ 1 / 2 D=\rho^{1/2} D = ρ 1/2 , L = O L C ( ρ ) L=\mathop{\mathrm{OLC}}(\rho) L = OLC ( ρ ) , S = ρ 1 / 4 A ρ 1 / 4 S=\rho^{1/4}A\rho^{1/4} S = ρ 1/4 A ρ 1/4 , T = ρ 1 / 4 B ρ 1 / 4 T=\rho^{1/4}B\rho^{1/4} T = ρ 1/4 B ρ 1/4 ; A ↦ S A\mapsto S A ↦ S is a bijection of M + \mathcal{M}_+ M + . By cyclicity φ ( A B ) = τ ( D S D − 1 T ) \varphi(AB)=\tau(DSD^{-1}T) φ ( A B ) = τ ( D S D − 1 T ) , and τ ( D S D − 1 T ) ‾ = τ ( D − 1 S D T ) \overline{\tau(DSD^{-1}T)}=\tau(D^{-1}SDT) τ ( D S D − 1 T ) = τ ( D − 1 S D T ) , so
R e φ ( A B ) = τ ( Φ ( S ) T ) , φ ( A 2 ) = τ ( Φ ( S ) S ) , φ ( B 2 ) = τ ( Φ ( T ) T ) , \mathop{\mathrm{Re}}\varphi(AB)=\tau(\Phi(S)T),\quad\varphi(A^2)=\tau(\Phi(S)S),\quad\varphi(B^2)=\tau(\Phi(T)T), Re φ ( A B ) = τ ( Φ ( S ) T ) , φ ( A 2 ) = τ ( Φ ( S ) S ) , φ ( B 2 ) = τ ( Φ ( T ) T ) , where Φ = 1 2 ( L D R D − 1 + L D − 1 R D ) \Phi=\frac12(L_DR_{D^{-1}}+L_{D^{-1}}R_D) Φ = 2 1 ( L D R D − 1 + L D − 1 R D ) acts on L 2 ( M , τ ) L^2(\mathcal{M},\tau) L 2 ( M , τ ) . Φ \Phi Φ is a function g ( L D , R D ) g(L_D,R_D) g ( L D , R D ) of two commuting positive operators with spectra in [ ∥ D − 1 ∥ − 1 , ∥ D ∥ ] [\|D^{-1}\|^{-1},\|D\|] [ ∥ D − 1 ∥ − 1 , ∥ D ∥ ] , where g ( s , t ) = 1 2 ( s / t + t / s ) ∈ [ 1 , cosh L ] g(s,t)=\frac12(s/t+t/s)\in[1,\cosh L] g ( s , t ) = 2 1 ( s / t + t / s ) ∈ [ 1 , cosh L ] , so σ ( Φ ) ⊂ [ 1 , cosh L ] \sigma(\Phi)\subset[1,\cosh L] σ ( Φ ) ⊂ [ 1 , cosh L ] on L 2 ( M , τ ) L^2(\mathcal{M},\tau) L 2 ( M , τ ) . Since ⟨ S , T ⟩ = τ ( T S ) = τ ( S 1 / 2 T S 1 / 2 ) ≥ 0 \langle S,T\rangle=\tau(TS)=\tau(S^{1/2}TS^{1/2})\ge0 ⟨ S , T ⟩ = τ ( T S ) = τ ( S 1/2 T S 1/2 ) ≥ 0 , Lemma 3.2 w:one gives − R e φ ( A B ) ≤ cosh L − 1 cosh L + 1 φ ( A 2 ) φ ( B 2 ) -\mathop{\mathrm{Re}}\varphi(AB)\le\frac{\cosh L-1}{\cosh L+1}\sqrt{\varphi(A^2)\varphi(B^2)} − Re φ ( A B ) ≤ c o s h L + 1 c o s h L − 1 φ ( A 2 ) φ ( B 2 ) , and cosh L − 1 cosh L + 1 = tanh 2 L 2 \frac{\cosh L-1}{\cosh L+1}=\tanh^2\frac L2 c o s h L + 1 c o s h L − 1 = tanh 2 2 L .
(R). σ z ( A , B ) = ( z 1 / 2 A , z 1 / 2 B ) \sigma_z(A,B)=(z^{1/2}A,z^{1/2}B) σ z ( A , B ) = ( z 1/2 A , z 1/2 B ) : τ ( ρ z 1 / 2 A z 1 / 2 B ) = τ ( z ρ A B ) \tau(\rho z^{1/2}Az^{1/2}B)=\tau(z\rho AB) τ ( ρ z 1/2 A z 1/2 B ) = τ ( z ρ A B ) , and similarly for A 2 A^2 A 2 , B 2 B^2 B 2 .
Example. In M 2 M_2 M 2 with ρ = d i a g ( κ , 1 ) \rho=\mathop{\mathrm{diag}}(\kappa,1) ρ = diag ( κ , 1 ) let A = ∣ a ⟩ ⟨ a ∣ A=|a\rangle\langle a| A = ∣ a ⟩ ⟨ a ∣ , B = ∣ b ⟩ ⟨ b ∣ B=|b\rangle\langle b| B = ∣ b ⟩ ⟨ b ∣ with a = ρ − 1 / 4 ( 1 , 1 ) a=\rho^{-1/4}(1,1) a = ρ − 1/4 ( 1 , 1 ) , b = ρ − 1 / 4 ( 1 , − 1 ) b=\rho^{-1/4}(1,-1) b = ρ − 1/4 ( 1 , − 1 ) , i.e. S = ( 1 1 1 1 ) S=\bigl(\begin{smallmatrix}1&1\\1&1\end{smallmatrix}\bigr) S = ( 1 1 1 1 ) , T = ( 1 − 1 − 1 1 ) T=\bigl(\begin{smallmatrix}1&-1\\-1&1\end{smallmatrix}\bigr) T = ( 1 − 1 − 1 1 ) . Φ \Phi Φ acts as the Schur multiplier with entries 1 1 1 on the diagonal and C = cosh L C=\cosh L C = cosh L off it, so τ ( Φ ( S ) T ) = 2 − 2 C \tau(\Phi(S)T)=2-2C τ ( Φ ( S ) T ) = 2 − 2 C and τ ( Φ ( S ) S ) = τ ( Φ ( T ) T ) = 2 + 2 C \tau(\Phi(S)S)=\tau(\Phi(T)T)=2+2C τ ( Φ ( S ) S ) = τ ( Φ ( T ) T ) = 2 + 2 C , and the ratio is ( C − 1 ) / ( C + 1 ) (C-1)/(C+1) ( C − 1 ) / ( C + 1 ) . The theorem follows from Theorem 4.15 applied to γ \gamma γ (with f ( κ ) = tanh 2 ( 1 4 ln κ ) f(\kappa)=\tanh^2(\frac14\ln\kappa) f ( κ ) = tanh 2 ( 4 1 ln κ ) , f ( 1 ) = 0 f(1)=0 f ( 1 ) = 0 , attained at A B = 0 AB=0 A B = 0 ). ◻
Corollary 5.15. Let K > 1 K>1 K > 1 . Then φ ( A 2 + B 2 + K ( A B + B A ) ) ≥ 0 \varphi(A^2+B^2+K(AB+BA))\ge0 φ ( A 2 + B 2 + K ( A B + B A )) ≥ 0 for all A , B ∈ M + A,B\in\mathcal{M}_+ A , B ∈ M + if and only if κ Z ( ρ ) ≤ ( K + 1 K − 1 ) 2 . \kappa_\mathcal{Z}(\rho)\le\Bigl(\frac{\sqrt K+1}{\sqrt K-1}\Bigr)^2. κ Z ( ρ ) ≤ ( K − 1 K + 1 ) 2 . In particular, for φ = τ ( ρ ⋅ ) \varphi=\tau(\rho\,\cdot) φ = τ ( ρ ⋅ ) with ρ ∈ M + + \rho\in\mathcal{M}_{++} ρ ∈ M ++ , condition (iii) of [104 Thm. 4.6] (there is K 0 > − 1 K_0>-1 K 0 > − 1 such that the inequality holds for all K > K 0 K>K_0 K > K 0 and all A , B ≥ 0 A,B\ge0 A , B ≥ 0 ) can be weakened to: the inequality holds for arbitrarily large K K K . Every non-tracial φ \varphi φ satisfies the inequality exactly on the bounded interval [ − 1 , K ∗ ( ρ ) ] [-1,K^*(\rho)] [ − 1 , K ∗ ( ρ )] .
Proof. K ≤ 1 / γ K\le1/\gamma K ≤ 1/ γ iff tanh ( L N C / 2 ) = ( κ Z − 1 ) / ( κ Z + 1 ) ≤ K − 1 / 2 \tanh(\mathop{\mathrm{LNC}}/2)=(\sqrt{\kappa_\mathcal{Z}}-1)/(\sqrt{\kappa_\mathcal{Z}}+1)\le K^{-1/2} tanh ( LNC /2 ) = ( κ Z − 1 ) / ( κ Z + 1 ) ≤ K − 1/2 . ◻
For projections, φ ( P 2 ) = φ ( P ) \varphi(P^2)=\varphi(P) φ ( P 2 ) = φ ( P ) , and Theorem (?) becomes a sharp lower bound for φ ( R e P Q ) \varphi(\mathrm{Re}\,PQ) φ ( Re P Q ) ; this is item 21.
Proposition 5.16 (Item 21). For all projections P , Q ∈ M P,Q\in\mathcal{M} P , Q ∈ M , − φ ( R e P Q ) ≤ γ ( ρ ) φ ( P ) φ ( Q ) ≤ γ ( ρ ) 2 ( φ ( P ) + φ ( Q ) ) , γ ( ρ ) = tanh 2 L N C ( ρ ) 2 = G Z − 1 G Z + 1 , -\varphi(\mathrm{Re}\,PQ)\le\gamma(\rho)\sqrt{\varphi(P)\varphi(Q)}\le\frac{\gamma(\rho)}2\bigl(\varphi(P)+\varphi(Q)\bigr),\qquad
\gamma(\rho)=\tanh^2\frac{\mathop{\mathrm{LNC}}(\rho)}2=\frac{G_\mathcal{Z}-1}{G_\mathcal{Z}+1}, − φ ( Re P Q ) ≤ γ ( ρ ) φ ( P ) φ ( Q ) ≤ 2 γ ( ρ ) ( φ ( P ) + φ ( Q ) ) , γ ( ρ ) = tanh 2 2 LNC ( ρ ) = G Z + 1 G Z − 1 , where γ ( ρ ) \gamma(\rho) γ ( ρ ) is the γ \gamma γ of Theorem (?) if M ≠ C 1 \mathcal{M}\ne\mathbb{C}1 M = C 1 (on C 1 \mathbb{C}1 C 1 we use the right-hand side), and γ ( ρ ) / 2 \gamma(\rho)/2 γ ( ρ ) /2 is the supremum of − φ ( R e P Q ) / ( φ ( P ) + φ ( Q ) ) -\varphi(\mathrm{Re}\,PQ)/(\varphi(P)+\varphi(Q)) − φ ( Re P Q ) / ( φ ( P ) + φ ( Q )) . In M n ( C ) M_n(\mathbb{C}) M n ( C ) it is attained and equals 1 2 tan 2 ( ψ / 2 ) \frac12\tan^2(\psi/2) 2 1 tan 2 ( ψ /2 ) , so V n = 1 2 tan 2 ( ψ / 2 ) = G − 1 2 ( G + 1 ) V_n=\frac12\tan^2(\psi/2)=\frac{G-1}{2(G+1)} V n = 2 1 tan 2 ( ψ /2 ) = 2 ( G + 1 ) G − 1 for item 21.
Proof. If M = C 1 \mathcal{M}=\mathbb{C}1 M = C 1 , then P , Q ∈ { 0 , 1 } P,Q\in\{0,1\} P , Q ∈ { 0 , 1 } , φ ( R e P Q ) ≥ 0 \varphi(\mathrm{Re}\,PQ)\ge0 φ ( Re P Q ) ≥ 0 and L N C ( ρ ) = 0 \mathop{\mathrm{LNC}}(\rho)=0 LNC ( ρ ) = 0 , so all claims hold trivially (the supremum is 0 0 0 , attained at P = 1 P=1 P = 1 , Q = 0 Q=0 Q = 0 ); let M ≠ C 1 \mathcal{M}\ne\mathbb{C}1 M = C 1 , so that Theorem (?) applies. Since φ ( R e P Q ) = R e φ ( P Q ) \varphi(\mathrm{Re}\,PQ)=\mathop{\mathrm{Re}}\varphi(PQ) φ ( Re P Q ) = Re φ ( P Q ) , P 2 = P P^2=P P 2 = P and Q 2 = Q Q^2=Q Q 2 = Q , the first inequality is Theorem (?) , and the second is the arithmetic–geometric mean inequality. For the lower bound let ρ = d i a g ( κ , 1 ) \rho=\mathop{\mathrm{diag}}(\kappa,1) ρ = diag ( κ , 1 ) in M 2 M_2 M 2 and let P , Q P,Q P , Q be the projections onto the vectors a = ρ − 1 / 4 ( 1 , 1 ) a=\rho^{-1/4}(1,1) a = ρ − 1/4 ( 1 , 1 ) and b = ρ − 1 / 4 ( 1 , − 1 ) b=\rho^{-1/4}(1,-1) b = ρ − 1/4 ( 1 , − 1 ) of the example in the proof of Theorem (?) . The ratio − R e φ ( A B ) / φ ( A 2 ) φ ( B 2 ) -\mathop{\mathrm{Re}}\varphi(AB)/\sqrt{\varphi(A^2)\varphi(B^2)} − Re φ ( A B ) / φ ( A 2 ) φ ( B 2 ) is invariant under positive scalings of A A A and of B B B , so it equals γ \gamma γ also for ( P , Q ) (P,Q) ( P , Q ) ; moreover ∥ a ∥ = ∥ b ∥ \|a\|=\|b\| ∥ a ∥ = ∥ b ∥ and ⟨ ρ a , a ⟩ = ⟨ ρ 1 / 2 ( 1 , 1 ) , ( 1 , 1 ) ⟩ = ⟨ ρ b , b ⟩ \langle\rho a,a\rangle=\langle\rho^{1/2}(1,1),(1,1)\rangle=\langle\rho b,b\rangle ⟨ ρ a , a ⟩ = ⟨ ρ 1/2 ( 1 , 1 ) , ( 1 , 1 )⟩ = ⟨ ρ b , b ⟩ , so φ ( P ) = φ ( Q ) \varphi(P)=\varphi(Q) φ ( P ) = φ ( Q ) and the arithmetic–geometric mean inequality is an equality. Hence the ratio − φ ( R e P Q ) / ( φ ( P ) + φ ( Q ) ) -\varphi(\mathrm{Re}\,PQ)/(\varphi(P)+\varphi(Q)) − φ ( Re P Q ) / ( φ ( P ) + φ ( Q )) equals γ ( d i a g ( κ , 1 ) ) / 2 \gamma(\mathop{\mathrm{diag}}(\kappa,1))/2 γ ( diag ( κ , 1 )) /2 . This ratio is a violation constant (two projections, F 1 = X 1 X 2 + X 2 X 1 F_1=X_1X_2+X_2X_1 F 1 = X 1 X 2 + X 2 X 1 , F 2 = X 1 F_2=X_1 F 2 = X 1 , F 3 = X 2 F_3=X_2 F 3 = X 2 , Ψ N ( z ) = − 1 2 R e z 1 \Psi_N(z)=-\frac12\mathop{\mathrm{Re}}z_1 Ψ N ( z ) = − 2 1 Re z 1 , Ψ D ( z ) = ∣ z 2 ∣ + ∣ z 3 ∣ \Psi_D(z)=|z_2|+|z_3| Ψ D ( z ) = ∣ z 2 ∣ + ∣ z 3 ∣ ), and Theorem 4.15 t:low and Remark 4.16 give the lower bound in M \mathcal{M} M and the attainment in M n M_n M n . Finally tan 2 ψ 2 = 1 − cos ψ 1 + cos ψ = G − 1 G + 1 \tan^2\frac\psi2=\frac{1-\cos\psi}{1+\cos\psi}=\frac{G-1}{G+1} tan 2 2 ψ = 1 + c o s ψ 1 − c o s ψ = G + 1 G − 1 , since cos ψ = 1 / G \cos\psi=1/G cos ψ = 1/ G . ◻
In the example, ∣ ⟨ a , b ⟩ ∣ / ( ∥ a ∥ ∥ b ∥ ) = ( κ − 1 ) / ( κ + 1 ) = tan ( ψ / 2 ) |\langle a,b\rangle|/(\|a\|\|b\|)=(\sqrt\kappa-1)/(\sqrt\kappa+1)=\tan(\psi/2) ∣ ⟨ a , b ⟩ ∣/ ( ∥ a ∥∥ b ∥ ) = ( κ − 1 ) / ( κ + 1 ) = tan ( ψ /2 ) : the two projections make the angle whose cosine is tan ( ψ / 2 ) \tan(\psi/2) tan ( ψ /2 ) , and their Bloch vectors are symmetric about the eigenvector of ρ \rho ρ for its smaller eigenvalue.
The second inequality of [104 ] that we consider, item 9, involves two projections and the compression Q P Q QPQ QP Q ; we determine its envelope on M 2 M_2 M 2 .
Proposition 5.17 (Item 9). For n = 2 n=2 n = 2 , item 9 has the envelope V 2 = s / ( 1 + s ) V_2=s/(1+s) V 2 = s / ( 1 + s ) , s = sin ( ψ / 2 ) s=\sin(\psi/2) s = sin ( ψ /2 ) . Thus V 2 ∼ δ / 2 V_2\sim\delta/2 V 2 ∼ δ /2 as δ → 0 \delta\to0 δ → 0 and V 2 → 2 − 1 V_2\to\sqrt2-1 V 2 → 2 − 1 as δ → 1 \delta\to1 δ → 1 .
Proof. If P P P or Q Q Q is 0 0 0 or 1 1 1 , or P = Q P=Q P = Q , then L = R \mathcal{L}=\mathcal{R} L = R . Otherwise P = u u ∗ P=uu^* P = u u ∗ , Q = v v ∗ Q=vv^* Q = v v ∗ with unit vectors; choose phases so that u = c v + σ v ⊥ u=cv+\sigma v_\perp u = c v + σ v ⊥ with c = cos θ c=\cos\theta c = cos θ , σ = sin θ \sigma=\sin\theta σ = sin θ , 0 < θ < π / 2 0<\theta<\pi/2 0 < θ < π /2 , and a unit vector v ⊥ ⊥ v v_\perp\perp v v ⊥ ⊥ v . Normalise T r ρ = 2 \mathop{\mathrm{Tr}}\rho=2 Tr ρ = 2 , so that ρ \rho ρ has the eigenvalues 1 ± δ 1\pm\delta 1 ± δ , and put β = v ∗ ρ v \beta=v^*\rho v β = v ∗ ρ v , β ′ = v ⊥ ∗ ρ v ⊥ = 2 − β \beta'=v_\perp^*\rho v_\perp=2-\beta β ′ = v ⊥ ∗ ρ v ⊥ = 2 − β , t = R e ( v ∗ ρ v ⊥ ) t=\mathrm{Re}\,(v^*\rho v_\perp) t = Re ( v ∗ ρ v ⊥ ) and μ = ∣ v ∗ ρ v ⊥ ∣ \mu=|v^*\rho v_\perp| μ = ∣ v ∗ ρ v ⊥ ∣ ; then β β ′ − μ 2 = det ρ = 1 − δ 2 \beta\beta'-\mu^2=\det \rho=1-\delta^2 β β ′ − μ 2 = det ρ = 1 − δ 2 . Since P Q + Q P = c ( u v ∗ + v u ∗ ) PQ+QP=c(uv^*+vu^*) P Q + QP = c ( u v ∗ + v u ∗ ) and Q P Q = c 2 Q QPQ=c^2Q QP Q = c 2 Q ,
L − R = 2 c R e ( v ∗ ρ u ) − 2 c 2 β = 2 c σ t , N = u ∗ ρ u + β = β ( 1 + c 2 ) + σ 2 β ′ + 2 c σ t . \mathcal{L}-\mathcal{R}=2c\mathrm{Re}\,(v^*\rho u)-2c^2\beta=2c\sigma t,\qquad \mathcal{N}=u^*\rho u+\beta=\beta(1+c^2)+\sigma^2\beta'+2c\sigma t . L − R = 2 c Re ( v ∗ ρ u ) − 2 c 2 β = 2 c σ t , N = u ∗ ρ u + β = β ( 1 + c 2 ) + σ 2 β ′ + 2 c σ t . With D = β ( 1 + c 2 ) + σ 2 β ′ > 0 D=\beta(1+c^2)+\sigma^2\beta'>0 D = β ( 1 + c 2 ) + σ 2 β ′ > 0 the quotient 2 c σ t / ( D + 2 c σ t ) 2c\sigma t/(D+2c\sigma t) 2 c σ t / ( D + 2 c σ t ) increases with t ≤ μ t\le\mu t ≤ μ , so it is at most r / ( 1 + r ) r/(1+r) r / ( 1 + r ) with
r = 2 c σ μ D = μ sin 2 θ β + 1 + ( β − 1 ) cos 2 θ ≤ μ ( β + 1 ) 2 − ( β − 1 ) 2 = μ 2 β , r=\frac{2c\sigma\mu}{D}=\frac{\mu\sin2\theta}{\beta+1+(\beta-1)\cos2\theta}\le\frac{\mu}{\sqrt{(\beta+1)^2-(\beta-1)^2}}
=\frac\mu{2\sqrt\beta}, r = D 2 c σ μ = β + 1 + ( β − 1 ) cos 2 θ μ sin 2 θ ≤ ( β + 1 ) 2 − ( β − 1 ) 2 μ = 2 β μ , using β ′ = 2 − β \beta'=2-\beta β ′ = 2 − β and max x sin x / ( p + q cos x ) = ( p 2 − q 2 ) − 1 / 2 \max_x\sin x/(p+q\cos x)=(p^2-q^2)^{-1/2} max x sin x / ( p + q cos x ) = ( p 2 − q 2 ) − 1/2 for p > ∣ q ∣ p>|q| p > ∣ q ∣ . Finally r 2 ≤ μ 2 / ( 4 β ) = 1 4 ( 2 − β − ( 1 − δ 2 ) / β ) r^2\le\mu^2/(4\beta)=\frac14\bigl(2-\beta-(1-\delta^2)/\beta\bigr) r 2 ≤ μ 2 / ( 4 β ) = 4 1 ( 2 − β − ( 1 − δ 2 ) / β ) , which is maximal at β = 1 − δ 2 = cos ψ \beta=\sqrt{1-\delta^2}=\cos\psi β = 1 − δ 2 = cos ψ with the value 1 2 ( 1 − cos ψ ) = sin 2 ( ψ / 2 ) \frac12(1-\cos\psi)=\sin^2(\psi/2) 2 1 ( 1 − cos ψ ) = sin 2 ( ψ /2 ) . All steps are equalities for v v v with v ∗ ρ v = cos ψ v^*\rho v=\cos\psi v ∗ ρ v = cos ψ , a phase of v ⊥ v_\perp v ⊥ with v ∗ ρ v ⊥ = μ > 0 v^*\rho v_\perp=\mu>0 v ∗ ρ v ⊥ = μ > 0 , and the optimal θ \theta θ , and r ↦ r / ( 1 + r ) r\mapsto r/(1+r) r ↦ r / ( 1 + r ) is increasing. ◻
Young’s inequality
Let 1 < p < ∞ 1<p<\infty 1 < p < ∞ , 1 / p + 1 / q = 1 1/p+1/q=1 1/ p + 1/ q = 1 , and
c p ( ρ ) = sup A , B ∈ M φ ( ∣ A B ∣ ) φ ( ∣ A ∣ p ) 1 / p φ ( ∣ B ∣ q ) 1 / q , c p + ( ρ ) = sup A , B ∈ M + φ ( ∣ A B ∣ ) φ ( A p ) 1 / p φ ( B q ) 1 / q . c_p(\rho)=\sup_{A,B\in\mathcal{M}}\frac{\varphi(|AB|)}{\varphi(|A|^p)^{1/p}\varphi(|B|^q)^{1/q}},\qquad
c_p^+(\rho)=\sup_{A,B\in\mathcal{M}_+}\frac{\varphi(|AB|)}{\varphi(A^p)^{1/p}\varphi(B^q)^{1/q}}. c p ( ρ ) = A , B ∈ M sup φ ( ∣ A ∣ p ) 1/ p φ ( ∣ B ∣ q ) 1/ q φ ( ∣ A B ∣ ) , c p + ( ρ ) = A , B ∈ M + sup φ ( A p ) 1/ p φ ( B q ) 1/ q φ ( ∣ A B ∣ ) .
Replacing ( A , B ) (A,B) ( A , B ) by ( s A , B / s ) (sA,B/s) ( s A , B / s ) and using inf s ( s p a / p + s − q b / q ) = a 1 / p b 1 / q \inf_s(s^pa/p+s^{-q}b/q)=a^{1/p}b^{1/q} inf s ( s p a / p + s − q b / q ) = a 1/ p b 1/ q , c p c_p c p is also the best constant c c c in φ ( ∣ A B ∣ ) ≤ c φ ( ∣ A ∣ p / p + ∣ B ∣ q / q ) \varphi(|AB|)\le c\,\varphi(|A|^p/p+|B|^q/q) φ ( ∣ A B ∣ ) ≤ c φ ( ∣ A ∣ p / p + ∣ B ∣ q / q ) , and similarly for c p + c_p^+ c p + . The Cho–Sano form φ ( ∣ A ∗ B ∣ ) \varphi(|A^*B|) φ ( ∣ A ∗ B ∣ ) gives the same constant: with A = U ∣ A ∣ A=U|A| A = U ∣ A ∣ , ∣ A ∗ B ∣ = ∣ ∣ A ∣ U ∗ B ∣ = ∣ A ( U ∗ B ) ∣ |A^*B|=||A|U^*B|=|A(U^*B)| ∣ A ∗ B ∣ = ∣∣ A ∣ U ∗ B ∣ = ∣ A ( U ∗ B ) ∣ and ∣ U ∗ B ∣ = ∣ B ∣ |U^*B|=|B| ∣ U ∗ B ∣ = ∣ B ∣ , so B ↦ U ∗ B B\mapsto U^*B B ↦ U ∗ B transforms one supremum into the other.
Lemma 5.18. For c , b ∈ M + c,b\in\mathcal{M}_+ c , b ∈ M + , τ ( ρ ∣ c b ∣ ) ≤ ∥ ρ ∥ 1 / p ∥ c ∥ p φ ( b q ) 1 / q \tau(\rho|cb|)\le\|\rho\|^{1/p}\,\|c\|_p\,\varphi(b^q)^{1/q} τ ( ρ ∣ c b ∣ ) ≤ ∥ ρ ∥ 1/ p ∥ c ∥ p φ ( b q ) 1/ q .
Proof. By norm continuity of b ↦ ∣ c b ∣ b\mapsto|cb| b ↦ ∣ c b ∣ we may assume b b b invertible. Let Y = ∣ c b ∣ Y=|cb| Y = ∣ c b ∣ , c b = W Y cb=WY c b = W Y the polar decomposition, S = b − 1 / 2 Y b − 1 / 2 ≥ 0 S=b^{-1/2}Yb^{-1/2}\ge0 S = b − 1/2 Y b − 1/2 ≥ 0 and Z = b 1 / 2 S b − 1 / 2 = Y b − 1 = W ∗ c Z=b^{1/2}Sb^{-1/2}=Yb^{-1}=W^*c Z = b 1/2 S b − 1/2 = Y b − 1 = W ∗ c , so ∥ Z ∥ p ≤ ∥ c ∥ p \|Z\|_p\le\|c\|_p ∥ Z ∥ p ≤ ∥ c ∥ p . The function F ( z ) = b − z Z b z F(z)=b^{-z}Zb^{z} F ( z ) = b − z Z b z is analytic and bounded on the strip 0 ≤ R e z ≤ 1 0\le\mathop{\mathrm{Re}}z\le1 0 ≤ Re z ≤ 1 , F ( 1 2 ) = S F(\frac12)=S F ( 2 1 ) = S , ∥ F ( i y ) ∥ p = ∥ Z ∥ p \|F(iy)\|_p=\|Z\|_p ∥ F ( i y ) ∥ p = ∥ Z ∥ p , and F ( 1 + i y ) = b − i y Z ∗ b i y F(1+iy)=b^{-iy}Z^*b^{iy} F ( 1 + i y ) = b − i y Z ∗ b i y because b − 1 Z b = b − 1 / 2 S b 1 / 2 = Z ∗ b^{-1}Zb=b^{-1/2}Sb^{1/2}=Z^* b − 1 Z b = b − 1/2 S b 1/2 = Z ∗ . By the three lines lemma applied to z ↦ τ ( F ( z ) V ) z\mapsto\tau(F(z)V) z ↦ τ ( F ( z ) V ) with ∥ V ∥ q ≤ 1 \|V\|_q\le1 ∥ V ∥ q ≤ 1 and by duality, ∥ S ∥ p ≤ ∥ Z ∥ p ≤ ∥ c ∥ p \|S\|_p\le\|Z\|_p\le\|c\|_p ∥ S ∥ p ≤ ∥ Z ∥ p ≤ ∥ c ∥ p . Next, τ ( ρ Y ) = τ ( S b 1 / 2 ρ b 1 / 2 ) ≤ ∥ S ∥ p ∥ b 1 / 2 ρ b 1 / 2 ∥ q \tau(\rho Y)=\tau(S\,b^{1/2}\rho b^{1/2})\le\|S\|_p\|b^{1/2}\rho b^{1/2}\|_q τ ( ρ Y ) = τ ( S b 1/2 ρ b 1/2 ) ≤ ∥ S ∥ p ∥ b 1/2 ρ b 1/2 ∥ q , and by (2) and ρ q ≤ ∥ ρ ∥ q − 1 ρ \rho^q\le\|\rho\|^{q-1}\rho ρ q ≤ ∥ ρ ∥ q − 1 ρ , ∥ b 1 / 2 ρ b 1 / 2 ∥ q q = τ ( ( ρ 1 / 2 b ρ 1 / 2 ) q ) ≤ τ ( ρ q b q ) ≤ ∥ ρ ∥ q − 1 φ ( b q ) \|b^{1/2}\rho b^{1/2}\|_q^q=\tau((\rho^{1/2}b\rho^{1/2})^q)\le\tau(\rho^qb^q)\le\|\rho\|^{q-1}\varphi(b^q) ∥ b 1/2 ρ b 1/2 ∥ q q = τ (( ρ 1/2 b ρ 1/2 ) q ) ≤ τ ( ρ q b q ) ≤ ∥ ρ ∥ q − 1 φ ( b q ) . Since ( q − 1 ) / q = 1 / p (q-1)/q=1/p ( q − 1 ) / q = 1/ p , the lemma follows. ◻
Theorem 5.19 (Item 29). c p ( ρ ) = κ Z 1 / p = e 2 L N C ( ρ ) / p c_p(\rho)=\kappa_\mathcal{Z}^{1/p}=e^{2\mathop{\mathrm{LNC}}(\rho)/p} c p ( ρ ) = κ Z 1/ p = e 2 LNC ( ρ ) / p ; that is, φ ( ∣ A B ∣ ) ≤ κ Z 1 / p φ ( ∣ A ∣ p ) 1 / p φ ( ∣ B ∣ q ) 1 / q ≤ κ Z 1 / p φ ( ∣ A ∣ p / p + ∣ B ∣ q / q ) \varphi(|AB|)\le\kappa_\mathcal{Z}^{1/p}\varphi(|A|^p)^{1/p}\varphi(|B|^q)^{1/q}\le\kappa_\mathcal{Z}^{1/p}\varphi(|A|^p/p+|B|^q/q) φ ( ∣ A B ∣ ) ≤ κ Z 1/ p φ ( ∣ A ∣ p ) 1/ p φ ( ∣ B ∣ q ) 1/ q ≤ κ Z 1/ p φ ( ∣ A ∣ p / p + ∣ B ∣ q / q ) for all A , B ∈ M A,B\in\mathcal{M} A , B ∈ M , with the optimal constant. In M n ( C ) M_n(\mathbb{C}) M n ( C ) it is attained.
Proof. Since ∣ A B ∣ 2 = B ∗ ∣ A ∣ 2 B |AB|^2=B^*|A|^2B ∣ A B ∣ 2 = B ∗ ∣ A ∣ 2 B , ∣ A B ∣ = ∣ ∣ A ∣ B ∣ |AB|=||A|B| ∣ A B ∣ = ∣∣ A ∣ B ∣ . With B = V ∣ B ∣ B=V|B| B = V ∣ B ∣ , V V V unitary, ∣ ∣ A ∣ V ∣ B ∣ ∣ = ∣ c ∣ B ∣ ∣ ||A|V|B||=|c\,|B|| ∣∣ A ∣ V ∣ B ∣∣ = ∣ c ∣ B ∣∣ for c = V ∗ ∣ A ∣ V c=V^*|A|V c = V ∗ ∣ A ∣ V , and τ ( c p ) = τ ( ∣ A ∣ p ) ≤ ∥ ρ − 1 ∥ φ ( ∣ A ∣ p ) \tau(c^p)=\tau(|A|^p)\le\|\rho^{-1}\|\varphi(|A|^p) τ ( c p ) = τ ( ∣ A ∣ p ) ≤ ∥ ρ − 1 ∥ φ ( ∣ A ∣ p ) . Lemma 5.18 gives c p ≤ κ 1 / p c_p\le\kappa^{1/p} c p ≤ κ 1/ p . (R) holds with σ z ( A , B ) = ( z 1 / p A , z 1 / q B ) \sigma_z(A,B)=(z^{1/p}A,z^{1/q}B) σ z ( A , B ) = ( z 1/ p A , z 1/ q B ) , since ∣ z 1 / p A z 1 / q B ∣ = z ∣ A B ∣ |z^{1/p}Az^{1/q}B|=z|AB| ∣ z 1/ p A z 1/ q B ∣ = z ∣ A B ∣ . In M 2 M_2 M 2 with ρ = d i a g ( κ , 1 ) \rho=\mathop{\mathrm{diag}}(\kappa,1) ρ = diag ( κ , 1 ) , A = e 12 A=e_{12} A = e 12 , B = e 21 B=e_{21} B = e 21 give ∣ A B ∣ = e 11 |AB|=e_{11} ∣ A B ∣ = e 11 , ∣ A ∣ p = e 22 |A|^p=e_{22} ∣ A ∣ p = e 22 , ∣ B ∣ q = e 11 |B|^q=e_{11} ∣ B ∣ q = e 11 and ratio κ / κ 1 / q = κ 1 / p \kappa/\kappa^{1/q}=\kappa^{1/p} κ / κ 1/ q = κ 1/ p . Theorem 4.15 and Remark 4.16 complete the proof. ◻
Corollary 5.20 (Item 8). For all A , B ∈ M A,B\in\mathcal{M} A , B ∈ M , φ ( ∣ A B ∗ ∣ ) ≤ κ Z φ ( ∣ A ∣ 2 ) 1 / 2 φ ( ∣ B ∣ 2 ) 1 / 2 ≤ κ Z ⋅ 1 2 ( φ ( ∣ A ∣ 2 ) + φ ( ∣ B ∣ 2 ) ) , \varphi(|AB^*|)\le\kappa_\mathcal{Z}\,\varphi(|A|^2)^{1/2}\varphi(|B|^2)^{1/2}\le\kappa_\mathcal{Z}\cdot\tfrac12\bigl(\varphi(|A|^2)+\varphi(|B|^2)\bigr), φ ( ∣ A B ∗ ∣ ) ≤ κ Z φ ( ∣ A ∣ 2 ) 1/2 φ ( ∣ B ∣ 2 ) 1/2 ≤ κ Z ⋅ 2 1 ( φ ( ∣ A ∣ 2 ) + φ ( ∣ B ∣ 2 ) ) , and κ Z \kappa_\mathcal{Z} κ Z is optimal in both inequalities. In M n ( C ) M_n(\mathbb{C}) M n ( C ) the constant κ \kappa κ is attained at A = B = e 1 e 2 ∗ A=B=e_1e_2^* A = B = e 1 e 2 ∗ , so V n = κ − 1 = 2 δ / ( 1 − δ ) V_n=\kappa-1=2\delta/(1-\delta) V n = κ − 1 = 2 δ / ( 1 − δ ) for item 8 in every dimension.
Proof. Theorem 5.19 with p = 2 p=2 p = 2 , applied to A A A and B ∗ B^* B ∗ , gives φ ( ∣ A B ∗ ∣ ) ≤ κ Z 1 / 2 φ ( ∣ A ∣ 2 ) 1 / 2 φ ( ∣ B ∗ ∣ 2 ) 1 / 2 \varphi(|AB^*|)\le\kappa_\mathcal{Z}^{1/2}\varphi(|A|^2)^{1/2}\varphi(|B^*|^2)^{1/2} φ ( ∣ A B ∗ ∣ ) ≤ κ Z 1/2 φ ( ∣ A ∣ 2 ) 1/2 φ ( ∣ B ∗ ∣ 2 ) 1/2 , and Theorem (?) with q = 2 q=2 q = 2 gives φ ( ∣ B ∗ ∣ 2 ) ≤ κ Z φ ( ∣ B ∣ 2 ) \varphi(|B^*|^2)\le\kappa_\mathcal{Z}\varphi(|B|^2) φ ( ∣ B ∗ ∣ 2 ) ≤ κ Z φ ( ∣ B ∣ 2 ) ; the second inequality is the arithmetic–geometric mean inequality. For the lower bound let ρ = d i a g ( κ , 1 ) \rho=\mathop{\mathrm{diag}}(\kappa,1) ρ = diag ( κ , 1 ) in M 2 M_2 M 2 and A = B = e 12 A=B=e_{12} A = B = e 12 . Then ∣ A B ∗ ∣ = e 11 |AB^*|=e_{11} ∣ A B ∗ ∣ = e 11 and ∣ A ∣ 2 = ∣ B ∣ 2 = e 22 |A|^2=|B|^2=e_{22} ∣ A ∣ 2 = ∣ B ∣ 2 = e 22 , so both ratios equal κ \kappa κ . The constants are violation constants (two arbitrary variables, F 1 = ∣ X 1 X 2 ∗ ∣ F_1=|X_1X_2^*| F 1 = ∣ X 1 X 2 ∗ ∣ , F 2 = ∣ X 1 ∣ 2 F_2=|X_1|^2 F 2 = ∣ X 1 ∣ 2 , F 3 = ∣ X 2 ∣ 2 F_3=|X_2|^2 F 3 = ∣ X 2 ∣ 2 ), so Theorem 4.15 t:low and Remark 4.16 complete the proof. ◻
With the contrast of the operators prescribed, the violation of item 8 is bounded below as follows.
Proposition 5.22 (Item 8 with two contrasts). Let O M C ( ρ ) = δ \mathop{\mathrm{OMC}}(\rho)=\delta OMC ( ρ ) = δ and let the moduli of A A A and B B B have contrast δ ′ \delta' δ ′ . On M 2 M_2 M 2 , for ρ = d i a g ( a , b ) \rho=\mathrm{diag}(a,b) ρ = diag ( a , b ) , T = d i a g ( ℓ ′ , 1 ) T=\mathrm{diag}(\ell',1) T = diag ( ℓ ′ , 1 ) with ℓ ′ = 1 − δ ′ 1 + δ ′ \ell'=\frac{1-\delta'}{1+\delta'} ℓ ′ = 1 + δ ′ 1 − δ ′ , U U U the swap of the two basis vectors and A = B = U T A=B=UT A = B = U T , the extended Young inequality has the relative violation κ + ℓ ′ 2 1 + κ ℓ ′ 2 − 1 \frac{\kappa+\ell'^2}{1+\kappa\ell'^2}-1 1 + κ ℓ ′2 κ + ℓ ′2 − 1 . Hence, for n = 2 n=2 n = 2 and, by direct sums, for every even n n n , V n ( δ , δ ′ ) ≥ κ + ℓ ′ 2 1 + κ ℓ ′ 2 − 1 = 4 δ δ ′ 1 + δ ′ 2 − 2 δ δ ′ . V_n(\delta,\delta')\ \ge\ \frac{\kappa+\ell'^2}{1+\kappa\ell'^2}-1=\frac{4\delta\delta'}{1+\delta'^2-2\delta\delta'}. V n ( δ , δ ′ ) ≥ 1 + κ ℓ ′2 κ + ℓ ′2 − 1 = 1 + δ ′2 − 2 δ δ ′ 4 δ δ ′ . For δ ′ = 1 \delta'=1 δ ′ = 1 this is the value of Corollary 5.20 .
Proof. For A = B A=B A = B , ∣ A B ∗ ∣ = ∣ A A ∗ ∣ = A A ∗ = U T 2 U ∗ = d i a g ( 1 , ℓ ′ 2 ) |AB^*|=|AA^*|=AA^*=UT^2U^*=\mathrm{diag}(1,\ell'^2) ∣ A B ∗ ∣ = ∣ A A ∗ ∣ = A A ∗ = U T 2 U ∗ = diag ( 1 , ℓ ′2 ) and ∣ A ∣ 2 = T 2 = d i a g ( ℓ ′ 2 , 1 ) |A|^2=T^2=\mathrm{diag}(\ell'^2,1) ∣ A ∣ 2 = T 2 = diag ( ℓ ′2 , 1 ) , so L = a + b ℓ ′ 2 \mathcal{L}=a+b\ell'^2 L = a + b ℓ ′2 and R = a ℓ ′ 2 + b \mathcal{R}=a\ell'^2+b R = a ℓ ′2 + b , and L / R = ( κ + ℓ ′ 2 ) / ( κ ℓ ′ 2 + 1 ) \mathcal{L}/\mathcal{R}=(\kappa+\ell'^2)/(\kappa\ell'^2+1) L / R = ( κ + ℓ ′2 ) / ( κ ℓ ′2 + 1 ) . Inserting κ = 1 + δ 1 − δ \kappa=\frac{1+\delta}{1-\delta} κ = 1 − δ 1 + δ and ℓ ′ \ell' ℓ ′ gives the second form. ◻
The numerical search over all A , B A,B A , B with this contrast finds no larger value on M 2 M_2 M 2 (agreement to four digits, Figure 16 ), and we conjecture equality for even n n n . Unlike Gardner’s inequality, the extended Young inequality fails linearly in each contrast, V ∼ 4 δ δ ′ V\sim4\delta\delta' V ∼ 4 δ δ ′ .
For positive operators the problem is different. For κ ≥ 1 \kappa\ge1 κ ≥ 1 put
y p ( κ ) = max v > 0 ( 1 + κ ) κ 1 / p v ( 1 + v 2 ) 1 p − 1 2 ( v 2 + κ 2 ) − 1 p − 1 2 . \mathsf y_p(\kappa)=\max_{v>0}\ (1+\kappa)\kappa^{1/p}\,v\,(1+v^2)^{\frac1p-\frac12}(v^2+\kappa^2)^{-\frac1p-\frac12}. y p ( κ ) = v > 0 max ( 1 + κ ) κ 1/ p v ( 1 + v 2 ) p 1 − 2 1 ( v 2 + κ 2 ) − p 1 − 2 1 .
Proposition 5.23 (Item 30).
c p + ( z ρ ) = c p + ( ρ ) c_p^+(z\rho)=c_p^+(\rho) c p + ( z ρ ) = c p + ( ρ ) for z ∈ Z + + z\in\mathcal{Z}_{++} z ∈ Z ++ , and y p ( κ Z ) ≤ c p + ( ρ ) ≤ κ Z 1 / p \mathsf y_p(\kappa_\mathcal{Z})\le c_p^+(\rho)\le\kappa_\mathcal{Z}^{1/p} y p ( κ Z ) ≤ c p + ( ρ ) ≤ κ Z 1/ p .
In M 2 ( C ) M_2(\mathbb{C}) M 2 ( C ) with ρ = d i a g ( κ , 1 ) \rho=\mathop{\mathrm{diag}}(\kappa,1) ρ = diag ( κ , 1 ) , y p ( κ ) \mathsf y_p(\kappa) y p ( κ ) is the supremum of the ratio over rank-one A , B A,B A , B . Moreover y p ( κ ) > 1 \mathsf y_p(\kappa)>1 y p ( κ ) > 1 for κ > 1 \kappa>1 κ > 1 , and y 2 ( κ ) = cosh ( 1 2 ln κ ) \mathsf y_2(\kappa)=\cosh(\frac12\ln\kappa) y 2 ( κ ) = cosh ( 2 1 ln κ ) .
If 1 < p ≤ 2 1<p\le2 1 < p ≤ 2 , then in M 2 ( C ) M_2(\mathbb{C}) M 2 ( C ) with ρ = d i a g ( κ , 1 ) \rho=\mathop{\mathrm{diag}}(\kappa,1) ρ = diag ( κ , 1 ) the supremum over A ≥ 0 A\ge0 A ≥ 0 and rank-one B ≥ 0 B\ge0 B ≥ 0 equals y p ( κ ) \mathsf y_p(\kappa) y p ( κ ) .
If p > 2 p>2 p > 2 and ρ ∉ Z \rho\notin\mathcal{Z} ρ ∈ / Z , then c p + ( ρ ) > y p ( κ Z ) c_p^+(\rho)>\mathsf y_p(\kappa_\mathcal{Z}) c p + ( ρ ) > y p ( κ Z ) .
Proof. yp:b For unit vectors x , y x,y x , y and A = ∣ x ⟩ ⟨ x ∣ A=|x\rangle\langle x| A = ∣ x ⟩ ⟨ x ∣ , B = ∣ y ⟩ ⟨ y ∣ B=|y\rangle\langle y| B = ∣ y ⟩ ⟨ y ∣ we have ∣ A B ∣ = ∣ ⟨ x , y ⟩ ∣ ∣ y ⟩ ⟨ y ∣ |AB|=|\langle x,y\rangle|\,|y\rangle\langle y| ∣ A B ∣ = ∣ ⟨ x , y ⟩ ∣ ∣ y ⟩ ⟨ y ∣ , so the ratio is R ( x , y ) = ∣ ⟨ x , y ⟩ ∣ ( ⟨ ρ y , y ⟩ / ⟨ ρ x , x ⟩ ) 1 / p R(x,y)=|\langle x,y\rangle|(\langle\rho y,y\rangle/\langle\rho x,x\rangle)^{1/p} R ( x , y ) = ∣ ⟨ x , y ⟩ ∣ (⟨ ρ y , y ⟩ / ⟨ ρ x , x ⟩ ) 1/ p . Replacing the coordinates by their moduli does not decrease R R R , so let x = ( cos s , sin s ) x=(\cos s,\sin s) x = ( cos s , sin s ) , y = ( cos t , sin t ) y=(\cos t,\sin t) y = ( cos t , sin t ) with s , t ∈ [ 0 , π / 2 ] s,t\in[0,\pi/2] s , t ∈ [ 0 , π /2 ] , and g ( t ) = κ cos 2 t + sin 2 t g(t)=\kappa\cos^2t+\sin^2t g ( t ) = κ cos 2 t + sin 2 t . On the boundary of the square R R R is not maximal unless R ≤ 1 R\le1 R ≤ 1 : e.g. at s = π / 2 s=\pi/2 s = π /2 , ∂ s ln R = − cot t < 0 \partial_s\ln R=-\cot t<0 ∂ s ln R = − cot t < 0 , and similarly on the other sides. At an interior critical point, tan ( t − s ) = 1 p g ′ ( s ) g ( s ) = 1 p g ′ ( t ) g ( t ) \tan(t-s)=\frac1p\frac{g'(s)}{g(s)}=\frac1p\frac{g'(t)}{g(t)} tan ( t − s ) = p 1 g ( s ) g ′ ( s ) = p 1 g ( t ) g ′ ( t ) , and g ′ ( t ) / g ( t ) = − 2 ( κ − 1 ) h ( tan t ) g'(t)/g(t)=-2(\kappa-1)h(\tan t) g ′ ( t ) / g ( t ) = − 2 ( κ − 1 ) h ( tan t ) with h ( v ) = v / ( κ + v 2 ) h(v)=v/(\kappa+v^2) h ( v ) = v / ( κ + v 2 ) . Now h ( v 1 ) = h ( v 2 ) h(v_1)=h(v_2) h ( v 1 ) = h ( v 2 ) iff v 1 = v 2 v_1=v_2 v 1 = v 2 or v 1 v 2 = κ v_1v_2=\kappa v 1 v 2 = κ ; the first case means s = t s=t s = t and R = 1 R=1 R = 1 . Substituting tan s = v \tan s=v tan s = v , tan t = κ / v \tan t=\kappa/v tan t = κ / v gives cos ( t − s ) = ( 1 + κ ) v / ( 1 + v 2 ) ( v 2 + κ 2 ) \cos(t-s)=(1+\kappa)v/\sqrt{(1+v^2)(v^2+\kappa^2)} cos ( t − s ) = ( 1 + κ ) v / ( 1 + v 2 ) ( v 2 + κ 2 ) and g ( t ) / g ( s ) = κ ( 1 + v 2 ) / ( v 2 + κ 2 ) g(t)/g(s)=\kappa(1+v^2)/(v^2+\kappa^2) g ( t ) / g ( s ) = κ ( 1 + v 2 ) / ( v 2 + κ 2 ) , i.e. R R R equals the function maximised in the definition of y p \mathsf y_p y p . That function tends to 0 0 0 as v → 0 v\to0 v → 0 or v → ∞ v\to\infty v → ∞ , equals 1 1 1 at v = κ v=\sqrt\kappa v = κ , and its logarithmic derivative there is 2 p κ − 1 κ ( κ + 1 ) > 0 \frac2p\frac{\kappa-1}{\sqrt\kappa(\kappa+1)}>0 p 2 κ ( κ + 1 ) κ − 1 > 0 ; hence its maximum exceeds 1 1 1 , so the maximum of R R R is interior and equals y p ( κ ) \mathsf y_p(\kappa) y p ( κ ) . For p = 2 p=2 p = 2 the maximum is at v = κ v=\kappa v = κ and equals ( 1 + κ ) / ( 2 κ ) (1+\kappa)/(2\sqrt\kappa) ( 1 + κ ) / ( 2 κ ) .
yp:a (R) holds with σ z ( A , B ) = ( z 1 / p A , z 1 / q B ) \sigma_z(A,B)=(z^{1/p}A,z^{1/q}B) σ z ( A , B ) = ( z 1/ p A , z 1/ q B ) . The upper bound follows from Theorem 5.19 , the lower bound from yp:b and Theorem 4.15 t:low .
yp:c For B = ∣ y ⟩ ⟨ y ∣ B=|y\rangle\langle y| B = ∣ y ⟩ ⟨ y ∣ and A = ∑ i a i ∣ x i ⟩ ⟨ x i ∣ A=\sum_ia_i|x_i\rangle\langle x_i| A = ∑ i a i ∣ x i ⟩ ⟨ x i ∣ (orthonormal x i x_i x i ), with w i = ∣ ⟨ x i , y ⟩ ∣ 2 w_i=|\langle x_i,y\rangle|^2 w i = ∣ ⟨ x i , y ⟩ ∣ 2 and r i = ⟨ ρ x i , x i ⟩ r_i=\langle\rho x_i,x_i\rangle r i = ⟨ ρ x i , x i ⟩ , the p p p -th power of the ratio is ( ∑ a i 2 w i ) p / 2 ⟨ ρ y , y ⟩ / ∑ a i p r i (\sum a_i^2w_i)^{p/2}\langle\rho y,y\rangle/\sum a_i^pr_i ( ∑ a i 2 w i ) p /2 ⟨ ρ y , y ⟩ / ∑ a i p r i . Since p / 2 ≤ 1 p/2\le1 p /2 ≤ 1 , ( ∑ a i 2 w i ) p / 2 ≤ ∑ a i p w i p / 2 ≤ max i ( w i p / 2 / r i ) ∑ a i p r i (\sum a_i^2w_i)^{p/2}\le\sum a_i^pw_i^{p/2}\le\max_i(w_i^{p/2}/r_i)\sum a_i^pr_i ( ∑ a i 2 w i ) p /2 ≤ ∑ a i p w i p /2 ≤ max i ( w i p /2 / r i ) ∑ a i p r i , so the ratio is at most max i R ( x i , y ) ≤ y p ( κ ) \max_iR(x_i,y)\le\mathsf y_p(\kappa) max i R ( x i , y ) ≤ y p ( κ ) .
yp:d Let ( x , y ) (x,y) ( x , y ) be a maximiser of R R R in M 2 M_2 M 2 with κ = κ Z \kappa=\kappa_\mathcal{Z} κ = κ Z , and x ′ ⊥ x x'\perp x x ′ ⊥ x . Then w 1 = ∣ ⟨ x , y ⟩ ∣ 2 ∈ ( 0 , 1 ) w_1=|\langle x,y\rangle|^2\in(0,1) w 1 = ∣ ⟨ x , y ⟩ ∣ 2 ∈ ( 0 , 1 ) , w 2 = 1 − w 1 > 0 w_2=1-w_1>0 w 2 = 1 − w 1 > 0 . For A = ∣ x ⟩ ⟨ x ∣ + a ∣ x ′ ⟩ ⟨ x ′ ∣ A=|x\rangle\langle x|+a|x'\rangle\langle
x'| A = ∣ x ⟩ ⟨ x ∣ + a ∣ x ′ ⟩ ⟨ x ′ ∣ and B = ∣ y ⟩ ⟨ y ∣ B=|y\rangle\langle y| B = ∣ y ⟩ ⟨ y ∣ the p p p -th power of the ratio equals R ( x , y ) p ( 1 + p 2 w 2 w 1 a 2 + o ( a 2 ) ) ( 1 − r 2 r 1 a p + o ( a p ) ) R(x,y)^p(1+\frac p2\frac{w_2}{w_1}a^2+o(a^2))(1-\frac{r_2}{r_1}a^p+o(a^p)) R ( x , y ) p ( 1 + 2 p w 1 w 2 a 2 + o ( a 2 )) ( 1 − r 1 r 2 a p + o ( a p )) , which exceeds R ( x , y ) p R(x,y)^p R ( x , y ) p for small a > 0 a>0 a > 0 because p > 2 p>2 p > 2 . Hence c p + ( d i a g ( κ Z , 1 ) ) > y p ( κ Z ) c_p^+(\mathop{\mathrm{diag}}(\kappa_\mathcal{Z},1))>\mathsf y_p(\kappa_\mathcal{Z}) c p + ( diag ( κ Z , 1 )) > y p ( κ Z ) ; apply Theorem 4.15 t:low . ◻
Conjecture 5.24. For 1 < p ≤ 2 1<p\le2 1 < p ≤ 2 , c p + ( ρ ) = y p ( κ Z ( ρ ) ) c_p^+(\rho)=\mathsf y_p(\kappa_\mathcal{Z}(\rho)) c p + ( ρ ) = y p ( κ Z ( ρ )) . In particular c 2 + ( ρ ) = cosh L N C ( ρ ) c_2^+(\rho)=\cosh\mathop{\mathrm{LNC}}(\rho) c 2 + ( ρ ) = cosh LNC ( ρ ) : for positive operators, the p = 2 p=2 p = 2 Young inequality is violated exactly as much as Gardner’s.
Numerical optimisation over general positive A , B A,B A , B in M 2 M_2 M 2 , M 3 M_3 M 3 and M 4 M_4 M 4 agrees with y p \mathsf y_p y p to about seven digits for p ≤ 2 p\le2 p ≤ 2 , and shows a small excess over y p \mathsf y_p y p for p > 2 p>2 p > 2 , in accordance with yp:d . By Proposition 5.23 , the theorem of Bikchentaev and Tikhonov [110 ] holds in every finite von Neumann algebra for φ = τ ( ρ ⋅ ) \varphi=\tau(\rho\,\cdot) φ = τ ( ρ ⋅ ) with ρ ∈ M + + \rho\in\mathcal{M}_{++} ρ ∈ M ++ ; densities that are unbounded or have unbounded inverses are not covered.
For p = 2 p=2 p = 2 and the normalisation of item 8, an independent BFGS search over A , B ≥ 0 A,B\ge0 A , B ≥ 0 on M 2 M_2 M 2 (30 random starts; a test of the repository) comes within 0.1 % 0.1\% 0.1% of G − 1 G-1 G − 1 and does not exceed it (0.1546 0.1546 0.1546 at δ = 0.5 \delta=0.5 δ = 0.5 and 1.2938 1.2938 1.2938 at δ = 0.9 \delta=0.9 δ = 0.9 , against G − 1 = 0.1547 G-1=0.1547 G − 1 = 0.1547 and 1.2942 1.2942 1.2942 ), in accordance with the case c 2 + = G Z c_2^+=G_\mathcal{Z} c 2 + = G Z of Conjecture 5.24 : for positive operators the violation stays at the order δ 2 \delta^2 δ 2 , and the linear failure of item 8 is a feature of the extension to arbitrary operators.
For p = 2 p=2 p = 2 and self-adjoint B B B the constant is ( 4 G Z 2 − 3 ) 1 / 2 > G Z (4G_\mathcal{Z}^2-3)^{1/2}>G_\mathcal{Z} ( 4 G Z 2 − 3 ) 1/2 > G Z . So if Conjecture 5.24 holds, the positivity of B B B cannot be weakened to self-adjointness without changing the constant.
Proposition (Item 31). For A ∈ M A\in\mathcal{M} A ∈ M and B ∈ M s a B\in\mathcal{M}_{sa} B ∈ M s a , φ ( ∣ A B ∣ ) ≤ ( κ Z − 1 + κ Z − 1 ) 1 / 2 φ ( ∣ A ∣ 2 ) 1 / 2 φ ( B 2 ) 1 / 2 . \varphi(|AB|)\le\bigl(\kappa_\mathcal{Z}-1+\kappa_\mathcal{Z}^{-1}\bigr)^{1/2}\varphi(|A|^2)^{1/2}\varphi(B^2)^{1/2}. φ ( ∣ A B ∣ ) ≤ ( κ Z − 1 + κ Z − 1 ) 1/2 φ ( ∣ A ∣ 2 ) 1/2 φ ( B 2 ) 1/2 . The constant ( κ Z − 1 + κ Z − 1 ) 1 / 2 = ( 4 G Z 2 − 3 ) 1 / 2 (\kappa_\mathcal{Z}-1+\kappa_\mathcal{Z}^{-1})^{1/2}=(4G_\mathcal{Z}^2-3)^{1/2} ( κ Z − 1 + κ Z − 1 ) 1/2 = ( 4 G Z 2 − 3 ) 1/2 is optimal, already for A ≥ 0 A\ge0 A ≥ 0 , and in M n ( C ) M_n(\mathbb{C}) M n ( C ) it is attained.
Proof. Let A B = W ∣ A B ∣ AB=W|AB| A B = W ∣ A B ∣ be the polar decomposition. Since ∣ A B ∣ = W ∗ A B |AB|=W^*AB ∣ A B ∣ = W ∗ A B and τ ( W Y W ∗ ) ≤ τ ( Y ) \tau(WYW^*)\le\tau(Y) τ ( W Y W ∗ ) ≤ τ ( Y ) for Y ≥ 0 Y\ge0 Y ≥ 0 ,
φ ( ∣ A B ∣ ) = τ ( A ρ 1 / 2 ⋅ ρ − 1 / 2 B ρ W ∗ ) ≤ ∥ A ρ 1 / 2 ∥ 2 ∥ ρ − 1 / 2 B ρ W ∗ ∥ 2 ≤ φ ( ∣ A ∣ 2 ) 1 / 2 τ ( ρ 2 B ρ − 1 B ) 1 / 2 . \varphi(|AB|)=\tau\bigl(A\rho^{1/2}\cdot\rho^{-1/2}B\rho W^*\bigr)
\le\|A\rho^{1/2}\|_2\,\|\rho^{-1/2}B\rho W^*\|_2\le\varphi(|A|^2)^{1/2}\,\tau(\rho^2B\rho^{-1}B)^{1/2}. φ ( ∣ A B ∣ ) = τ ( A ρ 1/2 ⋅ ρ − 1/2 B ρ W ∗ ) ≤ ∥ A ρ 1/2 ∥ 2 ∥ ρ − 1/2 B ρ W ∗ ∥ 2 ≤ φ ( ∣ A ∣ 2 ) 1/2 τ ( ρ 2 B ρ − 1 B ) 1/2 . On L 2 ( M , τ ) L^2(\mathcal{M},\tau) L 2 ( M , τ ) let J h = h ∗ Jh=h^* J h = h ∗ , so that J L x J = R x ∗ JL_xJ=R_{x^*} J L x J = R x ∗ . For B = B ∗ B=B^* B = B ∗ and the positive operator Y = L ρ 2 R ρ − 1 Y=L_{\rho^2}R_{\rho^{-1}} Y = L ρ 2 R ρ − 1 we have ⟨ Y B , B ⟩ = ⟨ J Y J B , B ⟩ \langle YB,B\rangle=\langle JYJB,B\rangle ⟨ Y B , B ⟩ = ⟨ J Y J B , B ⟩ . Hence τ ( ρ 2 B ρ − 1 B ) = ⟨ Y B , B ⟩ = ⟨ g ( L ρ , R ρ ) B , B ⟩ \tau(\rho^2B\rho^{-1}B)=\langle YB,B\rangle=\langle g(L_\rho,R_\rho)B,B\rangle τ ( ρ 2 B ρ − 1 B ) = ⟨ Y B , B ⟩ = ⟨ g ( L ρ , R ρ ) B , B ⟩ with g ( s , t ) = 1 2 ( s 2 / t + t 2 / s ) g(s,t)=\frac12(s^2/t+t^2/s) g ( s , t ) = 2 1 ( s 2 / t + t 2 / s ) , while φ ( B 2 ) = ⟨ h ( L ρ , R ρ ) B , B ⟩ \varphi(B^2)=\langle h(L_\rho,R_\rho)B,B\rangle φ ( B 2 ) = ⟨ h ( L ρ , R ρ ) B , B ⟩ with h ( s , t ) = 1 2 ( s + t ) h(s,t)=\frac12(s+t) h ( s , t ) = 2 1 ( s + t ) . On [ ∥ ρ − 1 ∥ − 1 , ∥ ρ ∥ ] 2 [\|\rho^{-1}\|^{-1},\|\rho\|]^2 [ ∥ ρ − 1 ∥ − 1 , ∥ ρ ∥ ] 2 we have g / h = s / t + t / s − 1 ≤ κ − 1 + κ − 1 g/h=s/t+t/s-1\le\kappa-1+\kappa^{-1} g / h = s / t + t / s − 1 ≤ κ − 1 + κ − 1 , which gives the bound with κ \kappa κ in place of κ Z \kappa_\mathcal{Z} κ Z . (R) holds with σ z ( A , B ) = ( z 1 / 2 A , z 1 / 2 B ) \sigma_z(A,B)=(z^{1/2}A,z^{1/2}B) σ z ( A , B ) = ( z 1/2 A , z 1/2 B ) , so Lemma 4.12 gives the bound with κ Z \kappa_\mathcal{Z} κ Z . For the lower bound let ρ = d i a g ( κ , 1 ) \rho=\mathop{\mathrm{diag}}(\kappa,1) ρ = diag ( κ , 1 ) in M 2 M_2 M 2 , A = d i a g ( κ − 1 , κ ) A=\mathop{\mathrm{diag}}(\kappa^{-1},\kappa) A = diag ( κ − 1 , κ ) and B = σ B=\sigma B = σ . Then ∣ A B ∣ = d i a g ( κ , κ − 1 ) |AB|=\mathop{\mathrm{diag}}(\kappa,\kappa^{-1}) ∣ A B ∣ = diag ( κ , κ − 1 ) , φ ( ∣ A B ∣ ) = φ ( A 2 ) = κ 2 + κ − 1 \varphi(|AB|)=\varphi(A^2)=\kappa^2+\kappa^{-1} φ ( ∣ A B ∣ ) = φ ( A 2 ) = κ 2 + κ − 1 and φ ( B 2 ) = κ + 1 \varphi(B^2)=\kappa+1 φ ( B 2 ) = κ + 1 , so the squared ratio is ( κ 2 + κ − 1 ) / ( κ + 1 ) = κ − 1 + κ − 1 (\kappa^2+\kappa^{-1})/(\kappa+1)=\kappa-1+\kappa^{-1} ( κ 2 + κ − 1 ) / ( κ + 1 ) = κ − 1 + κ − 1 . Theorem 4.15 and Remark 4.16 complete the proof, and κ + κ − 1 = 4 G Z 2 − 2 \kappa+\kappa^{-1}=4G_\mathcal{Z}^2-2 κ + κ − 1 = 4 G Z 2 − 2 . ◻
Monotonicity
For p > 1 p>1 p > 1 let
m p ( ρ ) = sup { φ ( A p ) φ ( B p ) : 0 ≤ A ≤ B , B ≠ 0 } , m exp ( ρ ) = sup { φ ( e A ) φ ( e B ) : A ≤ B in M s a } . m_p(\rho)=\sup\Bigl\{\frac{\varphi(A^p)}{\varphi(B^p)}:0\le A\le B,\ B\ne0\Bigr\},\qquad
m_{\exp}(\rho)=\sup\Bigl\{\frac{\varphi(e^A)}{\varphi(e^B)}:A\le B\text{ in }\mathcal{M}_{sa}\Bigr\}. m p ( ρ ) = sup { φ ( B p ) φ ( A p ) : 0 ≤ A ≤ B , B = 0 } , m e x p ( ρ ) = sup { φ ( e B ) φ ( e A ) : A ≤ B in M s a } .
Theorem 5.27 (Item 10). m 2 ( ρ ) = 1 1 − γ ( ρ ) 2 = ( G Z + 1 ) 2 4 G Z \displaystyle m_2(\rho)=\frac1{1-\gamma(\rho)^2}=\frac{(G_\mathcal{Z}+1)^2}{4G_\mathcal{Z}} m 2 ( ρ ) = 1 − γ ( ρ ) 2 1 = 4 G Z ( G Z + 1 ) 2 , where γ ( ρ ) = tanh 2 ( L N C ( ρ ) / 2 ) \gamma(\rho)=\tanh^2(\mathop{\mathrm{LNC}}(\rho)/2) γ ( ρ ) = tanh 2 ( LNC ( ρ ) /2 ) , which is the constant of Theorem (?) if M ≠ C 1 \mathcal{M}\ne\mathbb{C}1 M = C 1 . In M n ( C ) M_n(\mathbb{C}) M n ( C ) the supremum is attained.
Proof. If M = C 1 \mathcal{M}=\mathbb{C}1 M = C 1 , then R e φ ( A C ) ≥ 0 \mathop{\mathrm{Re}}\varphi(AC)\ge0 Re φ ( A C ) ≥ 0 for A , C ≥ 0 A,C\ge0 A , C ≥ 0 , so φ ( A 2 ) ≤ φ ( B 2 ) \varphi(A^2)\le\varphi(B^2) φ ( A 2 ) ≤ φ ( B 2 ) whenever 0 ≤ A ≤ B 0\le A\le B 0 ≤ A ≤ B , and m 2 = 1 = ( G Z + 1 ) 2 / ( 4 G Z ) m_2=1=(G_\mathcal{Z}+1)^2/(4G_\mathcal{Z}) m 2 = 1 = ( G Z + 1 ) 2 / ( 4 G Z ) since G Z = 1 G_\mathcal{Z}=1 G Z = 1 ; let M ≠ C 1 \mathcal{M}\ne\mathbb{C}1 M = C 1 . Let 0 ≤ A ≤ B 0\le A\le B 0 ≤ A ≤ B , C = B − A ≥ 0 C=B-A\ge0 C = B − A ≥ 0 , a = φ ( A 2 ) a=\varphi(A^2) a = φ ( A 2 ) , c = φ ( C 2 ) c=\varphi(C^2) c = φ ( C 2 ) . By Theorem (?) , φ ( B 2 ) = a + 2 R e φ ( A C ) + c ≥ a − 2 γ a c + c ≥ ( 1 − γ 2 ) a \varphi(B^2)=a+2\mathop{\mathrm{Re}}\varphi(AC)+c\ge a-2\gamma\sqrt{ac}+c\ge(1-\gamma^2)a φ ( B 2 ) = a + 2 Re φ ( A C ) + c ≥ a − 2 γ a c + c ≥ ( 1 − γ 2 ) a . For the converse, in M 2 M_2 M 2 with ρ = d i a g ( κ , 1 ) \rho=\mathop{\mathrm{diag}}(\kappa,1) ρ = diag ( κ , 1 ) , take the pair A = ∣ a ⟩ ⟨ a ∣ A=|a\rangle\langle a| A = ∣ a ⟩ ⟨ a ∣ , B ′ = ∣ b ⟩ ⟨ b ∣ B'=|b\rangle\langle b| B ′ = ∣ b ⟩ ⟨ b ∣ of the proof of Theorem (?) , for which R e φ ( A B ′ ) = − γ φ ( A 2 ) \mathop{\mathrm{Re}}\varphi(AB')=-\gamma\varphi(A^2) Re φ ( A B ′ ) = − γ φ ( A 2 ) and φ ( B ′ 2 ) = φ ( A 2 ) \varphi(B'^2)=\varphi(A^2) φ ( B ′2 ) = φ ( A 2 ) , and put B = A + γ B ′ B=A+\gamma B' B = A + γ B ′ . Then φ ( B 2 ) = ( 1 − γ 2 ) φ ( A 2 ) \varphi(B^2)=(1-\gamma^2)\varphi(A^2) φ ( B 2 ) = ( 1 − γ 2 ) φ ( A 2 ) . (R) holds with σ z ( A , B ) = ( z 1 / 2 A , z 1 / 2 B ) \sigma_z(A,B)=(z^{1/2}A,z^{1/2}B) σ z ( A , B ) = ( z 1/2 A , z 1/2 B ) , and Theorem 4.15 applies. Finally 1 − γ 2 = 4 G Z / ( G Z + 1 ) 2 1-\gamma^2=4G_\mathcal{Z}/(G_\mathcal{Z}+1)^2 1 − γ 2 = 4 G Z / ( G Z + 1 ) 2 since γ = ( G Z − 1 ) / ( G Z + 1 ) \gamma=(G_\mathcal{Z}-1)/(G_\mathcal{Z}+1) γ = ( G Z − 1 ) / ( G Z + 1 ) . ◻
Let m p ( κ ) = m p ( d i a g ( κ , 1 ) ) \mathsf m_p(\kappa)=m_p(\mathop{\mathrm{diag}}(\kappa,1)) m p ( κ ) = m p ( diag ( κ , 1 )) computed in M 2 ( C ) M_2(\mathbb{C}) M 2 ( C ) , and C p = ( p − 1 ) p − 1 / p / p p + 1 C_p=(p-1)^{p-1/p}/p^{p+1} C p = ( p − 1 ) p − 1/ p / p p + 1 . For h > 1 h>1 h > 1 and p > 1 p>1 p > 1 let
K ( h , p ) = h p − h ( p − 1 ) ( h − 1 ) ( p − 1 p h p − 1 h p − h ) p , S ( h ) = ( h − 1 ) h 1 / ( h − 1 ) e ln h , K(h,p)=\frac{h^p-h}{(p-1)(h-1)}\Bigl(\frac{p-1}p\,\frac{h^p-1}{h^p-h}\Bigr)^p,\qquad
S(h)=\frac{(h-1)h^{1/(h-1)}}{e\ln h}, K ( h , p ) = ( p − 1 ) ( h − 1 ) h p − h ( p p − 1 h p − h h p − 1 ) p , S ( h ) = e ln h ( h − 1 ) h 1/ ( h − 1 ) ,
and K ( 1 , p ) = S ( 1 ) = 1 K(1,p)=S(1)=1 K ( 1 , p ) = S ( 1 ) = 1 . These are the generalised Kantorovich constant (Ky Fan’s constant, see [112 ] ) and the Specht ratio. We call M \mathcal{M} M approximately finite-dimensional (AFD) if it contains an increasing net of finite-dimensional ∗ * ∗ -subalgebras containing 1 1 1 whose union is weakly dense. Examples are finite-dimensional algebras and the hyperfinite II1 _1 1 factor.
Proposition 5.29 (Item 32).
m p ( z ρ ) = m p ( ρ ) m_p(z\rho)=m_p(\rho) m p ( z ρ ) = m p ( ρ ) and m exp ( z ρ ) = m exp ( ρ ) m_{\exp}(z\rho)=m_{\exp}(\rho) m e x p ( z ρ ) = m e x p ( ρ ) for z ∈ Z + + z\in\mathcal{Z}_{++} z ∈ Z ++ .
p ↦ m p ( ρ ) p\mapsto m_p(\rho) p ↦ m p ( ρ ) is nondecreasing, and m exp ( ρ ) = sup p m p ( ρ ) = lim p → ∞ m p ( ρ ) m_{\exp}(\rho)=\sup_pm_p(\rho)=\lim_{p\to\infty}m_p(\rho) m e x p ( ρ ) = sup p m p ( ρ ) = lim p → ∞ m p ( ρ ) .
m p ( κ Z ) ≤ m p ( ρ ) ≤ κ Z \mathsf m_p(\kappa_\mathcal{Z})\le m_p(\rho)\le\kappa_\mathcal{Z} m p ( κ Z ) ≤ m p ( ρ ) ≤ κ Z , and m p ( κ ) > 1 \mathsf m_p(\kappa)>1 m p ( κ ) > 1 for κ > 1 \kappa>1 κ > 1 . Moreover m p ( κ ) ≥ C p κ 1 − 1 / p ( 1 + o ( 1 ) ) \mathsf m_p(\kappa)\ge C_p\kappa^{1-1/p}(1+o(1)) m p ( κ ) ≥ C p κ 1 − 1/ p ( 1 + o ( 1 )) as κ → ∞ \kappa\to\infty κ → ∞ .
For κ Z ≥ e \kappa_\mathcal{Z}\ge e κ Z ≥ e , κ Z e ( ln κ Z + ln ln κ Z ) ( 1 + 1 / ln κ Z ) ≤ m exp ( ρ ) ≤ κ Z ; \frac{\kappa_\mathcal{Z}}{e\,(\ln\kappa_\mathcal{Z}+\ln\ln\kappa_\mathcal{Z})(1+1/\ln\kappa_\mathcal{Z})}\le m_{\exp}(\rho)\le\kappa_\mathcal{Z}; e ( l n κ Z + l n l n κ Z ) ( 1 + 1/ l n κ Z ) κ Z ≤ m e x p ( ρ ) ≤ κ Z ; in particular m exp ( ρ ) ≥ ( 1 − o ( 1 ) ) κ Z / ( e ln κ Z ) m_{\exp}(\rho)\ge(1-o(1))\kappa_\mathcal{Z}/(e\ln\kappa_\mathcal{Z}) m e x p ( ρ ) ≥ ( 1 − o ( 1 )) κ Z / ( e ln κ Z ) as κ Z → ∞ \kappa_\mathcal{Z}\to\infty κ Z → ∞ .
If M \mathcal{M} M is AFD, then m p ( ρ ) ≤ K ( κ Z 1 / p , p ) m_p(\rho)\le K(\kappa_\mathcal{Z}^{1/p},p) m p ( ρ ) ≤ K ( κ Z 1/ p , p ) and m exp ( ρ ) ≤ S ( κ Z ) m_{\exp}(\rho)\le S(\kappa_\mathcal{Z}) m e x p ( ρ ) ≤ S ( κ Z ) . Consequently, on AFD algebras, as κ Z → ∞ \kappa_\mathcal{Z}\to\infty κ Z → ∞ , C p κ Z 1 − 1 / p ( 1 + o ( 1 ) ) ≤ m p ( ρ ) ≤ ( p − 1 ) p − 1 p p κ Z 1 − 1 / p ( 1 + o ( 1 ) ) , m exp ( ρ ) = ( 1 + o ( 1 ) ) κ Z e ln κ Z . C_p\kappa_\mathcal{Z}^{1-1/p}(1+o(1))\le m_p(\rho)\le\frac{(p-1)^{p-1}}{p^p}\,\kappa_\mathcal{Z}^{1-1/p}(1+o(1)),
\qquad m_{\exp}(\rho)=(1+o(1))\frac{\kappa_\mathcal{Z}}{e\ln\kappa_\mathcal{Z}}. C p κ Z 1 − 1/ p ( 1 + o ( 1 )) ≤ m p ( ρ ) ≤ p p ( p − 1 ) p − 1 κ Z 1 − 1/ p ( 1 + o ( 1 )) , m e x p ( ρ ) = ( 1 + o ( 1 )) e l n κ Z κ Z .
Proof. m:a Use ( A , B ) ↦ ( z 1 / p A , z 1 / p B ) (A,B)\mapsto(z^{1/p}A,z^{1/p}B) ( A , B ) ↦ ( z 1/ p A , z 1/ p B ) , resp. ( A + ln z , B + ln z ) (A+\ln z,B+\ln z) ( A + ln z , B + ln z ) . m:b If 1 < p < r 1<p<r 1 < p < r and 0 ≤ A ≤ B 0\le A\le B 0 ≤ A ≤ B , then A p / r ≤ B p / r A^{p/r}\le B^{p/r} A p / r ≤ B p / r by the Löwner–Heinz inequality, so φ ( A p ) / φ ( B p ) ≤ m r \varphi(A^p)/\varphi(B^p)\le m_r φ ( A p ) / φ ( B p ) ≤ m r . If 0 < A ≤ B 0<A\le B 0 < A ≤ B , then p ln A ≤ p ln B p\ln A\le p\ln B p ln A ≤ p ln B and φ ( A p ) / φ ( B p ) = φ ( e p ln A ) / φ ( e p ln B ) ≤ m exp \varphi(A^p)/\varphi(B^p)=\varphi(e^{p\ln A})/\varphi(e^{p\ln B})\le m_{\exp} φ ( A p ) / φ ( B p ) = φ ( e p l n A ) / φ ( e p l n B ) ≤ m e x p ; singular A A A is handled by A + ε A+\varepsilon A + ε . Conversely, for A ≤ B A\le B A ≤ B and large p p p , 0 ≤ 1 + A / p ≤ 1 + B / p 0\le1+A/p\le1+B/p 0 ≤ 1 + A / p ≤ 1 + B / p and ( 1 + A / p ) p → e A (1+A/p)^p\to e^A ( 1 + A / p ) p → e A in norm, so φ ( e A ) ≤ ( sup p m p ) φ ( e B ) \varphi(e^A)\le(\sup_pm_p)\varphi(e^B) φ ( e A ) ≤ ( sup p m p ) φ ( e B ) . m:c The lower bound is Theorem 4.15 t:low , and m p ( κ ) > 1 \mathsf m_p(\kappa)>1 m p ( κ ) > 1 for κ > 1 \kappa>1 κ > 1 by the theorem of Bikchentaev and Tikhonov [111 ] . For the upper bound, τ ( 1 Q A p ) ≤ τ ( 1 Q B p ) \tau(\mathbf 1_QA^p)\le\tau(\mathbf 1_QB^p) τ ( 1 Q A p ) ≤ τ ( 1 Q B p ) for every central projection Q Q Q by the monotonicity of τ ( f ( ⋅ ) ) \tau(f(\cdot)) τ ( f ( ⋅ )) for increasing f f f in the finite algebra Q M Q\mathcal{M} Q M [121 ] ; approximating central positive elements by positive combinations of central projections, τ ( u A p ) ≤ τ ( u B p ) \tau(uA^p)\le\tau(uB^p) τ ( u A p ) ≤ τ ( u B p ) , and by (3) φ ( A p ) ≤ τ ( u A p ) ≤ τ ( u B p ) ≤ κ Z τ ( l B p ) ≤ κ Z φ ( B p ) \varphi(A^p)\le\tau(uA^p)\le\tau(uB^p)\le\kappa_\mathcal{Z}\tau(lB^p)\le\kappa_\mathcal{Z}\varphi(B^p) φ ( A p ) ≤ τ ( u A p ) ≤ τ ( u B p ) ≤ κ Z τ ( l B p ) ≤ κ Z φ ( B p ) . For the asymptotics, let ρ = d i a g ( κ , 1 ) \rho=\mathop{\mathrm{diag}}(\kappa,1) ρ = diag ( κ , 1 ) , B = d i a g ( β , 1 ) B=\mathop{\mathrm{diag}}(\beta,1) B = diag ( β , 1 ) , w = ( c ′ , s ′ ) w=(c',s') w = ( c ′ , s ′ ) a unit vector and A = λ ∣ w ⟩ ⟨ w ∣ A=\lambda|w\rangle\langle w| A = λ ∣ w ⟩ ⟨ w ∣ with λ = ⟨ B − 1 w , w ⟩ − 1 \lambda=\langle B^{-1}w,w\rangle^{-1} λ = ⟨ B − 1 w , w ⟩ − 1 , the largest λ \lambda λ with A ≤ B A\le B A ≤ B . Then
φ ( A p ) φ ( B p ) = κ c ′ 2 + s ′ 2 ( c ′ 2 / β + s ′ 2 ) p ( κ β p + 1 ) . \begin{equation}
\frac{\varphi(A^p)}{\varphi(B^p)}=\frac{\kappa c'^2+s'^2}{(c'^2/\beta+s'^2)^p(\kappa\beta^p+1)}. \tag*{(7)}
\end{equation} φ ( B p ) φ ( A p ) = ( c ′2 / β + s ′2 ) p ( κ β p + 1 ) κ c ′2 + s ′2 . (7) With c ′ 2 = θ β c'^2=\theta\beta c ′2 = θ β , κ β p = y \kappa\beta^p=y κ β p = y , θ = y = 1 / ( p − 1 ) \theta=y=1/(p-1) θ = y = 1/ ( p − 1 ) and κ → ∞ \kappa\to\infty κ → ∞ this is θ y 1 / p κ 1 − 1 / p ( 1 + θ ) − p ( 1 + y ) − 1 ( 1 + o ( 1 ) ) = C p κ 1 − 1 / p ( 1 + o ( 1 ) ) \theta y^{1/p}\kappa^{1-1/p}(1+\theta)^{-p}(1+y)^{-1}(1+o(1))=C_p\kappa^{1-1/p}(1+o(1)) θ y 1/ p κ 1 − 1/ p ( 1 + θ ) − p ( 1 + y ) − 1 ( 1 + o ( 1 )) = C p κ 1 − 1/ p ( 1 + o ( 1 )) . m:d The upper bound follows from m:b and m:c . For the lower bound take β = e − σ / p \beta=e^{-\sigma/p} β = e − σ / p in (7) with fixed σ ≥ 1 \sigma\ge1 σ ≥ 1 and c ′ 2 = 1 / σ c'^2=1/\sigma c ′2 = 1/ σ ; as p → ∞ p\to\infty p → ∞ , ( c ′ 2 e σ / p + s ′ 2 ) p → e σ c ′ 2 = e (c'^2e^{\sigma/p}+s'^2)^p\to e^{\sigma c'^2}=e ( c ′2 e σ / p + s ′2 ) p → e σ c ′2 = e , so by m:b and m:c m exp ( ρ ) ≥ sup p m p ( κ Z ) ≥ ( κ Z / σ + 1 − 1 / σ ) / ( e ( κ Z e − σ + 1 ) ) m_{\exp}(\rho)\ge\sup_p\mathsf m_p(\kappa_\mathcal{Z})\ge(\kappa_\mathcal{Z}/\sigma+1-1/\sigma)/(e(\kappa_\mathcal{Z}e^{-\sigma}+1)) m e x p ( ρ ) ≥ sup p m p ( κ Z ) ≥ ( κ Z / σ + 1 − 1/ σ ) / ( e ( κ Z e − σ + 1 )) . Choose e − σ = 1 / ( κ Z ln κ Z ) e^{-\sigma}=1/(\kappa_\mathcal{Z}\ln\kappa_\mathcal{Z}) e − σ = 1/ ( κ Z ln κ Z ) , which gives σ ≥ 1 \sigma\ge1 σ ≥ 1 for κ Z ≥ e \kappa_\mathcal{Z}\ge e κ Z ≥ e .
m:e By m:a and Proposition 4.5 ch:resc we may replace ρ \rho ρ by z 0 ρ z_0\rho z 0 ρ and assume κ ( ρ ) = κ Z \kappa(\rho)=\kappa_\mathcal{Z} κ ( ρ ) = κ Z . Let μ − = min σ ( ρ ) \mu_-=\min\sigma(\rho) μ − = min σ ( ρ ) and μ + = max σ ( ρ ) \mu_+=\max\sigma(\rho) μ + = max σ ( ρ ) . We first show that
φ ( A p ) ≤ K ( κ 1 / p , p ) τ ( ( ρ 1 / 2 p A ρ 1 / 2 p ) p ) ( A ∈ M + ) \begin{equation}
\varphi(A^p)\le K(\kappa^{1/p},p)\,\tau\bigl((\rho^{1/2p}A\rho^{1/2p})^p\bigr)\qquad(A\in\mathcal{M}_+) \tag*{(8)}
\end{equation} φ ( A p ) ≤ K ( κ 1/ p , p ) τ ( ( ρ 1/2 p A ρ 1/2 p ) p ) ( A ∈ M + ) (8) for every finite-dimensional M \mathcal{M} M and every ρ \rho ρ with σ ( ρ ) ⊂ [ μ − , μ + ] \sigma(\rho)\subset[\mu_-,\mu_+] σ ( ρ ) ⊂ [ μ − , μ + ] , where κ = μ + / μ − \kappa=\mu_+/\mu_- κ = μ + / μ − . Then M = ⨁ i M n i ( C ) \mathcal{M}=\bigoplus_iM_{n_i}(\mathbb{C}) M = ⨁ i M n i ( C ) and τ = ∑ i t i T r i \tau=\sum_it_i\mathop{\mathrm{Tr}}_i τ = ∑ i t i Tr i with t i > 0 t_i>0 t i > 0 . Bourin’s reverse of Araki’s inequality [112 Thm. 2] states the following. If A ′ ≥ 0 A'\ge0 A ′ ≥ 0 and Z > 0 Z>0 Z > 0 in M N ( C ) M_N(\mathbb{C}) M N ( C ) , Z Z Z has extremal eigenvalues a , b a,b a , b , and p > 1 p>1 p > 1 , then there is a unitary V V V with A ′ p Z p A ′ p ≤ K ( b / a , p ) V ( A ′ Z A ′ ) p V ∗ A'^pZ^pA'^p\le K(b/a,p)\,V(A'ZA')^pV^* A ′ p Z p A ′ p ≤ K ( b / a , p ) V ( A ′ Z A ′ ) p V ∗ (the constant there is written K ( a , b , p ) = K ( b / a , p ) K(a,b,p)=K(b/a,p) K ( a , b , p ) = K ( b / a , p ) ). Apply it in M n i + 2 ( C ) M_{n_i+2}(\mathbb{C}) M n i + 2 ( C ) to A ′ = A i 1 / 2 ⊕ 0 A'=A_i^{1/2}\oplus0 A ′ = A i 1/2 ⊕ 0 and Z = ρ i 1 / p ⊕ d i a g ( μ − 1 / p , μ + 1 / p ) Z=\rho_i^{1/p}\oplus\mathop{\mathrm{diag}}(\mu_-^{1/p},\mu_+^{1/p}) Z = ρ i 1/ p ⊕ diag ( μ − 1/ p , μ + 1/ p ) , whose extremal eigenvalues are μ − 1 / p \mu_-^{1/p} μ − 1/ p and μ + 1 / p \mu_+^{1/p} μ + 1/ p , and take traces. This gives T r i ( ρ i A i p ) = T r i ( A i p / 2 ρ i A i p / 2 ) ≤ K ( κ 1 / p , p ) T r i ( ( A i 1 / 2 ρ i 1 / p A i 1 / 2 ) p ) \mathop{\mathrm{Tr}}_i(\rho_iA_i^p)=\mathop{\mathrm{Tr}}_i(A_i^{p/2}\rho_iA_i^{p/2})\le K(\kappa^{1/p},p)\mathop{\mathrm{Tr}}_i((A_i^{1/2}\rho_i^{1/p}A_i^{1/2})^p) Tr i ( ρ i A i p ) = Tr i ( A i p /2 ρ i A i p /2 ) ≤ K ( κ 1/ p , p ) Tr i (( A i 1/2 ρ i 1/ p A i 1/2 ) p ) , and the last trace is T r i ( ( ρ i 1 / 2 p A i ρ i 1 / 2 p ) p ) \mathop{\mathrm{Tr}}_i((\rho_i^{1/2p}A_i\rho_i^{1/2p})^p) Tr i (( ρ i 1/2 p A i ρ i 1/2 p ) p ) . Multiplying by t i t_i t i and summing gives (8) . If M \mathcal{M} M is AFD with finite-dimensional subalgebras N α N_\alpha N α , let E α E_\alpha E α be the τ \tau τ -preserving conditional expectations onto N α N_\alpha N α . Then E α ( x ) → x E_\alpha(x)\to x E α ( x ) → x in ∥ ⋅ ∥ 2 \|\cdot\|_2 ∥ ⋅ ∥ 2 , hence strongly, with ∥ E α ( x ) ∥ ≤ ∥ x ∥ \|E_\alpha(x)\|\le\|x\| ∥ E α ( x ) ∥ ≤ ∥ x ∥ , and σ ( E α ( ρ ) ) ⊂ [ μ − , μ + ] \sigma(E_\alpha(\rho))\subset[\mu_-,\mu_+] σ ( E α ( ρ )) ⊂ [ μ − , μ + ] . Apply (8) in N α N_\alpha N α to E α ( ρ ) E_\alpha(\rho) E α ( ρ ) and E α ( A ) E_\alpha(A) E α ( A ) , and pass to the limit; products, continuous functional calculus and τ \tau τ are strongly continuous on bounded sets. Now let 0 ≤ A ≤ B 0\le A\le B 0 ≤ A ≤ B . By (8) , by the monotonicity of X ↦ τ ( f ( X ) ) X\mapsto\tau(f(X)) X ↦ τ ( f ( X )) for increasing f f f [121 ] , and by (2) with r = p r=p r = p ,
φ ( A p ) ≤ K τ ( ( ρ 1 / 2 p A ρ 1 / 2 p ) p ) ≤ K τ ( ( ρ 1 / 2 p B ρ 1 / 2 p ) p ) ≤ K τ ( ρ 1 / 2 B p ρ 1 / 2 ) = K φ ( B p ) , \varphi(A^p)\le K\tau\bigl((\rho^{1/2p}A\rho^{1/2p})^p\bigr)\le K\tau\bigl((\rho^{1/2p}B\rho^{1/2p})^p\bigr)
\le K\tau(\rho^{1/2}B^p\rho^{1/2})=K\varphi(B^p), φ ( A p ) ≤ K τ ( ( ρ 1/2 p A ρ 1/2 p ) p ) ≤ K τ ( ( ρ 1/2 p B ρ 1/2 p ) p ) ≤ K τ ( ρ 1/2 B p ρ 1/2 ) = K φ ( B p ) , with K = K ( κ Z 1 / p , p ) K=K(\kappa_\mathcal{Z}^{1/p},p) K = K ( κ Z 1/ p , p ) . For h > 1 h>1 h > 1 and x = h 1 / p x=h^{1/p} x = h 1/ p , as p → ∞ p\to\infty p → ∞ we have ( h − x ) / ( ( p − 1 ) ( x − 1 ) ) → ( h − 1 ) / ln h (h-x)/((p-1)(x-1))\to(h-1)/\ln h ( h − x ) / (( p − 1 ) ( x − 1 )) → ( h − 1 ) / ln h , ( ( p − 1 ) / p ) p → e − 1 ((p-1)/p)^p\to e^{-1} (( p − 1 ) / p ) p → e − 1 and ( ( h − 1 ) / ( h − x ) ) p → h 1 / ( h − 1 ) ((h-1)/(h-x))^p\to h^{1/(h-1)} (( h − 1 ) / ( h − x ) ) p → h 1/ ( h − 1 ) . Hence K ( h 1 / p , p ) → S ( h ) K(h^{1/p},p)\to S(h) K ( h 1/ p , p ) → S ( h ) , and by m:b , m exp = lim p m p ≤ S ( κ Z ) m_{\exp}=\lim_pm_p\le S(\kappa_\mathcal{Z}) m e x p = lim p m p ≤ S ( κ Z ) . The asymptotics follow from m:c , m:d , K ( h , p ) ∼ ( p − 1 ) p − 1 p p h p − 1 K(h,p)\sim\frac{(p-1)^{p-1}}{p^p}h^{p-1} K ( h , p ) ∼ p p ( p − 1 ) p − 1 h p − 1 as h → ∞ h\to\infty h → ∞ , and S ( h ) ∼ h / ( e ln h ) S(h)\sim h/(e\ln h) S ( h ) ∼ h / ( e ln h ) . ◻
Weighted monotonicity
Proposition 5.31 (Item 25). For item 25 (φ ( A 4 ) ≤ φ ( A B 2 A ) \varphi(A^4)\le\varphi(AB^2A) φ ( A 4 ) ≤ φ ( A B 2 A ) for 0 ≤ A ≤ B 0\le A\le B 0 ≤ A ≤ B ), rank-one A A A gives no violation, and an explicit construction with A A A of rank two and B − A B-A B − A of rank one gives V n ≥ g 25 ( δ ) = δ 8 4096 ( 1 + O ( δ 2 ) ) in every dimension n ≥ 2. V_n\ge g_{25}(\delta)=\frac{\delta^8}{4096}\bigl(1+O(\delta^2)\bigr)\qquad\text{in every dimension }n\ge2 . V n ≥ g 25 ( δ ) = 4096 δ 8 ( 1 + O ( δ 2 ) ) in every dimension n ≥ 2. Conversely, V n ≤ κ − 1 = 2 δ / ( 1 − δ ) V_n\le\kappa-1=2\delta/(1-\delta) V n ≤ κ − 1 = 2 δ / ( 1 − δ ) , and in every finite von Neumann algebra the constant c 25 ( ρ ) = sup φ ( A 4 ) / φ ( A B 2 A ) c_{25}(\rho)=\sup\varphi(A^4)/\varphi(AB^2A) c 25 ( ρ ) = sup φ ( A 4 ) / φ ( A B 2 A ) over 0 ≤ A ≤ B 0\le A\le B 0 ≤ A ≤ B satisfies c 25 ( ρ ) ≤ κ Z ( ρ ) c_{25}(\rho)\le\kappa_\mathcal{Z}(\rho) c 25 ( ρ ) ≤ κ Z ( ρ ) .
Proof. Upper bound. For a finite trace τ \tau τ and 0 ≤ A ≤ B 0\le A\le B 0 ≤ A ≤ B we have τ ( A 4 ) ≤ τ ( A B 2 A ) \tau(A^4)\le\tau(AB^2A) τ ( A 4 ) ≤ τ ( A B 2 A ) (cf. [134 ] ): congruence by A 1 / 2 A^{1/2} A 1/2 twice gives A 3 ≤ A B A A^3\le ABA A 3 ≤ A B A , so τ ( A 4 ) ≤ τ ( A 3 / 2 B A 3 / 2 ) = τ ( A 2 ⋅ A B ) ≤ τ ( A 4 ) 1 / 2 τ ( A B 2 A ) 1 / 2 \tau(A^4)\le\tau(A^{3/2}BA^{3/2})=\tau(A^2\cdot AB)\le
\tau(A^4)^{1/2}\tau(AB^2A)^{1/2} τ ( A 4 ) ≤ τ ( A 3/2 B A 3/2 ) = τ ( A 2 ⋅ A B ) ≤ τ ( A 4 ) 1/2 τ ( A B 2 A ) 1/2 by the Cauchy–Schwarz inequality for τ \tau τ . Since also A B 2 A ≥ 0 AB^2A\ge0 A B 2 A ≥ 0 , φ ( A 4 ) ≤ ∥ ρ ∥ τ ( A 4 ) ≤ ∥ ρ ∥ τ ( A B 2 A ) ≤ κ ( ρ ) φ ( A B 2 A ) \varphi(A^4)\le\|\rho\|\tau(A^4)\le\|\rho\|\tau(AB^2A)\le\kappa(\rho)\varphi(AB^2A) φ ( A 4 ) ≤ ∥ ρ ∥ τ ( A 4 ) ≤ ∥ ρ ∥ τ ( A B 2 A ) ≤ κ ( ρ ) φ ( A B 2 A ) . Condition (R) holds with σ z ( A , B ) = ( z 1 / 4 A , z 1 / 4 B ) \sigma_z(A,B)=(z^{1/4}A,z^{1/4}B) σ z ( A , B ) = ( z 1/4 A , z 1/4 B ) , which preserves 0 ≤ A ≤ B 0\le A\le B 0 ≤ A ≤ B and multiplies A 4 A^4 A 4 and A B 2 A AB^2A A B 2 A by z z z , so Lemma 4.12 gives c 25 ≤ κ Z c_{25}\le\kappa_\mathcal{Z} c 25 ≤ κ Z ; on M n M_n M n this is V n ≤ κ − 1 V_n\le\kappa-1 V n ≤ κ − 1 . Lower bound. If A = α u u ∗ A=\alpha uu^* A = α u u ∗ has rank one and A ≤ B A\le B A ≤ B , then u ∗ B u ≥ α u^*Bu\ge\alpha u ∗ B u ≥ α and A B 2 A = α 2 ∥ B u ∥ 2 u u ∗ ≥ α 2 ( u ∗ B u ) 2 u u ∗ ≥ A 4 AB^2A=\alpha^2\|Bu\|^2uu^*\ge\alpha^2(u^*Bu)^2uu^*\ge
A^4 A B 2 A = α 2 ∥ B u ∥ 2 u u ∗ ≥ α 2 ( u ∗ B u ) 2 u u ∗ ≥ A 4 , so every positive φ \varphi φ satisfies the inequality. For the construction take ρ \rho ρ with T r ρ = 1 \mathop{\mathrm{Tr}}\rho=1 Tr ρ = 1 and eigenvalues ( 1 ± δ ) / 2 (1\pm\delta)/2 ( 1 ± δ ) /2 , and, in an orthonormal basis at the angle ϑ \vartheta ϑ to its eigenbasis, A = d i a g ( 1 , s ) A=\mathrm{diag}(1,s) A = diag ( 1 , s ) , 0 < s < 1 0<s<1 0 < s < 1 , and B = A + t v v ∗ B=A+tvv^* B = A + t v v ∗ with v = ( cos ϕ , sin ϕ ) v=(\cos\phi,\sin\phi) v = ( cos ϕ , sin ϕ ) ; in this basis ρ \rho ρ has the entries ρ 11 , 22 = 1 2 ( 1 ± δ cos 2 ϑ ) \rho_{11,22}=\frac12(1\pm\delta\cos2\vartheta) ρ 11 , 22 = 2 1 ( 1 ± δ cos 2 ϑ ) and ρ 12 = δ 2 sin 2 ϑ \rho_{12}=\frac\delta2\sin2\vartheta ρ 12 = 2 δ sin 2 ϑ . Then R − L = t T 1 + t 2 T 2 \mathcal{R}-\mathcal{L}=tT_1+t^2T_2 R − L = t T 1 + t 2 T 2 with T 1 = φ ( A 2 P v A + A P v A 2 ) T_1=\varphi(A^2P_vA+AP_vA^2) T 1 = φ ( A 2 P v A + A P v A 2 ) , T 2 = φ ( A P v A ) T_2=\varphi(AP_vA) T 2 = φ ( A P v A ) , P v = v v ∗ P_v=vv^* P v = v v ∗ , and whenever T 1 < 0 T_1<0 T 1 < 0 the best t = − T 1 / ( 2 T 2 ) t=-T_1/(2T_2) t = − T 1 / ( 2 T 2 ) gives ( L − R ) / R = Γ / ( L − Γ ) (\mathcal{L}-\mathcal{R})/\mathcal{R}=\Gamma/(\mathcal{L}-\Gamma) ( L − R ) / R = Γ/ ( L − Γ ) , Γ = T 1 2 / ( 4 T 2 ) \Gamma=T_1^2/(4T_2) Γ = T 1 2 / ( 4 T 2 ) . Let g 25 ( δ ) g_{25}(\delta) g 25 ( δ ) be the maximum over ( ϑ , s , ϕ ) (\vartheta,s,\phi) ( ϑ , s , ϕ ) ; it is attained, and by embedding C 2 ⊂ C n \mathbb{C}^2\subset\mathbb{C}^n C 2 ⊂ C n (with A A A , B B B extended by 0 0 0 ) it is a lower bound in every dimension. For the leading order put ϑ = π / 4 \vartheta=\pi/4 ϑ = π /4 , s = σ δ \sqrt s=\sigma\delta s = σ δ and tan ϕ = − ϱ s − 3 / 2 \tan\phi=-\varrho
s^{-3/2} tan ϕ = − ϱ s − 3/2 . Then T 1 = cos 2 ϕ ( 1 + ϱ 2 − ( 1 + s ) ϱ / σ ) T_1=\cos^2\phi\,(1+\varrho^2-(1+s)\varrho/\sigma) T 1 = cos 2 ϕ ( 1 + ϱ 2 − ( 1 + s ) ϱ / σ ) , T 2 = cos 2 ϕ ( 1 2 + ϱ 2 2 s − ϱ σ ) T_2=\cos^2\phi\,(\frac12+\frac{\varrho^2}{2s}-\frac\varrho\sigma) T 2 = cos 2 ϕ ( 2 1 + 2 s ϱ 2 − σ ϱ ) and cos 2 ϕ = s 3 / ( s 3 + ϱ 2 ) \cos^2\phi=s^3/(s^3+\varrho^2) cos 2 ϕ = s 3 / ( s 3 + ϱ 2 ) , so that
\frac{\mathcal{L}-\mathcal{R}}\mathcal{R}=\bigl(\sigma^2(\varsigma-\varsigma^2)-\sigma^4\bigr)^2\delta^8+O(\delta^{10}),\qquad\varsigma=\sigma/\varrho,
which is maximal for ς = 1 2 \varsigma=\frac12 ς = 2 1 , σ 2 = 1 8 \sigma^2=\frac18 σ 2 = 8 1 , with the value δ 8 / 4096 \delta^8/4096 δ 8 /4096 . ◻
The numerical maximisation gives g 25 ( 0.1 ) / ( 0.1 8 / 4096 ) = 1.020 g_{25}(0.1)/(0.1^8/4096)=1.020 g 25 ( 0.1 ) / ( 0. 1 8 /4096 ) = 1.020 . The same scaling for a general exponent r ≥ 0 r\ge0 r ≥ 0 in φ ( A 2 r + 2 ) ≤ φ ( A r B 2 A r ) \varphi(A^{2r+2})\le\varphi(A^rB^2A^r) φ ( A 2 r + 2 ) ≤ φ ( A r B 2 A r ) gives the order δ 4 ( r + 1 ) \delta^{4(r+1)} δ 4 ( r + 1 ) ; for r = 0 r=0 r = 0 the supremum of the leading coefficient is 1 16 \frac1{16} 16 1 , that of item 10.
Projection inequalities
For projections P , Q ∈ M P,Q\in\mathcal{M} P , Q ∈ M let
c P Q P = sup φ ( P Q P ) φ ( Q ) , c P Q = sup Q P ≠ 0 φ ( P Q P ) φ ( Q P Q ) , c s = sup Q P ≠ 0 φ ( s ( P Q P ) ) φ ( s ( Q P Q ) ) , c ∨ = sup φ ( P ∨ Q ) φ ( P ) + φ ( Q ) . c_{PQP}=\sup\frac{\varphi(PQP)}{\varphi(Q)},\quad c_{PQ}=\sup_{QP\ne0}\frac{\varphi(PQP)}{\varphi(QPQ)},\quad
c_s=\sup_{QP\ne0}\frac{\varphi(s(PQP))}{\varphi(s(QPQ))},\quad c_\vee=\sup\frac{\varphi(P\vee Q)}{\varphi(P)+\varphi(Q)}. c P QP = sup φ ( Q ) φ ( P QP ) , c P Q = QP = 0 sup φ ( QP Q ) φ ( P QP ) , c s = QP = 0 sup φ ( s ( QP Q )) φ ( s ( P QP )) , c ∨ = sup φ ( P ) + φ ( Q ) φ ( P ∨ Q ) .
They correspond to the characterisations φ ( P Q P ) ≤ φ ( Q ) \varphi(PQP)\le\varphi(Q) φ ( P QP ) ≤ φ ( Q ) , φ ( P Q P ) = φ ( Q P Q ) \varphi(PQP)=\varphi(QPQ) φ ( P QP ) = φ ( QP Q ) and φ ( s ( P Q P ) ) = φ ( s ( Q P Q ) ) \varphi(s(PQP))=\varphi(s(QPQ)) φ ( s ( P QP )) = φ ( s ( QP Q )) in Bikchentaev’s series [95 , 96 , 97 , 100 ] and to the inequality φ ( P ∨ Q ) ≤ φ ( P ) + φ ( Q ) \varphi(P\vee Q)\le\varphi(P)+\varphi(Q) φ ( P ∨ Q ) ≤ φ ( P ) + φ ( Q ) , the formal specialisation of Tikhonov’s subadditivity [164 ] to f = 1 ( 0 , ∞ ) f=\mathbf 1_{(0,\infty)} f = 1 ( 0 , ∞ ) and projections (for traces it follows from Kaplansky’s formula). The constant for φ ( ∣ P − Q ∣ ) ≤ φ ( P + Q ) \varphi(|P-Q|)\le\varphi(P+Q) φ ( ∣ P − Q ∣ ) ≤ φ ( P + Q ) is c p r o j = G Z c_{\rm proj}=G_\mathcal{Z} c proj = G Z by Theorem 5.4 .
Theorem 5.32 (Items 17, 23 and 33).
P ρ P ≤ G Z 2 ρ P\rho P\le G_\mathcal{Z}^2\rho P ρP ≤ G Z 2 ρ for every projection P P P , and c P Q P = G Z 2 = ( 1 + κ Z ) 2 4 κ Z c_{PQP}=G_\mathcal{Z}^2=\frac{(1+\kappa_\mathcal{Z})^2}{4\kappa_\mathcal{Z}} c P QP = G Z 2 = 4 κ Z ( 1 + κ Z ) 2 ; in M n ( C ) M_n(\mathbb{C}) M n ( C ) the supremum is attained (on M n ( C ) M_n(\mathbb{C}) M n ( C ) the operator inequality is contained in [113 Cor. 1.5] ).
c P Q = c s = κ Z c_{PQ}=c_s=\kappa_\mathcal{Z} c P Q = c s = κ Z ; the corresponding infima are κ Z − 1 \kappa_\mathcal{Z}^{-1} κ Z − 1 . If M = M n ( C ) \mathcal{M}=M_n(\mathbb{C}) M = M n ( C ) and ρ \rho ρ is not scalar, neither c P Q c_{PQ} c P Q nor c s c_s c s is attained.
φ ( P ∨ Q ) ≤ min { φ ( Q ) + κ Z φ ( P ) , φ ( P ) + κ Z φ ( Q ) } \varphi(P\vee Q)\le\min\{\varphi(Q)+\kappa_\mathcal{Z}\varphi(P),\varphi(P)+\kappa_\mathcal{Z}\varphi(Q)\} φ ( P ∨ Q ) ≤ min { φ ( Q ) + κ Z φ ( P ) , φ ( P ) + κ Z φ ( Q )} and c ∨ = 1 + κ Z 2 c_\vee=\frac{1+\kappa_\mathcal{Z}}2 c ∨ = 2 1 + κ Z .
Proof. p:a By Lemma 3.3 for T = z 0 ρ T=z_0\rho T = z 0 ρ with z 0 z_0 z 0 as in Lemma 4.12 , P z 0 ρ P ≤ cosh 2 O L C ( z 0 ρ ) z 0 ρ = G Z 2 z 0 ρ Pz_0\rho P\le\cosh^2\mathop{\mathrm{OLC}}(z_0\rho)z_0\rho=G_\mathcal{Z}^2z_0\rho P z 0 ρP ≤ cosh 2 OLC ( z 0 ρ ) z 0 ρ = G Z 2 z 0 ρ ; multiply by z 0 − 1 / 2 z_0^{-1/2} z 0 − 1/2 on both sides. Then φ ( P Q P ) = τ ( Q P ρ P Q ) ≤ G Z 2 τ ( Q ρ Q ) = G Z 2 φ ( Q ) \varphi(PQP)=\tau(QP\rho PQ)\le G_\mathcal{Z}^2\tau(Q\rho Q)=G_\mathcal{Z}^2\varphi(Q) φ ( P QP ) = τ ( QP ρP Q ) ≤ G Z 2 τ ( QρQ ) = G Z 2 φ ( Q ) . For the lower bound, in M 2 M_2 M 2 with ρ = d i a g ( κ , 1 ) \rho=\mathop{\mathrm{diag}}(\kappa,1) ρ = diag ( κ , 1 ) let P = ∣ w ⟩ ⟨ w ∣ P=|w\rangle\langle w| P = ∣ w ⟩ ⟨ w ∣ , w = ( 1 , 1 ) / 2 w=(1,1)/\sqrt2 w = ( 1 , 1 ) / 2 , and Q = ∣ q ⟩ ⟨ q ∣ Q=|q\rangle\langle q| Q = ∣ q ⟩ ⟨ q ∣ with q = ρ − 1 w / ∥ ρ − 1 w ∥ q=\rho^{-1}w/\|\rho^{-1}w\| q = ρ − 1 w /∥ ρ − 1 w ∥ . Then P Q P = ∣ ⟨ w , q ⟩ ∣ 2 P PQP=|\langle w,q\rangle|^2P P QP = ∣ ⟨ w , q ⟩ ∣ 2 P and the ratio is ⟨ ρ − 1 w , w ⟩ ⟨ ρ w , w ⟩ = ( κ + 1 ) 2 / ( 4 κ ) \langle\rho^{-1}w,w\rangle\langle\rho w,w\rangle=(\kappa+1)^2/(4\kappa) ⟨ ρ − 1 w , w ⟩ ⟨ ρw , w ⟩ = ( κ + 1 ) 2 / ( 4 κ ) , the Kantorovich expression, which here attains its maximum G 2 G^2 G 2 . Apply Theorem 4.15 t:low .
p:b With X = Q P X=QP X = QP , ∣ X ∣ 2 = P Q P |X|^2=PQP ∣ X ∣ 2 = P QP and ∣ X ∗ ∣ 2 = Q P Q |X^*|^2=QPQ ∣ X ∗ ∣ 2 = QP Q , so c P Q ≤ κ Z c_{PQ}\le\kappa_\mathcal{Z} c P Q ≤ κ Z by Theorem (?) . If X = w ∣ X ∣ X=w|X| X = w ∣ X ∣ , then s ( P Q P ) = w ∗ w s(PQP)=w^*w s ( P QP ) = w ∗ w , s ( Q P Q ) = w w ∗ s(QPQ)=ww^* s ( QP Q ) = w w ∗ , and φ ( w ∗ w ) ≤ κ Z φ ( w w ∗ ) \varphi(w^*w)\le\kappa_\mathcal{Z}\varphi(ww^*) φ ( w ∗ w ) ≤ κ Z φ ( w w ∗ ) by Theorem (?) with U = w ∗ U=w^* U = w ∗ , Y = w w ∗ Y=ww^* Y = w w ∗ . For the lower bound, in M 2 M_2 M 2 with ρ = d i a g ( κ , 1 ) \rho=\mathop{\mathrm{diag}}(\kappa,1) ρ = diag ( κ , 1 ) , let P , Q P,Q P , Q be the projections onto ( cos θ , sin θ ) (\cos\theta,\sin\theta) ( cos θ , sin θ ) and ( sin θ , cos θ ) (\sin\theta,\cos\theta) ( sin θ , cos θ ) , 0 < θ < π / 4 0<\theta<\pi/4 0 < θ < π /4 . Then P Q P = sin 2 2 θ P PQP=\sin^22\theta\,P P QP = sin 2 2 θ P , Q P Q = sin 2 2 θ Q QPQ=\sin^22\theta\,Q QP Q = sin 2 2 θ Q , s ( P Q P ) = P s(PQP)=P s ( P QP ) = P , s ( Q P Q ) = Q s(QPQ)=Q s ( QP Q ) = Q , and both ratios equal ( κ cos 2 θ + sin 2 θ ) / ( κ sin 2 θ + cos 2 θ ) → κ (\kappa\cos^2\theta+\sin^2\theta)/(\kappa\sin^2\theta+\cos^2\theta)\to\kappa ( κ cos 2 θ + sin 2 θ ) / ( κ sin 2 θ + cos 2 θ ) → κ as θ → 0 \theta\to0 θ → 0 . Non-attainment: equality in τ ( ρ P Q P ) ≤ ∥ ρ ∥ τ ( P Q P ) = ∥ ρ ∥ τ ( Q P Q ) ≤ κ τ ( ρ Q P Q ) \tau(\rho PQP)\le\|\rho\|\tau(PQP)=\|\rho\|\tau(QPQ)\le\kappa\tau(\rho QPQ) τ ( ρP QP ) ≤ ∥ ρ ∥ τ ( P QP ) = ∥ ρ ∥ τ ( QP Q ) ≤ κ τ ( ρQP Q ) forces ran P Q P ⊂ E M \operatorname{ran}PQP\subset E_M ran P QP ⊂ E M and ran Q P Q ⊂ E m \operatorname{ran}QPQ\subset E_m ran QP Q ⊂ E m , the eigenspaces of M = max σ ( ρ ) M=\max\sigma(\rho) M = max σ ( ρ ) and m = min σ ( ρ ) m=\min\sigma(\rho) m = min σ ( ρ ) . Take x ∈ ran Q x\in\operatorname{ran}Q x ∈ ran Q with P x ≠ 0 Px\ne0 P x = 0 , put T = Q P Q T=QPQ T = QP Q and y = T x ∈ E m y=Tx\in E_m y = T x ∈ E m ; y ≠ 0 y\ne0 y = 0 since ⟨ T x , x ⟩ = ∥ P x ∥ 2 \langle Tx,x\rangle=\|Px\|^2 ⟨ T x , x ⟩ = ∥ P x ∥ 2 . Then P y = P Q P x ∈ E M ⊥ E m Py=PQP x\in E_M\perp
E_m P y = P QP x ∈ E M ⊥ E m , so ∥ P y ∥ 2 = ⟨ P y , y ⟩ = 0 \|Py\|^2=\langle Py,y\rangle=0 ∥ P y ∥ 2 = ⟨ P y , y ⟩ = 0 , whereas ⟨ P y , y ⟩ = ⟨ T 3 x , x ⟩ = ∥ T 3 / 2 x ∥ 2 > 0 \langle Py,y\rangle=\langle T^3x,x\rangle=\|T^{3/2}x\|^2>0 ⟨ P y , y ⟩ = ⟨ T 3 x , x ⟩ = ∥ T 3/2 x ∥ 2 > 0 . For c s c_s c s , equality in τ ( ρ s ( P Q P ) ) ≤ ∥ ρ ∥ τ ( s ( P Q P ) ) = ∥ ρ ∥ τ ( s ( Q P Q ) ) ≤ κ τ ( ρ s ( Q P Q ) ) \tau(\rho\,s(PQP))\le\|\rho\|\tau(s(PQP))=\|\rho\|\tau(s(QPQ))\le\kappa\tau(\rho\,s(QPQ)) τ ( ρ s ( P QP )) ≤ ∥ ρ ∥ τ ( s ( P QP )) = ∥ ρ ∥ τ ( s ( QP Q )) ≤ κ τ ( ρ s ( QP Q )) forces ran s ( P Q P ) = ran P Q P ⊂ E M \operatorname{ran}s(PQP)=\operatorname{ran}PQP\subset E_M ran s ( P QP ) = ran P QP ⊂ E M and ran s ( Q P Q ) = ran Q P Q ⊂ E m \operatorname{ran}s(QPQ)=\operatorname{ran}QPQ\subset E_m ran s ( QP Q ) = ran QP Q ⊂ E m , which gives the same contradiction.
p:c By Kaplansky’s formula [136 Ch. 6] there is a partial isometry w w w with w w ∗ = P ∨ Q − Q ww^*=P\vee Q-Q w w ∗ = P ∨ Q − Q and w ∗ w = P − P ∧ Q ≤ P w^*w=P-P\wedge Q\le P w ∗ w = P − P ∧ Q ≤ P . By Theorem (?) , φ ( P ∨ Q ) = φ ( Q ) + φ ( w w ∗ ) ≤ φ ( Q ) + κ Z φ ( P ) \varphi(P\vee Q)=\varphi(Q)+\varphi(ww^*)\le\varphi(Q)+\kappa_\mathcal{Z}\varphi(P) φ ( P ∨ Q ) = φ ( Q ) + φ ( w w ∗ ) ≤ φ ( Q ) + κ Z φ ( P ) ; symmetrise and average to get c ∨ ≤ ( 1 + κ Z ) / 2 c_\vee\le(1+\kappa_\mathcal{Z})/2 c ∨ ≤ ( 1 + κ Z ) /2 . In M 2 M_2 M 2 with ρ = d i a g ( κ , 1 ) \rho=\mathop{\mathrm{diag}}(\kappa,1) ρ = diag ( κ , 1 ) , P = e 22 P=e_{22} P = e 22 and Q θ Q_\theta Q θ the projection onto ( sin θ , cos θ ) (\sin\theta,\cos\theta) ( sin θ , cos θ ) give P ∨ Q θ = 1 P\vee Q_\theta=1 P ∨ Q θ = 1 for θ ≠ 0 \theta\ne0 θ = 0 and ratio ( κ + 1 ) / ( 1 + κ sin 2 θ + cos 2 θ ) → ( 1 + κ ) / 2 (\kappa+1)/(1+\kappa\sin^2\theta+\cos^2\theta)\to(1+\kappa)/2 ( κ + 1 ) / ( 1 + κ sin 2 θ + cos 2 θ ) → ( 1 + κ ) /2 . Apply Theorem 4.15 t:low . ◻
The five projection constants c p r o j , c P Q P , c ∨ , c P Q , c s c_{\rm proj},c_{PQP},c_\vee,c_{PQ},c_s c proj , c P QP , c ∨ , c P Q , c s take four different values, G Z ≤ G Z 2 ≤ 1 + κ Z 2 ≤ κ Z G_\mathcal{Z}\le G_\mathcal{Z}^2\le\frac{1+\kappa_\mathcal{Z}}2\le\kappa_\mathcal{Z} G Z ≤ G Z 2 ≤ 2 1 + κ Z ≤ κ Z ; near the centre they behave like 1 + 1 2 L N C 2 1+\frac12\mathop{\mathrm{LNC}}^2 1 + 2 1 LNC 2 , 1 + L N C 2 1+\mathop{\mathrm{LNC}}^2 1 + LNC 2 , 1 + L N C 1+\mathop{\mathrm{LNC}} 1 + LNC and 1 + 2 L N C 1+2\mathop{\mathrm{LNC}} 1 + 2 LNC . The quadratic ones come from the Krein and Wielandt inequalities, the linear ones from comparing equivalent projections.
Proposition 5.33 (Item 22). For all projections P , Q ∈ M P,Q\in\mathcal{M} P , Q ∈ M , φ ( R e P Q ) ≤ G Z φ ( ∣ P Q ∣ ) \varphi(\mathrm{Re}\,PQ)\le G_\mathcal{Z}\,\varphi(|PQ|) φ ( Re P Q ) ≤ G Z φ ( ∣ P Q ∣ ) , and the constant is optimal. In M n ( C ) M_n(\mathbb{C}) M n ( C ) it is G G G and attained by rank-one projections, so V n = G − 1 V_n=G-1 V n = G − 1 for item 22.
Proof. φ ( R e P Q ) = R e φ ( P Q ) ≤ ∣ φ ( P Q ) ∣ ≤ G Z φ ( ∣ P Q ∣ ) \varphi(\mathrm{Re}\,PQ)=\mathop{\mathrm{Re}}\varphi(PQ)\le|\varphi(PQ)|\le G_\mathcal{Z}\varphi(|PQ|) φ ( Re P Q ) = Re φ ( P Q ) ≤ ∣ φ ( P Q ) ∣ ≤ G Z φ ( ∣ P Q ∣ ) by Theorem 5.1 . For the lower bound let ρ = d i a g ( a , b ) \rho=\mathop{\mathrm{diag}}(a,b) ρ = diag ( a , b ) in M 2 M_2 M 2 , v = b / ( a + b ) e 1 + a / ( a + b ) e 2 v=\sqrt{b/(a+b)}\,e_1+\sqrt{a/(a+b)}\,e_2 v = b / ( a + b ) e 1 + a / ( a + b ) e 2 , u = ρ v / ∥ ρ v ∥ u=\rho v/\|\rho v\| u = ρ v /∥ ρ v ∥ , P = u u ∗ P=uu^* P = u u ∗ and Q = v v ∗ Q=vv^* Q = v v ∗ . Then P Q = ( u ∗ v ) u v ∗ PQ=(u^*v)\,uv^* P Q = ( u ∗ v ) u v ∗ with u ∗ v = ⟨ ρ v , v ⟩ / ∥ ρ v ∥ > 0 u^*v=\langle\rho v,v\rangle/\|\rho v\|>0 u ∗ v = ⟨ ρ v , v ⟩ /∥ ρ v ∥ > 0 , so ∣ P Q ∣ = ( u ∗ v ) Q |PQ|=(u^*v)Q ∣ P Q ∣ = ( u ∗ v ) Q , φ ( P Q ) = ( u ∗ v ) ∥ ρ v ∥ \varphi(PQ)=(u^*v)\|\rho v\| φ ( P Q ) = ( u ∗ v ) ∥ ρ v ∥ and φ ( ∣ P Q ∣ ) = ( u ∗ v ) ⟨ ρ v , v ⟩ \varphi(|PQ|)=(u^*v)\langle\rho v,v\rangle φ ( ∣ P Q ∣ ) = ( u ∗ v ) ⟨ ρ v , v ⟩ . The ratio is ∥ ρ v ∥ / ⟨ ρ v , v ⟩ = a b ( a + b ) / ( 2 a b ) = G ( ρ ) \|\rho v\|/\langle\rho v,v\rangle=\sqrt{ab}\,(a+b)/(2ab)=G(\rho) ∥ ρ v ∥/ ⟨ ρ v , v ⟩ = ab ( a + b ) / ( 2 ab ) = G ( ρ ) , the equality case of Lemma 3.1 . The constant is a violation constant (two projections, F 1 = 1 2 ( X 1 X 2 + X 2 X 1 ) F_1=\frac12(X_1X_2+X_2X_1) F 1 = 2 1 ( X 1 X 2 + X 2 X 1 ) , F 2 = ∣ X 1 X 2 ∣ F_2=|X_1X_2| F 2 = ∣ X 1 X 2 ∣ ), and Theorem 4.15 t:low and Remark 4.16 complete the proof. ◻
The violation constants of items 17, 18, 22 and 23 are functions of κ Z \kappa_\mathcal{Z} κ Z ; on M n M_n M n they do not depend on the dimension. For items 19 and 20, the remaining projection inequalities of [96 ] and [97 ] the normalisation contains φ ( 1 ) \varphi(1) φ ( 1 ) , and we determine the envelopes on M n M_n M n .
Theorem 5.34 (Items 19, 20 and 23 on M n M_n M n ). Let ρ > 0 \rho>0 ρ > 0 in M n M_n M n , n ≥ 2 n\ge2 n ≥ 2 , with O M C ( ρ ) = δ \mathop{\mathrm{OMC}}(\rho)=\delta OMC ( ρ ) = δ , and let P , Q P,Q P , Q be projections. With N \mathcal{N} N as in Table 3 :
V n = δ V_n=\delta V n = δ for φ ( P + Q ) ≤ φ ( 1 + P ∧ Q ) \varphi(P+Q)\le\varphi(1+P\wedge Q) φ ( P + Q ) ≤ φ ( 1 + P ∧ Q ) if n n n is even, and V n = ( n − 1 ) δ / ( n − δ ) V_n=(n-1)\delta/(n-\delta) V n = ( n − 1 ) δ / ( n − δ ) if n n n is odd;
for φ ( e Q P Q ) ≤ φ ( e P ) \varphi(e^{QPQ})\le\varphi(e^P) φ ( e QP Q ) ≤ φ ( e P ) and even n n n , V n = max 0 < g < π λ ( δ , g ) V_n=\max_{0<g<\pi}\lambda(\delta,g) V n = max 0 < g < π λ ( δ , g ) , where λ ( δ , g ) \lambda(\delta,g) λ ( δ , g ) is the larger root of the quadratic (9) below, and V n = ( e − 1 ) 2 e ( e + 1 ) δ 2 + O ( δ 3 ) V_n=\frac{(e-1)^2}{e(e+1)}\delta^2+O(\delta^3) V n = e ( e + 1 ) ( e − 1 ) 2 δ 2 + O ( δ 3 ) ;
the value V n = δ / ( 1 − δ ) V_n=\delta/(1-\delta) V n = δ / ( 1 − δ ) of φ ( P ∨ Q ) ≤ φ ( P ) + φ ( Q ) \varphi(P\vee Q)\le\varphi(P)+\varphi(Q) φ ( P ∨ Q ) ≤ φ ( P ) + φ ( Q ) (Theorem 5.32 p:c ) is approached but not attained.
The value is attained in 20 by rank-one P , Q P,Q P , Q , and approached in 19 and 23 as P ≠ Q P\ne Q P = Q tend to the same eigenvector of ρ \rho ρ . For odd n n n , V n V_n V n in item 20 lies between the M 2 M_2 M 2 value and the same maximum with ( 1 − δ ) / ( n − 1 ) (1-\delta)/(n-1) ( 1 − δ ) / ( n − 1 ) added to β \beta β in (9) .
Proof. Step 1: reduction to rank-one projections on M 2 M_2 M 2 . By Lemma 3.4 , C n = ⨁ j H j \mathbb{C}^n=\bigoplus_jH_j C n = ⨁ j H j with dim H j ∈ { 1 , 2 } \dim H_j\in\{1,2\} dim H j ∈ { 1 , 2 } , every H j H_j H j invariant under P P P and Q Q Q . Every operator in L \mathcal{L} L , R \mathcal{R} R and N \mathcal{N} N (namely P + Q P+Q P + Q , 1 + P ∧ Q 1+P\wedge Q 1 + P ∧ Q , e Q P Q e^{QPQ} e QP Q , e P e^P e P , P ∨ Q P\vee Q P ∨ Q ) is then block diagonal, with the block obtained from P j , Q j P_j,Q_j P j , Q j ; for the lattice operations because r a n P ∩ r a n Q = ⨁ j ( r a n P j ∩ r a n Q j ) \mathrm{ran}\,P\cap\mathrm{ran}\,Q=\bigoplus_j(\mathrm{ran}\,
P_j\cap\mathrm{ran}\,Q_j) ran P ∩ ran Q = ⨁ j ( ran P j ∩ ran Q j ) , the projections onto the H j H_j H j commuting with P P P and Q Q Q . Hence L = ∑ j L j \mathcal{L}=\sum_jL_j L = ∑ j L j , R = ∑ j R j \mathcal{R}=\sum_jR_j R = ∑ j R j and N = ∑ j N j \mathcal{N}=\sum_jN_j N = ∑ j N j , where L j \mathcal{L}_j L j is the same expression for P j , Q j P_j,Q_j P j , Q j and the functional T r ( ρ j ⋅ ) \mathop{\mathrm{Tr}}(\rho_j\,\cdot\,) Tr ( ρ j ⋅ ) on H j H_j H j , with the compression ρ j \rho_j ρ j of ρ \rho ρ to H j H_j H j ; its spectrum lies in [ b , a ] [b,a] [ b , a ] , so O M C ( ρ j ) ≤ δ \mathop{\mathrm{OMC}}(\rho_j)\le\delta OMC ( ρ j ) ≤ δ . On a one-dimensional H j H_j H j , P j , Q j ∈ { 0 , 1 } P_j,Q_j\in\{0,1\} P j , Q j ∈ { 0 , 1 } commute and L j ≤ R j \mathcal{L}_j\le \mathcal{R}_j L j ≤ R j in all three items (p + q ≤ 1 + p q p+q\le1+pq p + q ≤ 1 + pq , e p q ≤ e p e^{pq}\le e^p e pq ≤ e p , max ( p , q ) ≤ p + q \max(p,q)\le p+q max ( p , q ) ≤ p + q ). On a two-dimensional H j H_j H j , P j ≠ Q j P_j\ne Q_j P j = Q j are rank-one projections. Since N j ≥ 0 \mathcal{N}_j\ge0 N j ≥ 0 , and L j ≤ R j \mathcal{L}_j\le \mathcal{R}_j L j ≤ R j wherever N j = 0 \mathcal{N}_j=0 N j = 0 ,
\frac{\mathcal{L}-\mathcal{R}}\mathcal{N}\le\max_j\frac{\mathcal{L}_j-\mathcal{R}_j}{\mathcal{N}_j},
and the quotients do not change when ρ j \rho_j ρ j is scaled. It therefore suffices to bound ( L − R ) / N (\mathcal{L}-\mathcal{R})/\mathcal{N} ( L − R ) / N for ρ = 1 2 ( 1 + t σ 3 ) \rho=\frac12(1+t\sigma_3) ρ = 2 1 ( 1 + t σ 3 ) , 0 ≤ t ≤ δ 0\le t\le\delta 0 ≤ t ≤ δ , and P = 1 2 ( 1 + μ ⋅ σ ) P=\frac12(1+\mu\cdot\sigma) P = 2 1 ( 1 + μ ⋅ σ ) , Q = 1 2 ( 1 + ν ⋅ σ ) Q=\frac12(1+\nu\cdot\sigma) Q = 2 1 ( 1 + ν ⋅ σ ) with unit Bloch vectors μ ≠ ν \mu\ne\nu μ = ν in R 3 \mathbb{R}^3 R 3 (σ = ( σ 1 , σ 2 , σ 3 ) \sigma=(\sigma_1,\sigma_2,\sigma_3) σ = ( σ 1 , σ 2 , σ 3 ) the Pauli matrices). For fixed μ , ν \mu,\nu μ , ν the quotient is a linear-fractional function of t t t (all three quantities are affine in ρ \rho ρ ) with positive denominator, hence monotone on [ − δ , δ ] [-\delta,\delta] [ − δ , δ ] ; conjugation by σ 1 \sigma_1 σ 1 maps t t t to − t -t − t and ( μ , ν ) (\mu,\nu) ( μ , ν ) to reflected vectors. So the bound for t = δ t=\delta t = δ covers every t ≤ δ t\le\delta t ≤ δ , and the M 2 M_2 M 2 envelope bounds the envelope in every dimension. In item 23, N \mathcal{N} N vanishes on blocks where P = Q = 0 P=Q=0 P = Q = 0 , so the strict bound below holds on M n M_n M n . In 19 and 20, N \mathcal{N} N contains φ ( 1 ) \varphi(1) φ ( 1 ) , which we treat at the end.
Step 2: formulas on M 2 M_2 M 2 . Write φ ( x 0 + x ⋅ σ ) = x 0 + δ x 3 \varphi(x_0+x\cdot\sigma)=x_0+\delta x_3 φ ( x 0 + x ⋅ σ ) = x 0 + δ x 3 , μ ⋅ ν = cos 2 θ \mu\cdot\nu=\cos2\theta μ ⋅ ν = cos 2 θ (0 < θ ≤ π / 2 0<\theta\le\pi/2 0 < θ ≤ π /2 ) and c = cos θ = ∣ ⟨ u , v ⟩ ∣ c=\cos\theta=|\langle u,v\rangle| c = cos θ = ∣ ⟨ u , v ⟩ ∣ for unit vectors u , v u,v u , v with P = u u ∗ P=uu^* P = u u ∗ , Q = v v ∗ Q=vv^* Q = v v ∗ . Then P ∧ Q = 0 P\wedge Q=0 P ∧ Q = 0 , P ∨ Q = 1 P\vee Q=1 P ∨ Q = 1 , e Q P Q = 1 + ( e c 2 − 1 ) Q e^{QPQ}=1+(e^{c^2}-1)Q e QP Q = 1 + ( e c 2 − 1 ) Q and e P = 1 + ( e − 1 ) P e^P=1+(e-1)P e P = 1 + ( e − 1 ) P . We use the following elementary fact: for β > ∣ z ∣ \beta>|z| β > ∣ z ∣ , max ∣ ξ ∣ = 1 ( α + y ⋅ ξ ) / ( β + z ⋅ ξ ) \max_{|\xi|=1}(\alpha+y\cdot\xi)/(\beta+z\cdot\xi) max ∣ ξ ∣ = 1 ( α + y ⋅ ξ ) / ( β + z ⋅ ξ ) is the unique zero λ \lambda λ of α − λ β + ∣ y − λ z ∣ \alpha-\lambda\beta+|y-\lambda z| α − λ β + ∣ y − λ z ∣ , since ( α + y ⋅ ξ ) − λ ( β + z ⋅ ξ ) ≤ α − λ β + ∣ y − λ z ∣ (\alpha+y\cdot\xi)-\lambda(\beta+z\cdot\xi)\le\alpha-\lambda\beta+|y-\lambda z| ( α + y ⋅ ξ ) − λ ( β + z ⋅ ξ ) ≤ α − λ β + ∣ y − λ z ∣ with equality for ξ ∥ y − λ z \xi\parallel y-\lambda z ξ ∥ y − λ z , and the right-hand side decreases strictly in λ \lambda λ .
(19) ( L − R ) / N = φ ( P + Q ) − 1 = δ 2 ( μ 3 + ν 3 ) < δ (\mathcal{L}-\mathcal{R})/\mathcal{N}=\varphi(P+Q)-1=\frac\delta2(\mu_3+\nu_3)<\delta ( L − R ) / N = φ ( P + Q ) − 1 = 2 δ ( μ 3 + ν 3 ) < δ , approached as μ , ν → e 3 \mu,\nu\to e_3 μ , ν → e 3 .
(20) For fixed g = 2 θ g=2\theta g = 2 θ , the pair ( μ 3 , ν 3 ) (\mu_3,\nu_3) ( μ 3 , ν 3 ) ranges over the image of the unit sphere under ξ ↦ ( μ ⋅ ξ , ν ⋅ ξ ) \xi\mapsto(\mu\cdot
\xi,\nu\cdot\xi) ξ ↦ ( μ ⋅ ξ , ν ⋅ ξ ) , a filled ellipse, and the linear-fractional quotient is maximal on its boundary, μ 3 = cos x \mu_3=\cos x μ 3 = cos x , ν 3 = cos ( x + g ) \nu_3=\cos(x+g) ν 3 = cos ( x + g ) . With K = e − 1 K=e-1 K = e − 1 and E = e cos 2 ( g / 2 ) − 1 E=e^{\cos^2(g/2)}-1 E = e c o s 2 ( g /2 ) − 1 ,
\frac{\mathcal{L}-\mathcal{R}}\mathcal{N}=\frac{\alpha+y\cdot\xi_x}{\beta+z\cdot\xi_x},\quad \alpha=\tfrac{E-K}2,\ y=\tfrac\delta2(E\cos g-K,\,-E\sin g),
\ \beta=1+\tfrac K2,\ z=(\tfrac{\delta K}2,0),
ξ x = ( cos x , sin x ) \xi_x=(\cos x,\sin x) ξ x = ( cos x , sin x ) . By the fact, its maximum λ ( δ , g ) \lambda(\delta,g) λ ( δ , g ) is the larger root of
( β 2 − ∣ z ∣ 2 ) λ 2 − 2 ( α β − y ⋅ z ) λ + α 2 − ∣ y ∣ 2 = 0. \begin{equation}
(\beta^2-|z|^2)\lambda^2-2(\alpha\beta-y\cdot z)\lambda+\alpha^2-|y|^2=0 . \tag*{(9)}
\end{equation} ( β 2 − ∣ z ∣ 2 ) λ 2 − 2 ( α β − y ⋅ z ) λ + α 2 − ∣ y ∣ 2 = 0. (9) For small δ \delta δ the maximiser has g = 2 K δ / e + O ( δ 2 ) g=2K\delta/e+O(\delta^2) g = 2 K δ / e + O ( δ 2 ) , x = π / 2 x=\pi/2 x = π /2 , and L − R = − e 8 g 2 + K δ 2 g + O ( δ 3 ) \mathcal{L}-\mathcal{R}=-\frac e8g^2+\frac{K\delta}2g+O(\delta^3) L − R = − 8 e g 2 + 2 K δ g + O ( δ 3 ) , which gives V = K 2 δ 2 / ( e ( e + 1 ) ) + O ( δ 3 ) V=K^2\delta^2/(e(e+1))+O(\delta^3) V = K 2 δ 2 / ( e ( e + 1 )) + O ( δ 3 ) .
(23) L / R = 1 / ( 1 + δ 2 ( μ 3 + ν 3 ) ) < 1 / ( 1 − δ ) \mathcal{L}/\mathcal{R}=1/(1+\frac\delta2(\mu_3+\nu_3))<1/(1-\delta) L / R = 1/ ( 1 + 2 δ ( μ 3 + ν 3 )) < 1/ ( 1 − δ ) , approached as μ , ν → − e 3 \mu,\nu\to-e_3 μ , ν → − e 3 .
Step 3: items 19 and 20 in dimension n n n . For even n n n take ρ \rho ρ with the eigenvalues a a a and b b b , each n / 2 n/2 n /2 times, and n / 2 n/2 n /2 copies of the M 2 M_2 M 2 extremiser, one in each block spanned by an eigenvector for a a a and one for b b b ; this attains (approaches) the M 2 M_2 M 2 value, which is an upper bound by Step 1. For odd n n n at least one block of Lemma 3.4 is one-dimensional, so there are k ′ ≤ k = ( n − 1 ) / 2 k'\le k=(n-1)/2 k ′ ≤ k = ( n − 1 ) /2 two-dimensional blocks. In item 19, N ≥ φ ( 1 ) = T r ρ \mathcal{N}\ge\varphi(1)=\mathop{\mathrm{Tr}}\rho N ≥ φ ( 1 ) = Tr ρ ; by Step 2 a two-dimensional block contributes L j − R j = T r ρ j ⋅ O M C ( ρ j ) 2 ( μ 3 + ν 3 ) < λ 1 ( ρ j ) − λ 2 ( ρ j ) \mathcal{L}_j-\mathcal{R}_j=\mathop{\mathrm{Tr}}\rho_j\cdot\frac{\mathop{\mathrm{OMC}}(\rho_j)}2(\mu_3+\nu_3)
<\lambda_1(\rho_j)-\lambda_2(\rho_j) L j − R j = Tr ρ j ⋅ 2 OMC ( ρ j ) ( μ 3 + ν 3 ) < λ 1 ( ρ j ) − λ 2 ( ρ j ) , and a one-dimensional one L j − R j ≤ 0 \mathcal{L}_j-\mathcal{R}_j\le0 L j − R j ≤ 0 . By Ky Fan’s maximum principle [93 Sect. III.1] , ∑ j λ 1 ( ρ j ) \sum_j\lambda_1(\rho_j) ∑ j λ 1 ( ρ j ) is at most the sum of the k ′ k' k ′ largest and ∑ j λ 2 ( ρ j ) \sum_j\lambda_2(\rho_j) ∑ j λ 2 ( ρ j ) at least the sum of the k ′ k' k ′ smallest eigenvalues of ρ \rho ρ . The resulting bound for ( L − R ) / N (\mathcal{L}-\mathcal{R})/\mathcal{N} ( L − R ) / N is a linear-fractional function of the spectrum of ρ \rho ρ and is maximal for k ′ = k k'=k k ′ = k and the spectrum a ( k ) , b ( k + 1 ) a^{(k)},b^{(k+1)} a ( k ) , b ( k + 1 ) ; hence ( L − R ) / N < k ( a − b ) / ( k ( a + b ) + b ) = ( n − 1 ) δ / ( n − δ ) (\mathcal{L}-\mathcal{R})/\mathcal{N}<k(a-b)/(k(a+b)+b)=(n-1)\delta/(n-\delta) ( L − R ) / N < k ( a − b ) / ( k ( a + b ) + b ) = ( n − 1 ) δ / ( n − δ ) , which is approached for this spectrum with P = 1 P=1 P = 1 , Q = 0 Q=0 Q = 0 on the remaining eigenvector (a block with L j = R j \mathcal{L}_j=\mathcal{R}_j L j = R j ). In item 20 the same configuration, with P = Q = 0 P=Q=0 P = Q = 0 on the remaining eigenvector, adds b / k b/k b / k per block to N \mathcal{N} N , i.e. ( 1 − δ ) / ( n − 1 ) (1-\delta)/(n-1) ( 1 − δ ) / ( n − 1 ) to β \beta β (normalisation T r ρ j = 1 \mathop{\mathrm{Tr}}\rho_j=1 Tr ρ j = 1 ); this gives the lower bound, and the M 2 M_2 M 2 value the upper bound. ◻
Proposition 5.35 (Item 14). For every n ≥ 2 n\ge2 n ≥ 2 , V n ≥ δ 2 / 36 + O ( δ 3 ) V_n\ge\delta^2/36+O(\delta^3) V n ≥ δ 2 /36 + O ( δ 3 ) for item 14, and more generally V n ≥ ( 1 − 2 p ) 2 δ 2 / 4 + O ( δ 3 ) V_n\ge(1-\frac2p)^2\delta^2/4+O(\delta^3) V n ≥ ( 1 − p 2 ) 2 δ 2 /4 + O ( δ 3 ) for Hölder’s inequality ∣ φ ( P Q ) ∣ ≤ φ ( P ) 1 / p φ ( Q ) 1 / q |\varphi(PQ)|\le\varphi(P)^{1/p}\varphi(Q)^{1/q} ∣ φ ( P Q ) ∣ ≤ φ ( P ) 1/ p φ ( Q ) 1/ q .
Proof. Take ρ = 1 2 ( 1 + δ D ) \rho=\frac12(1+\delta D) ρ = 2 1 ( 1 + δ D ) , D = d i a g ( 1 , − 1 ) D=\mathrm{diag}(1,-1) D = diag ( 1 , − 1 ) , and real unit vectors u , v u,v u , v at the angles ϑ ± ε / 2 \vartheta\pm\varepsilon/2 ϑ ± ε /2 to e 1 e_1 e 1 , P = u u ∗ P=uu^* P = u u ∗ , Q = v v ∗ Q=vv^* Q = v v ∗ . Then φ ( P Q ) = 1 2 ( cos 2 ε + δ cos ε cos 2 ϑ ) \varphi(PQ)=\frac12(\cos^2\varepsilon+\delta\cos\varepsilon\cos2\vartheta) φ ( P Q ) = 2 1 ( cos 2 ε + δ cos ε cos 2 ϑ ) and φ ( P ) 1 / p φ ( Q ) 1 / q = 1 2 ( 1 + δ cos 2 ϑ + δ ( 1 q − 1 p ) ε sin 2 ϑ ) + O ( δ 2 ε 2 ) \varphi(P)^{1/p}\varphi(Q)^{1/q}=\frac12\bigl(1+\delta\cos2\vartheta+\delta(\frac1q-\frac1p)\varepsilon\sin2\vartheta\bigr)+O(\delta^2\varepsilon^2) φ ( P ) 1/ p φ ( Q ) 1/ q = 2 1 ( 1 + δ cos 2 ϑ + δ ( q 1 − p 1 ) ε sin 2 ϑ ) + O ( δ 2 ε 2 ) , so L − R = 1 2 ( − ε 2 − δ ( 1 − 2 p ) ε sin 2 ϑ ) + O ( δ 3 ) \mathcal{L}-\mathcal{R}=\frac12\bigl(-\varepsilon^2-\delta(1-\frac2p)\varepsilon\sin2\vartheta\bigr)+O(\delta^3) L − R = 2 1 ( − ε 2 − δ ( 1 − p 2 ) ε sin 2 ϑ ) + O ( δ 3 ) for ε = O ( δ ) \varepsilon=O(\delta) ε = O ( δ ) . The choice ϑ = π / 4 \vartheta=\pi/4 ϑ = π /4 , ε = − δ 2 ( 1 − 2 p ) \varepsilon=-\frac\delta2(1-\frac2p) ε = − 2 δ ( 1 − p 2 ) gives L − R = 1 8 ( 1 − 2 p ) 2 δ 2 + O ( δ 3 ) \mathcal{L}-\mathcal{R}=\frac18(1-\frac2p)^2\delta^2+O(\delta^3) L − R = 8 1 ( 1 − p 2 ) 2 δ 2 + O ( δ 3 ) and R = 1 2 + O ( δ ) \mathcal{R}=\frac12+O(\delta) R = 2 1 + O ( δ ) . ◻
The canonical trace
Items 11 and 24 characterise the canonical trace among the positive functionals with φ ( 1 ) = n \varphi(1)=n φ ( 1 ) = n on M n M_n M n ; they are not violation constants in the sense of Definition 4.11 , and we determine them on M n M_n M n .
Proposition 5.36 (Item 11). Let T r ρ = n \mathop{\mathrm{Tr}}\rho=n Tr ρ = n . For item 11, V ( ρ ) = 1 − G M / A M V(\rho)=1-\mathrm{GM}/\mathrm{AM} V ( ρ ) = 1 − GM / AM of the eigenvalues of ρ \rho ρ , attained at A = ρ − 1 A=\rho^{-1} A = ρ − 1 . Over the densities with contrast δ \delta δ the maximum is attained at a two-level spectrum: V n ( δ ) = 1 − min 1 ≤ m < n n ℓ ( n − m ) / n m + ( n − m ) ℓ , ℓ = 1 κ = 1 − δ 1 + δ . V_n(\delta)=1-\min_{1\le m<n}\frac{n\,\ell^{(n-m)/n}}{m+(n-m)\ell},\qquad\ell=\frac1\kappa=\frac{1-\delta}{1+\delta}. V n ( δ ) = 1 − min 1 ≤ m < n m + ( n − m ) ℓ n ℓ ( n − m ) / n , ℓ = κ 1 = 1 + δ 1 − δ . In particular V 2 = 1 − 1 − δ 2 V_2=1-\sqrt{1-\delta^2} V 2 = 1 − 1 − δ 2 , and V n ↑ 1 − 1 / S ( κ ) V_n\uparrow1-1/S(\kappa) V n ↑ 1 − 1/ S ( κ ) as n → ∞ n\to\infty n → ∞ , where S ( κ ) = ( κ − 1 ) κ 1 / ( κ − 1 ) e log κ S(\kappa)=\frac{(\kappa-1)\kappa^{1/(\kappa-1)}}{e\log \kappa} S ( κ ) = e l o g κ ( κ − 1 ) κ 1/ ( κ − 1 ) is the Specht ratio.
Proof. By the AM–GM inequality for the eigenvalues of ρ 1 / 2 A ρ 1 / 2 \rho^{1/2}A\rho^{1/2} ρ 1/2 A ρ 1/2 , φ ρ ( A ) / n ≥ ( det ρ det A ) 1 / n \varphi_\rho(A)/n\ge(\det \rho\det A)^{1/n} φ ρ ( A ) / n ≥ ( det ρ det A ) 1/ n , with equality for A = ρ − 1 A=\rho^{-1} A = ρ − 1 , so inf A > 0 φ ρ ( A ) / ( n ( det A ) 1 / n ) = ( det ρ ) 1 / n \inf_{A>0}\varphi_\rho(A)/(n(\det A)^{1/n})=(\det \rho)^{1/n} inf A > 0 φ ρ ( A ) / ( n ( det A ) 1/ n ) = ( det ρ ) 1/ n and V ( ρ ) = 1 − ( det ρ ) 1 / n V(\rho)=1-(\det \rho)^{1/n} V ( ρ ) = 1 − ( det ρ ) 1/ n , which for T r ρ = n \mathop{\mathrm{Tr}}\rho=n Tr ρ = n is 1 − G M / A M 1-\mathrm{GM}/\mathrm{AM} 1 − GM / AM . In the coordinates x i = log w i x_i=\log w_i x i = log w i , where w 1 , … , w n w_1,\dots,w_n w 1 , … , w n are the eigenvalues of ρ \rho ρ , A M / G M = 1 n ∑ i e x i − x ˉ \mathrm{AM}/\mathrm{GM}=\frac1n\sum_i
e^{x_i-\bar x} AM / GM = n 1 ∑ i e x i − x ˉ is convex, so on the box of spectra in [ b , a ] [b,a] [ b , a ] it is maximal at a vertex, a two-level spectrum with m m m eigenvalues a a a and n − m n-m n − m eigenvalues b b b ; this gives the formula. Its limit is Specht’s theorem [158 ] : sup A M / G M \sup\mathrm{AM}/\mathrm{GM} sup AM / GM over all finite families in [ b , a ] [b,a] [ b , a ] is S ( κ ) S(\kappa) S ( κ ) . ◻
Proposition 5.37 (Item 24). Let T r ρ = n \mathop{\mathrm{Tr}}\rho=n Tr ρ = n and let w 1 , … , w n w_1,\dots,w_n w 1 , … , w n be the eigenvalues of ρ \rho ρ , w min = min i w i w_{\min}=\min_iw_i w m i n = min i w i . For item 24, V ( ρ ) = w min − 1 / 2 − 1 V(\rho)=w_{\min}^{-1/2}-1 V ( ρ ) = w m i n − 1/2 − 1 , and V n ( δ ) = ℓ + n − 1 n ℓ − 1 , ℓ = 1 − δ 1 + δ ; V_n(\delta)=\sqrt{\frac{\ell+n-1}{n\ell}}-1,\qquad\ell=\frac{1-\delta}{1+\delta}; V n ( δ ) = n ℓ ℓ + n − 1 − 1 , ℓ = 1 + δ 1 − δ ; V 2 = ( 1 − δ ) − 1 / 2 − 1 ∼ δ / 2 V_2=(1-\delta)^{-1/2}-1\sim\delta/2 V 2 = ( 1 − δ ) − 1/2 − 1 ∼ δ /2 , V n ∼ n − 1 n δ V_n\sim\frac{n-1}n\delta V n ∼ n n − 1 δ , and V n ↑ κ − 1 V_n\uparrow\sqrt \kappa-1 V n ↑ κ − 1 as n → ∞ n\to\infty n → ∞ .
Proof. For A ≥ 0 A\ge0 A ≥ 0 , φ ( A 2 ) = T r ( A 1 / 2 ρ A 1 / 2 A ) ≤ ∥ A ∥ φ ( A ) ≤ T r ( A ) φ ( A ) ≤ φ ( A ) 2 / w min \varphi(A^2)=\mathop{\mathrm{Tr}}(A^{1/2}\rho A^{1/2}A)\le\|A\|\varphi(A)\le\mathop{\mathrm{Tr}}(A)\varphi(A)\le\varphi(A)^2/w_{\min} φ ( A 2 ) = Tr ( A 1/2 ρ A 1/2 A ) ≤ ∥ A ∥ φ ( A ) ≤ Tr ( A ) φ ( A ) ≤ φ ( A ) 2 / w m i n , with equality for the projection onto an eigenvector for w min w_{\min} w m i n . With T r ρ = n \mathop{\mathrm{Tr}}\rho=n Tr ρ = n and the extreme eigenvalues in the ratio κ \kappa κ , w min = n b / ∑ i w i ≥ n b / ( b + ( n − 1 ) a ) = n ℓ / ( ℓ + n − 1 ) w_{\min}=nb/\sum_iw_i\ge nb/(b+(n-1)a)=n\ell/(\ell+n-1) w m i n = nb / ∑ i w i ≥ nb / ( b + ( n − 1 ) a ) = n ℓ / ( ℓ + n − 1 ) , with equality for the spectrum { a , … , a , b } \{a,\dots,a,b\} { a , … , a , b } . For n = 2 n=2 n = 2 , ( ℓ + 1 ) / ( 2 ℓ ) = 1 / ( 1 − δ ) (\ell+1)/(2\ell)=1/(1-\delta) ( ℓ + 1 ) / ( 2 ℓ ) = 1/ ( 1 − δ ) . ◻
Subadditivity of powers
For 0 < p < 1 0<p<1 0 < p < 1 let c p T ( ρ ) = sup A , B ∈ M + φ ( ( A + B ) p ) / ( φ ( A p ) + φ ( B p ) ) c_p^T(\rho)=\sup_{A,B\in\mathcal{M}_+}\varphi((A+B)^p)/(\varphi(A^p)+\varphi(B^p)) c p T ( ρ ) = sup A , B ∈ M + φ (( A + B ) p ) / ( φ ( A p ) + φ ( B p )) , and
t p ( κ ) = max 0 ≤ x ≤ 1 2 p − 1 κ x p + ( 1 − x ) p κ x + 1 − x . \mathsf t_p(\kappa)=\max_{0\le x\le1}2^{p-1}\frac{\kappa x^p+(1-x)^p}{\kappa x+1-x}. t p ( κ ) = 0 ≤ x ≤ 1 max 2 p − 1 κ x + 1 − x κ x p + ( 1 − x ) p .
Lemma 5.39. Let χ \chi χ be a normal positive tracial functional on M \mathcal{M} M , A , B ∈ M + A,B\in\mathcal{M}_+ A , B ∈ M + and 0 < p < 1 0<p<1 0 < p < 1 . Then χ ( ( A + B ) p ) ≤ χ ( A p ) + χ ( B p ) \chi((A+B)^p)\le\chi(A^p)+\chi(B^p) χ (( A + B ) p ) ≤ χ ( A p ) + χ ( B p ) .
Proof. For s > 0 s>0 s > 0 let h s ( t ) = t / ( t + s ) h_s(t)=t/(t+s) h s ( t ) = t / ( t + s ) and C = A + B C=A+B C = A + B . Then h s ( C ) = ( C + s ) − 1 / 2 ( A + B ) ( C + s ) − 1 / 2 h_s(C)=(C+s)^{-1/2}(A+B)(C+s)^{-1/2} h s ( C ) = ( C + s ) − 1/2 ( A + B ) ( C + s ) − 1/2 and, by the trace property and ( C + s ) − 1 ≤ ( A + s ) − 1 (C+s)^{-1}\le(A+s)^{-1} ( C + s ) − 1 ≤ ( A + s ) − 1 , χ ( ( C + s ) − 1 / 2 A ( C + s ) − 1 / 2 ) = χ ( A 1 / 2 ( C + s ) − 1 A 1 / 2 ) ≤ χ ( A 1 / 2 ( A + s ) − 1 A 1 / 2 ) = χ ( h s ( A ) ) \chi((C+s)^{-1/2}A(C+s)^{-1/2})=\chi(A^{1/2}(C+s)^{-1}A^{1/2})\le\chi(A^{1/2}(A+s)^{-1}A^{1/2})=\chi(h_s(A)) χ (( C + s ) − 1/2 A ( C + s ) − 1/2 ) = χ ( A 1/2 ( C + s ) − 1 A 1/2 ) ≤ χ ( A 1/2 ( A + s ) − 1 A 1/2 ) = χ ( h s ( A )) ; similarly for B B B . Hence χ ( h s ( A + B ) ) ≤ χ ( h s ( A ) ) + χ ( h s ( B ) ) \chi(h_s(A+B))\le\chi(h_s(A))+\chi(h_s(B)) χ ( h s ( A + B )) ≤ χ ( h s ( A )) + χ ( h s ( B )) . Integrate against sin p π π s p − 1 d s \frac{\sin p\pi}\pi s^{p-1}ds π s i n p π s p − 1 d s over ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) : for X ∈ M + X\in\mathcal{M}_+ X ∈ M + , X p = sin p π π ∫ 0 ∞ s p − 1 h s ( X ) d s X^p=\frac{\sin p\pi}\pi\int_0^\infty s^{p-1}h_s(X)\,ds X p = π s i n p π ∫ 0 ∞ s p − 1 h s ( X ) d s as a norm convergent integral. ◻
For matrices this is Rotfel’d’s trace inequality [154 ] ; the short proof above is included for completeness.
Theorem 5.40 (Items 12 and 34).
c p T ( z ρ ) = c p T ( ρ ) c_p^T(z\rho)=c_p^T(\rho) c p T ( z ρ ) = c p T ( ρ ) for z ∈ Z + + z\in\mathcal{Z}_{++} z ∈ Z ++ .
φ ( ( A + B ) p ) ≤ min { φ ( B p ) + κ Z φ ( A p ) , φ ( A p ) + κ Z φ ( B p ) } \varphi((A+B)^p)\le\min\{\varphi(B^p)+\kappa_\mathcal{Z}\varphi(A^p),\varphi(A^p)+\kappa_\mathcal{Z}\varphi(B^p)\} φ (( A + B ) p ) ≤ min { φ ( B p ) + κ Z φ ( A p ) , φ ( A p ) + κ Z φ ( B p )} for A , B ∈ M + A,B\in\mathcal{M}_+ A , B ∈ M + .
t p ( κ Z ) ≤ c p T ( ρ ) ≤ 1 + κ Z 2 \mathsf t_p(\kappa_\mathcal{Z})\le c_p^T(\rho)\le\frac{1+\kappa_\mathcal{Z}}2 t p ( κ Z ) ≤ c p T ( ρ ) ≤ 2 1 + κ Z , and t p ( κ ) > 1 \mathsf t_p(\kappa)>1 t p ( κ ) > 1 for κ > 1 \kappa>1 κ > 1 .
Proof. (a) Use ( A , B ) ↦ ( z 1 / p A , z 1 / p B ) (A,B)\mapsto(z^{1/p}A,z^{1/p}B) ( A , B ) ↦ ( z 1/ p A , z 1/ p B ) . (b) Y = ( A + B ) p − B p ≥ 0 Y=(A+B)^p-B^p\ge0 Y = ( A + B ) p − B p ≥ 0 by the Löwner–Heinz inequality and Z = A p Z=A^p Z = A p . Lemma 5.39 applied to χ = τ ( 1 Q ⋅ ) \chi=\tau(\mathbf 1_Q\,\cdot) χ = τ ( 1 Q ⋅ ) for central projections Q Q Q gives τ ( 1 Q Y ) ≤ τ ( 1 Q Z ) \tau(\mathbf 1_QY)\le\tau(\mathbf 1_QZ) τ ( 1 Q Y ) ≤ τ ( 1 Q Z ) , hence τ ( u Y ) ≤ τ ( u Z ) \tau(uY)\le\tau(uZ) τ ( u Y ) ≤ τ ( u Z ) , and by (3) φ ( Y ) ≤ τ ( u Y ) ≤ τ ( u Z ) ≤ κ Z τ ( l Z ) ≤ κ Z φ ( Z ) \varphi(Y)\le\tau(uY)\le\tau(uZ)\le\kappa_\mathcal{Z}\tau(lZ)\le\kappa_\mathcal{Z}\varphi(Z) φ ( Y ) ≤ τ ( u Y ) ≤ τ ( u Z ) ≤ κ Z τ ( l Z ) ≤ κ Z φ ( Z ) . (c) The upper bound follows from (b) by averaging. For the lower bound, in M 2 M_2 M 2 with ρ = d i a g ( κ , 1 ) \rho=\mathop{\mathrm{diag}}(\kappa,1) ρ = diag ( κ , 1 ) , let A , B A,B A , B be the projections onto ( ± sin θ , cos θ ) (\pm\sin\theta,\cos\theta) ( ± sin θ , cos θ ) and x = sin 2 θ x=\sin^2\theta x = sin 2 θ . Then A + B = 2 d i a g ( x , 1 − x ) A+B=2\mathop{\mathrm{diag}}(x,1-x) A + B = 2 diag ( x , 1 − x ) , A p = A A^p=A A p = A , B p = B B^p=B B p = B , and the ratio is 2 p − 1 ( κ x p + ( 1 − x ) p ) / ( κ x + 1 − x ) 2^{p-1}(\kappa x^p+(1-x)^p)/(\kappa x+1-x) 2 p − 1 ( κ x p + ( 1 − x ) p ) / ( κ x + 1 − x ) ; apply Theorem 4.15 t:low . This function equals 1 1 1 at x = 1 2 x=\frac12 x = 2 1 and its logarithmic derivative there is 2 ( p − 1 ) ( κ − 1 ) / ( κ + 1 ) < 0 2(p-1)(\kappa-1)/(\kappa+1)<0 2 ( p − 1 ) ( κ − 1 ) / ( κ + 1 ) < 0 , so t p ( κ ) > 1 \mathsf t_p(\kappa)>1 t p ( κ ) > 1 . ◻
As p → 0 p\to0 p → 0 both bounds tend to ( 1 + κ Z ) / 2 (1+\kappa_\mathcal{Z})/2 ( 1 + κ Z ) /2 , the constant c ∨ c_\vee c ∨ of Theorem 5.32 p:c . Numerically the optimisers in M 2 M_2 M 2 and M 3 M_3 M 3 are always the symmetric pairs of rank-one projections used above, which suggests c p T ( ρ ) = t p ( κ Z ) c_p^T(\rho)=\mathsf t_p(\kappa_\mathcal{Z}) c p T ( ρ ) = t p ( κ Z ) . For a non-homogeneous function f f f in place of t p t^p t p , condition (R) fails and the constant of the f f f -version need not be a function of κ Z \kappa_\mathcal{Z} κ Z .
Items 12 and 13
Item 12 is Theorem 5.40 with p = 1 2 p=\frac12 p = 2 1 : t 1 / 2 ( κ Z ) ≤ c ≤ 1 + κ Z 2 \mathsf t_{1/2}(\kappa_\mathcal{Z})\le c\le\frac{1+\kappa_\mathcal{Z}}2 t 1/2 ( κ Z ) ≤ c ≤ 2 1 + κ Z . On M 2 M_2 M 2 the computed envelope of item 12 (Table 5 ) agrees with t 1 / 2 ( κ ) − 1 \mathsf t_{1/2}(\kappa)-1 t 1/2 ( κ ) − 1 to three decimal places at δ = 0.1 \delta=0.1 δ = 0.1 , 0.5 0.5 0.5 and 0.9 0.9 0.9 , and its leading coefficient 1 2 \frac12 2 1 (§ 10 ) is that of t 1 / 2 ( κ ) − 1 \mathsf t_{1/2}(\kappa)-1 t 1/2 ( κ ) − 1 ; this supports the conjecture c p T = t p ( κ Z ) c_p^T=\mathsf t_p(\kappa_\mathcal{Z}) c p T = t p ( κ Z ) for p = 1 2 p=\frac12 p = 2 1 . With the contrast of the operators prescribed, item 12 cannot be violated at all below a sharp threshold (Proposition 9.5 ).
For the Powers–Størmer inequality (item 13) only numerical values are available (§ 10 ). Its normalisation is chosen as follows.
The spectral geometric mean
For A , B ∈ M + + A,B\in\mathcal{M}_{++} A , B ∈ M ++ the spectral geometric mean is A ♮ B = ( A − 1 # B ) 1 / 2 A ( A − 1 # B ) 1 / 2 A\natural B=(A^{-1}\#B)^{1/2}A(A^{-1}\#B)^{1/2} A ♮ B = ( A − 1 # B ) 1/2 A ( A − 1 # B ) 1/2 , where X # Y = X 1 / 2 ( X − 1 / 2 Y X − 1 / 2 ) 1 / 2 X 1 / 2 X\#Y=X^{1/2}(X^{-1/2}YX^{-1/2})^{1/2}X^{1/2} X # Y = X 1/2 ( X − 1/2 Y X − 1/2 ) 1/2 X 1/2 is the geometric mean. Let c 6 ( ρ ) = sup { φ ( A ♮ B ) / φ ( A ) φ ( B ) : A , B ∈ M + + } c_6(\rho)=\sup\{\varphi(A\natural B)/\sqrt{\varphi(A)\varphi(B)}:A,B\in\mathcal{M}_{++}\} c 6 ( ρ ) = sup { φ ( A ♮ B ) / φ ( A ) φ ( B ) : A , B ∈ M ++ } .
Proposition 5.43 (Item 6). G Z ≤ c 6 ( ρ ) ≤ κ Z G_\mathcal{Z}\le c_6(\rho)\le\kappa_\mathcal{Z} G Z ≤ c 6 ( ρ ) ≤ κ Z in every finite von Neumann algebra. In M n ( C ) M_n(\mathbb{C}) M n ( C ) , G − 1 ≤ V n ≤ κ − 1 G-1\le V_n\le\kappa-1 G − 1 ≤ V n ≤ κ − 1 for item 6, and the value G G G of the quotient is approached by the explicit family in the proof.
Proof. Upper bound. Let Y = A 1 / 2 ( A − 1 # B ) 1 / 2 Y=A^{1/2}(A^{-1}\#B)^{1/2} Y = A 1/2 ( A − 1 # B ) 1/2 . Then A ♮ B = Y ∗ Y A\natural B=Y^*Y A ♮ B = Y ∗ Y and, since
A − 1 # B = A − 1 / 2 ( A 1 / 2 B A 1 / 2 ) 1 / 2 A − 1 / 2 , we have Y Y ∗ = ( A 1 / 2 B A 1 / 2 ) 1 / 2 = ∣ B 1 / 2 A 1 / 2 ∣ . A^{-1}\#B=A^{-1/2}(A^{1/2}BA^{1/2})^{1/2}A^{-1/2},\qquad\text{we have}\qquad
YY^*=(A^{1/2}BA^{1/2})^{1/2}=|B^{1/2}A^{1/2}|. A − 1 # B = A − 1/2 ( A 1/2 B A 1/2 ) 1/2 A − 1/2 , we have Y Y ∗ = ( A 1/2 B A 1/2 ) 1/2 = ∣ B 1/2 A 1/2 ∣. By the trace property and Hölder’s inequality, τ ( A ♮ B ) = τ ( ∣ B 1 / 2 A 1 / 2 ∣ ) ≤ ∥ B 1 / 2 ∥ 2 ∥ A 1 / 2 ∥ 2 = ( τ ( A ) τ ( B ) ) 1 / 2 \tau(A\natural B)=\tau(|B^{1/2}A^{1/2}|)\le\|B^{1/2}\|_2\|A^{1/2}\|_2=(\tau(A)\tau(B))^{1/2} τ ( A ♮ B ) = τ ( ∣ B 1/2 A 1/2 ∣ ) ≤ ∥ B 1/2 ∥ 2 ∥ A 1/2 ∥ 2 = ( τ ( A ) τ ( B ) ) 1/2 . Since τ ≤ ∥ ρ − 1 ∥ φ \tau\le\|\rho^{-1}\|\varphi τ ≤ ∥ ρ − 1 ∥ φ on M + \mathcal{M}_+ M + ,
φ ( A ♮ B ) ≤ ∥ ρ ∥ τ ( A ♮ B ) ≤ ∥ ρ ∥ ( τ ( A ) τ ( B ) ) 1 / 2 ≤ κ ( ρ ) ( φ ( A ) φ ( B ) ) 1 / 2 . \varphi(A\natural B)\le\|\rho\|\tau(A\natural B)\le\|\rho\|(\tau(A)\tau(B))^{1/2}\le\kappa(\rho)(\varphi(A)\varphi(B))^{1/2}. φ ( A ♮ B ) ≤ ∥ ρ ∥ τ ( A ♮ B ) ≤ ∥ ρ ∥ ( τ ( A ) τ ( B ) ) 1/2 ≤ κ ( ρ ) ( φ ( A ) φ ( B ) ) 1/2 . (R). Let z ∈ Z + + z\in\mathcal{Z}_{++} z ∈ Z ++ . As central elements commute with M \mathcal{M} M and with the functional calculus of its elements, ( s X ) # ( t Y ) = ( s t ) 1 / 2 ( X # Y ) (sX)\#(tY)=(st)^{1/2}(X\#Y) ( s X ) # ( t Y ) = ( s t ) 1/2 ( X # Y ) for s , t ∈ Z + + s,t\in\mathcal{Z}_{++} s , t ∈ Z ++ ; hence ( z A ) − 1 # ( z B ) = ( z − 1 A − 1 ) # ( z B ) = A − 1 # B (zA)^{-1}\#(zB)=(z^{-1}A^{-1})\#(zB)=A^{-1}\#B ( z A ) − 1 # ( z B ) = ( z − 1 A − 1 ) # ( z B ) = A − 1 # B and therefore ( z A ) ♮ ( z B ) = z ( A ♮ B ) (zA)\natural(zB)=z(A\natural B) ( z A ) ♮ ( z B ) = z ( A ♮ B ) . So the ratio for z ρ z\rho z ρ at ( A , B ) (A,B) ( A , B ) equals the ratio for ρ \rho ρ at ( z A , z B ) (zA,zB) ( z A , z B ) , and c 6 ( z ρ ) = c 6 ( ρ ) c_6(z\rho)=c_6(\rho) c 6 ( z ρ ) = c 6 ( ρ ) . By Lemma 4.12 , c 6 ( ρ ) = c 6 ( z 0 ρ ) ≤ κ ( z 0 ρ ) = κ Z ( ρ ) c_6(\rho)=c_6(z_0\rho)\le\kappa(z_0\rho)=\kappa_\mathcal{Z}(\rho) c 6 ( ρ ) = c 6 ( z 0 ρ ) ≤ κ ( z 0 ρ ) = κ Z ( ρ ) . Lower bound in M 2 M_2 M 2 . Let ρ = d i a g ( a , b ) \rho=\mathop{\mathrm{diag}}(a,b) ρ = diag ( a , b ) , e ± = ( e 1 ± e 2 ) / 2 e_\pm=(e_1\pm e_2)/\sqrt2 e ± = ( e 1 ± e 2 ) / 2 , v = e + v=e_+ v = e + and x = ρ − 1 v x=\rho^{-1}v x = ρ − 1 v . For invertible A A A and a rank-one B = v v ∗ B=vv^* B = v v ∗ , ∥ v ∥ = 1 \|v\|=1 ∥ v ∥ = 1 , we have A − 1 # v v ∗ = ( v ∗ A v ) − 1 / 2 v v ∗ A^{-1}\#vv^*=(v^*Av)^{-1/2}vv^* A − 1 # v v ∗ = ( v ∗ A v ) − 1/2 v v ∗ , hence ( A − 1 # B ) 1 / 2 = ( v ∗ A v ) − 1 / 4 v v ∗ (A^{-1}\#B)^{1/2}=(v^*Av)^{-1/4}vv^* ( A − 1 # B ) 1/2 = ( v ∗ A v ) − 1/4 v v ∗ and A ♮ B = ( v ∗ A v ) 1 / 2 v v ∗ A\natural B=(v^*Av)^{1/2}vv^* A ♮ B = ( v ∗ A v ) 1/2 v v ∗ . With A = x x ∗ + ε A=xx^*+\varepsilon A = x x ∗ + ε ,
φ ( A ♮ B ) 2 φ ( A ) φ ( B ) = ( v ∗ A v ) ( v ∗ ρ v ) T r ( ρ A ) → ε → 0 ∣ v ∗ x ∣ 2 v ∗ ρ v x ∗ ρ x = ( v ∗ ρ − 1 v ) ( v ∗ ρ v ) = 1 4 ( a − 1 + b − 1 ) ( a + b ) = G 2 . \frac{\varphi(A\natural B)^2}{\varphi(A)\varphi(B)}=\frac{(v^*Av)(v^*\rho v)}{\mathop{\mathrm{Tr}}(\rho A)}\ \xrightarrow[\varepsilon\to0]{}\
\frac{|v^*x|^2\,v^*\rho v}{x^*\rho x}=(v^*\rho^{-1}v)(v^*\rho v)=\tfrac14(a^{-1}+b^{-1})(a+b)=G^2 . φ ( A ) φ ( B ) φ ( A ♮ B ) 2 = Tr ( ρ A ) ( v ∗ A v ) ( v ∗ ρ v ) ε → 0 x ∗ ρ x ∣ v ∗ x ∣ 2 v ∗ ρ v = ( v ∗ ρ − 1 v ) ( v ∗ ρ v ) = 4 1 ( a − 1 + b − 1 ) ( a + b ) = G 2 . For fixed ε \varepsilon ε the quotient at ( A , v v ∗ + ε ′ ) (A,vv^*+\varepsilon') ( A , v v ∗ + ε ′ ) tends to the quotient at ( A , v v ∗ ) (A,vv^*) ( A , v v ∗ ) as ε ′ → 0 \varepsilon'\to0 ε ′ → 0 , by continuity; letting then ε → 0 \varepsilon\to0 ε → 0 , the quotient approaches G G G for positive definite A A A and B B B . (The limits must be taken in this order; for instance ε ′ = ε \varepsilon'=\varepsilon ε ′ = ε does not give G G G .) Transfer. Let ρ ∉ Z \rho\notin\mathcal{Z} ρ ∈ / Z and let π \pi π , t t t and R ε → d i a g ( α , β ) R_\varepsilon\to\mathop{\mathrm{diag}}(\alpha,\beta) R ε → diag ( α , β ) , α / β = κ Z ( ρ ) \alpha/\beta=\kappa_\mathcal{Z}(\rho) α / β = κ Z ( ρ ) , be as in the proof of Theorem 4.15 t:low , and e = π ( 1 ) e=\pi(1) e = π ( 1 ) . If e = 1 e=1 e = 1 we use π \pi π itself, a unital injective ∗ * ∗ -homomorphism, and drop the terms with s s s below. Otherwise the map π ′ ( X , λ ) = π ( X ) + λ ( 1 − e ) \pi'(X,\lambda)=\pi(X)+\lambda(1-e) π ′ ( X , λ ) = π ( X ) + λ ( 1 − e ) is a unital injective ∗ * ∗ -homomorphism M 2 ⊕ C → M M_2\oplus\mathbb{C}\to\mathcal{M} M 2 ⊕ C → M , so it commutes with inverses, square roots and products, and π ′ ( A 0 , s ) ♮ π ′ ( B 0 , s ) = π ′ ( A 0 ♮ B 0 , s ) \pi'(A_0,s)\natural\pi'(B_0,s)=\pi'(A_0\natural B_0,s) π ′ ( A 0 , s ) ♮ π ′ ( B 0 , s ) = π ′ ( A 0 ♮ B 0 , s ) for positive definite A 0 , B 0 ∈ M 2 A_0,B_0\in M_2 A 0 , B 0 ∈ M 2 and s > 0 s>0 s > 0 . Hence
φ ( π ′ ( A 0 , s ) ♮ π ′ ( B 0 , s ) ) φ ( π ′ ( A 0 , s ) ) φ ( π ′ ( B 0 , s ) ) = t T r ( R ε ( A 0 ♮ B 0 ) ) + s φ ( 1 − e ) ( t T r ( R ε A 0 ) + s φ ( 1 − e ) ) ( t T r ( R ε B 0 ) + s φ ( 1 − e ) ) , \frac{\varphi(\pi'(A_0,s)\natural\pi'(B_0,s))}{\sqrt{\varphi(\pi'(A_0,s))\varphi(\pi'(B_0,s))}}
=\frac{t\mathop{\mathrm{Tr}}(R_\varepsilon(A_0\natural B_0))+s\varphi(1-e)}{\sqrt{(t\mathop{\mathrm{Tr}}(R_\varepsilon A_0)+s\varphi(1-e))(t\mathop{\mathrm{Tr}}(R_\varepsilon B_0)+s\varphi(1-e))}}, φ ( π ′ ( A 0 , s )) φ ( π ′ ( B 0 , s )) φ ( π ′ ( A 0 , s ) ♮ π ′ ( B 0 , s )) = ( t Tr ( R ε A 0 ) + s φ ( 1 − e )) ( t Tr ( R ε B 0 ) + s φ ( 1 − e )) t Tr ( R ε ( A 0 ♮ B 0 )) + s φ ( 1 − e ) , which tends to the ratio for d i a g ( α , β ) \mathop{\mathrm{diag}}(\alpha,\beta) diag ( α , β ) at ( A 0 , B 0 ) (A_0,B_0) ( A 0 , B 0 ) as s → 0 s\to0 s → 0 and then ε → 0 \varepsilon\to0 ε → 0 . With the M 2 M_2 M 2 examples above, c 6 ( ρ ) ≥ G ( d i a g ( α , β ) ) = G Z ( ρ ) c_6(\rho)\ge G(\mathop{\mathrm{diag}}(\alpha,\beta))=G_\mathcal{Z}(\rho) c 6 ( ρ ) ≥ G ( diag ( α , β )) = G Z ( ρ ) . For central ρ \rho ρ both bounds are 1 1 1 . In M n M_n M n the statements follow with κ Z = κ \kappa_\mathcal{Z}=\kappa κ Z = κ . ◻
Numerically, for n = 3 n=3 n = 3 and 4 4 4 the computed envelope of item 6 exceeds G − 1 G-1 G − 1 at some δ \delta δ (§ 10 ). This indicates that the constant on M n M_n M n , n ≥ 3 n\ge3 n ≥ 3 , is larger than G G G , but we do not use it as a proof. For n = 2 n=2 n = 2 we conjecture that the constant is G G G (Conjecture 11.1 ).
Further inequalities
The Alhasan–Fawwaz inequality
Let
a ( ρ ) = sup { φ ( A ) φ ( ∣ A + i B ∣ ) : A ∈ M + , B ∈ M s a , A + i B ≠ 0 } . a(\rho)=\sup\Bigl\{\frac{\varphi(A)}{\varphi(|A+iB|)}:A\in\mathcal{M}_+,\ B\in\mathcal{M}_{sa},\ A+iB\ne0\Bigr\}. a ( ρ ) = sup { φ ( ∣ A + i B ∣ ) φ ( A ) : A ∈ M + , B ∈ M s a , A + i B = 0 } .
Since φ ( A ) \varphi(A) φ ( A ) and φ ( B ) \varphi(B) φ ( B ) are real, φ ( A ) = R e φ ( X ) \varphi(A)=\mathop{\mathrm{Re}}\varphi(X) φ ( A ) = Re φ ( X ) for X = A + i B X=A+iB X = A + i B . So a ( ρ ) a(\rho) a ( ρ ) is the supremum of R e φ ( X ) / φ ( ∣ X ∣ ) \mathop{\mathrm{Re}}\varphi(X)/\varphi(|X|) Re φ ( X ) / φ ( ∣ X ∣ ) over accretive X X X , i.e. over X X X with R e X ≥ 0 \mathop{\mathrm{Re}}X\ge0 Re X ≥ 0 . By [86 ] , a ( ρ ) = 1 a(\rho)=1 a ( ρ ) = 1 if and only if ρ ∈ Z \rho\in\mathcal{Z} ρ ∈ Z ; this also follows from af:a below.
Proposition 5.44 (Item 35).
a ( z ρ ) = a ( ρ ) a(z\rho)=a(\rho) a ( z ρ ) = a ( ρ ) for z ∈ Z + + z\in\mathcal{Z}_{++} z ∈ Z ++ , and a M 2 ( d i a g ( κ Z , 1 ) ) ≤ a ( ρ ) ≤ G Z a_{M_2}(\mathop{\mathrm{diag}}(\kappa_\mathcal{Z},1))\le a(\rho)\le G_\mathcal{Z} a M 2 ( diag ( κ Z , 1 )) ≤ a ( ρ ) ≤ G Z . If ρ ∉ Z \rho\notin\mathcal{Z} ρ ∈ / Z , then a ( ρ ) > 1 a(\rho)>1 a ( ρ ) > 1 .
If M = M n ( C ) \mathcal{M}=M_n(\mathbb{C}) M = M n ( C ) and ρ \rho ρ is not scalar, the supremum is attained and a ( ρ ) < cosh O L C ( ρ ) a(\rho)<\cosh\mathop{\mathrm{OLC}}(\rho) a ( ρ ) < cosh OLC ( ρ ) . In particular, G Z G_\mathcal{Z} G Z is never attained in M n ( C ) M_n(\mathbb{C}) M n ( C ) .
If X ∈ M 2 ( C ) X\in M_2(\mathbb{C}) X ∈ M 2 ( C ) and R e X ≥ 0 \mathop{\mathrm{Re}}X\ge0 Re X ≥ 0 , then R e X ≤ 2 ∣ X ∣ \mathop{\mathrm{Re}}X\le2|X| Re X ≤ 2∣ X ∣ . Consequently a ( ρ ) < 2 a(\rho)<2 a ( ρ ) < 2 for every ρ ∈ M 2 ( C ) + + \rho\in M_2(\mathbb{C})_{++} ρ ∈ M 2 ( C ) ++ .
For ρ = d i a g ( 10 4 , 1 , 1 ) ∈ M 3 ( C ) \rho=\mathop{\mathrm{diag}}(10^4,1,1)\in M_3(\mathbb{C}) ρ = diag ( 1 0 4 , 1 , 1 ) ∈ M 3 ( C ) , a ( ρ ) > 2.68 a(\rho)>2.68 a ( ρ ) > 2.68 . Hence a a a is not a function of κ Z \kappa_\mathcal{Z} κ Z , even on finite factors: d i a g ( 10 4 , 1 ) ∈ M 2 ( C ) \mathop{\mathrm{diag}}(10^4,1)\in M_2(\mathbb{C}) diag ( 1 0 4 , 1 ) ∈ M 2 ( C ) and d i a g ( 10 4 , 1 , 1 ) ∈ M 3 ( C ) \mathop{\mathrm{diag}}(10^4,1,1)\in M_3(\mathbb{C}) diag ( 1 0 4 , 1 , 1 ) ∈ M 3 ( C ) both have κ Z = 10 4 \kappa_\mathcal{Z}=10^4 κ Z = 1 0 4 .
Proof. af:a (R) holds with σ z ( A , B ) = ( z A , z B ) \sigma_z(A,B)=(zA,zB) σ z ( A , B ) = ( z A , z B ) . The upper bound is Theorem 5.1 : φ ( A ) = R e φ ( X ) ≤ ∣ φ ( X ) ∣ ≤ G Z φ ( ∣ X ∣ ) \varphi(A)=\mathop{\mathrm{Re}}\varphi(X)\le|\varphi(X)|\le G_\mathcal{Z}\varphi(|X|) φ ( A ) = Re φ ( X ) ≤ ∣ φ ( X ) ∣ ≤ G Z φ ( ∣ X ∣ ) . The lower bound is Theorem 4.15 t:low , with the test class X 1 ∈ M + X_1\in\mathcal{M}_+ X 1 ∈ M + , X 2 ∈ M s a X_2\in\mathcal{M}_{sa} X 2 ∈ M s a and F 1 = X 1 F_1=X_1 F 1 = X 1 , F 2 = ∣ X 1 + i X 2 ∣ F_2=|X_1+iX_2| F 2 = ∣ X 1 + i X 2 ∣ . Let ρ ∉ Z \rho\notin\mathcal{Z} ρ ∈ / Z . Choose S ∈ M s a S\in\mathcal{M}_{sa} S ∈ M s a with C = [ S , ρ ] ≠ 0 C=[S,\rho]\ne0 C = [ S , ρ ] = 0 ; then C ∗ = − C C^*=-C C ∗ = − C . For λ > 0 \lambda>0 λ > 0 let R = − λ i C ∈ M s a R=-\lambda iC\in\mathcal{M}_{sa} R = − λi C ∈ M s a and X t = 1 + t ( R + i S ) X_t=1+t(R+iS) X t = 1 + t ( R + i S ) . For small t > 0 t>0 t > 0 we have R e X t = 1 + t R ≥ 0 \mathop{\mathrm{Re}}X_t=1+tR\ge0 Re X t = 1 + tR ≥ 0 . Since X t ∗ X t = 1 + 2 t R + t 2 ( R 2 + S 2 + i [ R , S ] ) X_t^*X_t=1+2tR+t^2(R^2+S^2+i[R,S]) X t ∗ X t = 1 + 2 tR + t 2 ( R 2 + S 2 + i [ R , S ]) , ∣ X t ∣ = 1 + t R + t 2 2 ( S 2 + i [ R , S ] ) + O ( t 3 ) |X_t|=1+tR+\frac{t^2}2(S^2+i[R,S])+O(t^3) ∣ X t ∣ = 1 + tR + 2 t 2 ( S 2 + i [ R , S ]) + O ( t 3 ) , and
φ ( R e X t ) φ ( ∣ X t ∣ ) = 1 − t 2 2 φ ( 1 ) φ ( S 2 + i [ R , S ] ) + O ( t 3 ) . \frac{\varphi(\mathop{\mathrm{Re}}X_t)}{\varphi(|X_t|)}=1-\frac{t^2}{2\varphi(1)}\,\varphi\bigl(S^2+i[R,S]\bigr)+O(t^3). φ ( ∣ X t ∣ ) φ ( Re X t ) = 1 − 2 φ ( 1 ) t 2 φ ( S 2 + i [ R , S ] ) + O ( t 3 ) . Here φ ( i [ R , S ] ) = i τ ( R [ S , ρ ] ) = λ τ ( C 2 ) = − λ τ ( C ∗ C ) \varphi(i[R,S])=i\tau(R[S,\rho])=\lambda\tau(C^2)=-\lambda\tau(C^*C) φ ( i [ R , S ]) = i τ ( R [ S , ρ ]) = λ τ ( C 2 ) = − λ τ ( C ∗ C ) . So for λ > φ ( S 2 ) / τ ( C ∗ C ) \lambda>\varphi(S^2)/\tau(C^*C) λ > φ ( S 2 ) / τ ( C ∗ C ) the ratio exceeds 1 1 1 for small t t t .
af:b The ratio is invariant under positive scalings of ( A , B ) (A,B) ( A , B ) . On the compact set { ∥ A + i B ∥ 2 = 1 } \{\|A+iB\|_2=1\} { ∥ A + i B ∥ 2 = 1 } it is continuous, and its denominator is at least λ min ( ρ ) T r ∣ X ∣ > 0 \lambda_{\min}(\rho)\mathop{\mathrm{Tr}}|X|>0 λ m i n ( ρ ) Tr ∣ X ∣ > 0 , so the supremum is attained. Suppose the ratio equals cosh O L C ( ρ ) \cosh\mathop{\mathrm{OLC}}(\rho) cosh OLC ( ρ ) at X = U ∣ X ∣ X=U|X| X = U ∣ X ∣ , with U U U unitary, and let h = ∣ X ∣ 1 / 2 h=|X|^{1/2} h = ∣ X ∣ 1/2 . Then the chain R e ⟨ U h , ρ h ⟩ ≤ ∥ h ∥ 2 ∥ ρ h ∥ 2 ≤ cosh O L C ( ρ ) ⟨ ρ h , h ⟩ \mathop{\mathrm{Re}}\langle Uh,\rho h\rangle\le\|h\|_2\|\rho h\|_2\le\cosh\mathop{\mathrm{OLC}}(\rho)\langle\rho h,h\rangle Re ⟨ U h , ρ h ⟩ ≤ ∥ h ∥ 2 ∥ ρ h ∥ 2 ≤ cosh OLC ( ρ ) ⟨ ρ h , h ⟩ from the proof of Theorem 5.1 consists of equalities. Let M > m M>m M > m be the extreme eigenvalues of ρ \rho ρ , P M , P m P_M,P_m P M , P m the corresponding spectral projections and E = ran ( P M + P m ) E=\operatorname{ran}(P_M+P_m) E = ran ( P M + P m ) . Equality in Lemma 3.1 for L ρ L_\rho L ρ means ( L ρ − m ) ( M − L ρ ) h = 0 (L_\rho-m)(M-L_\rho)h=0 ( L ρ − m ) ( M − L ρ ) h = 0 and ∥ ρ h ∥ 2 2 = m M ∥ h ∥ 2 2 \|\rho h\|_2^2=mM\|h\|_2^2 ∥ ρ h ∥ 2 2 = m M ∥ h ∥ 2 2 . Hence V : = ran h ⊂ E V:=\operatorname{ran}h\subset E V := ran h ⊂ E and M ∥ P M h ∥ 2 2 = m ∥ P m h ∥ 2 2 M\|P_Mh\|_2^2=m\|P_mh\|_2^2 M ∥ P M h ∥ 2 2 = m ∥ P m h ∥ 2 2 , and both norms are nonzero. Equality in the Cauchy–Schwarz inequality gives U h = λ ρ h Uh=\lambda\rho h U h = λ ρ h , and φ ( X ) = λ ∥ ρ h ∥ 2 2 ≥ 0 \varphi(X)=\lambda\|\rho h\|_2^2\ge0 φ ( X ) = λ ∥ ρ h ∥ 2 2 ≥ 0 gives λ > 0 \lambda>0 λ > 0 . So X = λ ρ K X=\lambda\rho K X = λ ρ K with K = h 2 = ∣ X ∣ K=h^2=|X| K = h 2 = ∣ X ∣ . Also h 2 = h U ∗ U h = λ 2 h ρ 2 h h^2=hU^*Uh=\lambda^2h\rho^2h h 2 = h U ∗ U h = λ 2 h ρ 2 h , so the compression of ρ 2 = M 2 P M + m 2 P m \rho^2=M^2P_M+m^2P_m ρ 2 = M 2 P M + m 2 P m to V V V is λ − 2 \lambda^{-2} λ − 2 . Thus P V P M P V = t P V P_VP_MP_V=tP_V P V P M P V = t P V and P V P m P V = ( 1 − t ) P V P_VP_mP_V=(1-t)P_V P V P m P V = ( 1 − t ) P V for some t t t , and t ∈ ( 0 , 1 ) t\in(0,1) t ∈ ( 0 , 1 ) because P M h ≠ 0 ≠ P m h P_Mh\ne0\ne P_mh P M h = 0 = P m h . Take x ∈ V x\in V x ∈ V with ⟨ K x , x ⟩ > 0 \langle Kx,x\rangle>0 ⟨ K x , x ⟩ > 0 , and ξ = P M x − c P m x \xi=P_Mx-cP_mx ξ = P M x − c P m x with t / ( 1 − t ) < c < M t / ( m ( 1 − t ) ) t/(1-t)<c<Mt/(m(1-t)) t / ( 1 − t ) < c < M t / ( m ( 1 − t )) . Then P V ξ = ( t − c ( 1 − t ) ) x P_V\xi=(t-c(1-t))x P V ξ = ( t − c ( 1 − t )) x and P V ρ ξ = ( M t − c m ( 1 − t ) ) x P_V\rho\xi=(Mt-cm(1-t))x P V ρ ξ = ( M t − c m ( 1 − t )) x . Since K = P V K P V K=P_VKP_V K = P V K P V ,
⟨ ( R e X ) ξ , ξ ⟩ = λ R e ⟨ K ξ , ρ ξ ⟩ = λ ( t − c ( 1 − t ) ) ( M t − c m ( 1 − t ) ) ⟨ K x , x ⟩ < 0 , \langle(\mathop{\mathrm{Re}}X)\xi,\xi\rangle=\lambda\mathop{\mathrm{Re}}\langle K\xi,\rho\xi\rangle
=\lambda\bigl(t-c(1-t)\bigr)\bigl(Mt-cm(1-t)\bigr)\langle Kx,x\rangle<0, ⟨( Re X ) ξ , ξ ⟩ = λ Re ⟨ K ξ , ρ ξ ⟩ = λ ( t − c ( 1 − t ) ) ( M t − c m ( 1 − t ) ) ⟨ K x , x ⟩ < 0 , which contradicts R e X ≥ 0 \mathop{\mathrm{Re}}X\ge0 Re X ≥ 0 .
af:c Let first X X X be invertible, X = U P X=UP X = U P with P = ∣ X ∣ P=|X| P = ∣ X ∣ , and T = P − 1 / 2 U P 1 / 2 T=P^{-1/2}UP^{1/2} T = P − 1/2 U P 1/2 . Then X = P 1 / 2 T P 1 / 2 X=P^{1/2}TP^{1/2} X = P 1/2 T P 1/2 , so R e X = P 1 / 2 ( R e T ) P 1 / 2 \mathop{\mathrm{Re}}X=P^{1/2}(\mathop{\mathrm{Re}}T)P^{1/2} Re X = P 1/2 ( Re T ) P 1/2 and R e T ≥ 0 \mathop{\mathrm{Re}}T\ge0 Re T ≥ 0 . As T T T is similar to U U U , it is diagonalisable with eigenvalues λ 1 , λ 2 \lambda_1,\lambda_2 λ 1 , λ 2 of modulus 1 1 1 . Let c = 1 2 ( λ 1 + λ 2 ) c=\frac12(\lambda_1+\lambda_2) c = 2 1 ( λ 1 + λ 2 ) . The traceless matrix T − c T-c T − c is unitarily similar to a matrix with zero diagonal, and d i a g ( 1 , − 1 ) \mathop{\mathrm{diag}}(1,-1) diag ( 1 , − 1 ) conjugates such a matrix to its negative. Hence the numerical range satisfies W ( T ) = 2 c − W ( T ) W(T)=2c-W(T) W ( T ) = 2 c − W ( T ) . Since W ( T ) ⊂ { R e z ≥ 0 } W(T)\subset\{\mathop{\mathrm{Re}}z\ge0\} W ( T ) ⊂ { Re z ≥ 0 } , we get R e z ≤ 2 R e c \mathop{\mathrm{Re}}z\le2\mathop{\mathrm{Re}}c Re z ≤ 2 Re c on W ( T ) W(T) W ( T ) , i.e. R e T ≤ 2 R e c ≤ 2 \mathop{\mathrm{Re}}T\le2\mathop{\mathrm{Re}}c\le2 Re T ≤ 2 Re c ≤ 2 . If R e c = 1 \mathop{\mathrm{Re}}c=1 Re c = 1 , then λ 1 = λ 2 = 1 \lambda_1=\lambda_2=1 λ 1 = λ 2 = 1 , so T = 1 T=1 T = 1 and R e X = P \mathop{\mathrm{Re}}X=P Re X = P . In all cases R e X ≤ c ′ P \mathop{\mathrm{Re}}X\le c'P Re X ≤ c ′ P for some c ′ < 2 c'<2 c ′ < 2 . If X ≠ 0 X\ne0 X = 0 is singular, then X = ∣ u ⟩ ⟨ v ∣ X=|u\rangle\langle v| X = ∣ u ⟩ ⟨ v ∣ . Here u = μ v u=\mu v u = μv with R e μ ≥ 0 \mathop{\mathrm{Re}}\mu\ge0 Re μ ≥ 0 , since otherwise ⟨ ( R e X ) ξ , ξ ⟩ = R e ( ⟨ ξ , v ⟩ ⟨ u , ξ ⟩ ) < 0 \langle(\mathop{\mathrm{Re}}X)\xi,\xi\rangle=\mathop{\mathrm{Re}}(\langle\xi,v\rangle\langle u,\xi\rangle)<0 ⟨( Re X ) ξ , ξ ⟩ = Re (⟨ ξ , v ⟩ ⟨ u , ξ ⟩) < 0 for a suitable ξ \xi ξ ; hence R e X ≤ ∣ X ∣ \mathop{\mathrm{Re}}X\le|X| Re X ≤ ∣ X ∣ . Thus φ ( R e X ) < 2 φ ( ∣ X ∣ ) \varphi(\mathop{\mathrm{Re}}X)<2\varphi(|X|) φ ( Re X ) < 2 φ ( ∣ X ∣ ) for every accretive X ≠ 0 X\ne0 X = 0 , and a ( ρ ) < 2 a(\rho)<2 a ( ρ ) < 2 because the supremum is attained (by af:b ; for scalar ρ \rho ρ , a = 1 a=1 a = 1 ).
af:d Let
P = ( 2 19 20 19 637 491 20 491 416 ) , U = 1 10485 ( 9965 − 1180 3040 2020 9907 − 2776 − 2560 3224 9643 ) , P=\begin{pmatrix}2&19&20\\19&637&491\\20&491&416\end{pmatrix},\qquad
U=\frac1{10485}\begin{pmatrix}9965&-1180&3040\\2020&9907&-2776\\-2560&3224&9643\end{pmatrix}, P = 2 19 20 19 637 491 20 491 416 , U = 10485 1 9965 2020 − 2560 − 1180 9907 3224 3040 − 2776 9643 , and X = U P X=UP X = U P . The matrix U U U is orthogonal; it is the Cayley transform ( 1 − N ) ( 1 + N ) − 1 (1-N)(1+N)^{-1} ( 1 − N ) ( 1 + N ) − 1 of the skew-symmetric N N N with N 12 = 2 25 N_{12}=\frac2{25} N 12 = 25 2 , N 13 = − 7 50 N_{13}=-\frac7{50} N 13 = − 50 7 , N 23 = 3 20 N_{23}=\frac3{20} N 23 = 20 3 . The matrix P P P is positive definite, since its leading principal minors are 2 2 2 , 913 913 913 and 16006 16006 16006 . Hence ∣ X ∣ = P |X|=P ∣ X ∣ = P . An exact computation in rational arithmetic, which can be checked by hand, shows that the leading principal minors of R e X \mathop{\mathrm{Re}}X Re X are 11662 2097 \frac{11662}{2097} 2097 11662 , 6082302974 109935225 \frac{6082302974}{109935225} 109935225 6082302974 and 36535906 7329015 \frac{36535906}{7329015} 7329015 36535906 , so R e X > 0 \mathop{\mathrm{Re}}X>0 Re X > 0 . It also gives φ ( R e X ) = 39575293 699 \varphi(\mathop{\mathrm{Re}}X)=\frac{39575293}{699} φ ( Re X ) = 699 39575293 and φ ( ∣ X ∣ ) = φ ( P ) = 21053 \varphi(|X|)=\varphi(P)=21053 φ ( ∣ X ∣ ) = φ ( P ) = 21053 , so φ ( R e X ) / φ ( ∣ X ∣ ) = 39575293 14716047 > 2.68 \varphi(\mathop{\mathrm{Re}}X)/\varphi(|X|)=\frac{39575293}{14716047}>2.68 φ ( Re X ) / φ ( ∣ X ∣ ) = 14716047 39575293 > 2.68 . Together with af:c this proves the claim. ◻
The constant 2 2 2 in af:c is sharp, but not attained. Every T ∈ M 2 ( C ) T\in M_2(\mathbb{C}) T ∈ M 2 ( C ) that is similar to a unitary can be written as T = P − 1 / 2 U P 1 / 2 T=P^{-1/2}UP^{1/2} T = P − 1/2 U P 1/2 with P > 0 P>0 P > 0 and U U U unitary: if T = S D S − 1 T=SDS^{-1} T = S D S − 1 with D D D unitary and S = ∣ S ∗ ∣ V S=|S^*|V S = ∣ S ∗ ∣ V is the polar decomposition, take P = ∣ S ∗ ∣ − 2 P=|S^*|^{-2} P = ∣ S ∗ ∣ − 2 and U = V D V ∗ U=VDV^* U = V D V ∗ . Then X = U P X=UP X = U P satisfies ∣ X ∣ = P |X|=P ∣ X ∣ = P and R e X = P 1 / 2 ( R e T ) P 1 / 2 \mathop{\mathrm{Re}}X=P^{1/2}(\mathop{\mathrm{Re}}T)P^{1/2} Re X = P 1/2 ( Re T ) P 1/2 . For T ϵ = ( e i ϵ 2 cos ϵ 0 e − i ϵ ) T_\epsilon=\bigl(\begin{smallmatrix}e^{i\epsilon}&2\cos\epsilon\\0&e^{-i\epsilon}\end{smallmatrix}\bigr) T ϵ = ( e i ϵ 0 2 c o s ϵ e − i ϵ ) , 0 < ϵ < π / 2 0<\epsilon<\pi/2 0 < ϵ < π /2 , the eigenvalues e ± i ϵ e^{\pm i\epsilon} e ± i ϵ are distinct and unimodular, and R e T ϵ = cos ϵ ( 1 1 1 1 ) ≥ 0 \mathop{\mathrm{Re}}T_\epsilon=\cos\epsilon\bigl(\begin{smallmatrix}1&1\\1&1\end{smallmatrix}\bigr)\ge0 Re T ϵ = cos ϵ ( 1 1 1 1 ) ≥ 0 has largest eigenvalue 2 cos ϵ 2\cos\epsilon 2 cos ϵ . So the best constant c 2 c_2 c 2 in the operator inequality R e X ≤ c 2 ∣ X ∣ \mathop{\mathrm{Re}}X\le c_2|X| Re X ≤ c 2 ∣ X ∣ for accretive X ∈ M 2 ( C ) X\in M_2(\mathbb{C}) X ∈ M 2 ( C ) equals 2 2 2 .
Exponential inequalities
Bikchentaev’s characterisation of traces by φ ( e P Q P ) ≤ φ ( e Q ) \varphi(e^{PQP})\le\varphi(e^Q) φ ( e P QP ) ≤ φ ( e Q ) concerns projections P , Q P,Q P , Q (see [99 ] for the precise statement). For traces it remains true for all Q ≥ 0 Q\ge0 Q ≥ 0 (Lemma 5.46 ), but it fails for self-adjoint Q Q Q : for Q = − t 1 Q=-t1 Q = − t 1 with t > 0 t>0 t > 0 and 0 ≠ P ≠ 1 0\ne P\ne1 0 = P = 1 , τ ( e P Q P ) = e − t τ ( P ) + τ ( 1 − P ) > τ ( e Q ) \tau(e^{PQP})=e^{-t}\tau(P)+\tau(1-P)>\tau(e^Q) τ ( e P QP ) = e − t τ ( P ) + τ ( 1 − P ) > τ ( e Q ) . Let
c exp ( ρ ) = sup φ ( e P Q P ) φ ( e Q ) , b ( ρ ) = sup φ ( P e P Q P ) φ ( e Q ) . c_{\exp}(\rho)=\sup\frac{\varphi(e^{PQP})}{\varphi(e^Q)},\qquad
b(\rho)=\sup\frac{\varphi(Pe^{PQP})}{\varphi(e^Q)}. c e x p ( ρ ) = sup φ ( e Q ) φ ( e P QP ) , b ( ρ ) = sup φ ( e Q ) φ ( P e P QP ) .
In c exp c_{\exp} c e x p , P P P runs through the projections and Q Q Q through M + \mathcal{M}_+ M + . Let c exp p r c_{\exp}^{\rm pr} c e x p pr be the same supremum with Q Q Q also a projection. In b b b , P ≠ 0 P\ne0 P = 0 runs through the projections and Q Q Q through M s a \mathcal{M}_{sa} M s a . The quantity b b b is invariant under Q ↦ Q + s Q\mapsto Q+s Q ↦ Q + s . Let b 2 ( κ ) b_2(\kappa) b 2 ( κ ) denote b ( d i a g ( κ , 1 ) ) b(\mathop{\mathrm{diag}}(\kappa,1)) b ( diag ( κ , 1 )) computed in M 2 ( C ) M_2(\mathbb{C}) M 2 ( C ) .
Lemma 5.46. Let τ \tau τ be a faithful normal finite trace on M \mathcal{M} M , P P P a projection and Q ∈ M s a Q\in\mathcal{M}_{sa} Q ∈ M s a . Then τ ( P e P Q P ) ≤ τ ( P e Q P ) ≤ τ ( e Q ) \tau(Pe^{PQP})\le\tau(Pe^QP)\le\tau(e^Q) τ ( P e P QP ) ≤ τ ( P e Q P ) ≤ τ ( e Q ) . If Q ≥ 0 Q\ge0 Q ≥ 0 , then τ ( e P Q P ) ≤ τ ( e Q ) \tau(e^{PQP})\le\tau(e^Q) τ ( e P QP ) ≤ τ ( e Q ) .
Proof. The first inequality is Jensen’s trace inequality for the unital compression to P M P P\mathcal{M}P P M P (cf. [114 ] ); we give a direct proof. Let ε > 0 \varepsilon>0 ε > 0 , let I 1 , … , I N I_1,\dots,I_N I 1 , … , I N be disjoint intervals of length < ε <\varepsilon < ε covering σ ( P Q P ) \sigma(PQP) σ ( P QP ) , with left endpoints λ j \lambda_j λ j , and let E j = P 1 I j ( P Q P ) E_j=P\mathbf 1_{I_j}(PQP) E j = P 1 I j ( P QP ) . The E j E_j E j are projections, they commute with P Q P PQP P QP , and ∑ j E j = P \sum_jE_j=P ∑ j E j = P . Moreover E j P Q P E j ≥ λ j E j E_jPQPE_j\ge\lambda_jE_j E j P QP E j ≥ λ j E j and E j e P Q P ≤ e λ j + ε E j E_je^{PQP}\le e^{\lambda_j+\varepsilon}E_j E j e P QP ≤ e λ j + ε E j . For E j ≠ 0 E_j\ne0 E j = 0 , ω j = τ ( E j ⋅ E j ) / τ ( E j ) \omega_j=\tau(E_j\,\cdot\,E_j)/\tau(E_j) ω j = τ ( E j ⋅ E j ) / τ ( E j ) is a normal state. Jensen’s inequality for the spectral measure of Q Q Q in ω j \omega_j ω j gives ω j ( e Q ) ≥ e ω j ( Q ) \omega_j(e^Q)\ge e^{\omega_j(Q)} ω j ( e Q ) ≥ e ω j ( Q ) , and ω j ( Q ) = τ ( E j P Q P E j ) / τ ( E j ) ≥ λ j \omega_j(Q)=\tau(E_jPQPE_j)/\tau(E_j)\ge\lambda_j ω j ( Q ) = τ ( E j P QP E j ) / τ ( E j ) ≥ λ j . Hence τ ( P e Q P ) = ∑ j τ ( E j e Q E j ) ≥ ∑ j e λ j τ ( E j ) ≥ e − ε ∑ j τ ( E j e P Q P ) = e − ε τ ( P e P Q P ) \tau(Pe^QP)=\sum_j\tau(E_je^QE_j)\ge\sum_je^{\lambda_j}\tau(E_j)\ge e^{-\varepsilon}\sum_j\tau(E_je^{PQP})
=e^{-\varepsilon}\tau(Pe^{PQP}) τ ( P e Q P ) = ∑ j τ ( E j e Q E j ) ≥ ∑ j e λ j τ ( E j ) ≥ e − ε ∑ j τ ( E j e P QP ) = e − ε τ ( P e P QP ) ; let ε → 0 \varepsilon\to0 ε → 0 . The second inequality holds because e Q ≥ 0 e^Q\ge0 e Q ≥ 0 . If Q ≥ 0 Q\ge0 Q ≥ 0 , then e Q ≥ 1 e^Q\ge1 e Q ≥ 1 and τ ( e P Q P ) = τ ( P e P Q P ) + τ ( 1 − P ) ≤ τ ( P e Q P ) + τ ( ( 1 − P ) e Q ( 1 − P ) ) = τ ( e Q ) \tau(e^{PQP})=\tau(Pe^{PQP})+\tau(1-P)\le\tau(Pe^QP)+\tau((1-P)e^Q(1-P))=\tau(e^Q) τ ( e P QP ) = τ ( P e P QP ) + τ ( 1 − P ) ≤ τ ( P e Q P ) + τ (( 1 − P ) e Q ( 1 − P )) = τ ( e Q ) . ◻
Proposition 5.47 (Items 20 and 36).
c exp p r ≤ c exp ≤ κ Z c^{\rm pr}_{\exp}\le c_{\exp}\le\kappa_\mathcal{Z} c e x p pr ≤ c e x p ≤ κ Z , b ≤ κ Z b\le\kappa_\mathcal{Z} b ≤ κ Z , and b ≤ c exp ≤ 1 + b b\le c_{\exp}\le1+b b ≤ c e x p ≤ 1 + b .
b ( z ρ ) = b ( ρ ) b(z\rho)=b(\rho) b ( z ρ ) = b ( ρ ) for z ∈ Z + + z\in\mathcal{Z}_{++} z ∈ Z ++ , and b ( ρ ) ≥ b 2 ( κ Z ) b(\rho)\ge b_2(\kappa_\mathcal{Z}) b ( ρ ) ≥ b 2 ( κ Z ) ; moreover b 2 ( κ ) > 1 b_2(\kappa)>1 b 2 ( κ ) > 1 for κ > 1 \kappa>1 κ > 1 . If ρ ∉ Z \rho\notin\mathcal{Z} ρ ∈ / Z , then c exp p r ( ρ ) > 1 c^{\rm pr}_{\exp}(\rho)>1 c e x p pr ( ρ ) > 1 .
Neither c exp c_{\exp} c e x p nor c exp p r c^{\rm pr}_{\exp} c e x p pr is a function of κ Z \kappa_\mathcal{Z} κ Z . For κ > 1 \kappa>1 κ > 1 and ρ λ = d i a g ( κ , 1 ) ⊕ λ 1 2 ∈ M 2 ( C ) ⊕ M 2 ( C ) \rho_\lambda=\mathop{\mathrm{diag}}(\kappa,1)\oplus\lambda1_2\in M_2(\mathbb{C})\oplus M_2(\mathbb{C}) ρ λ = diag ( κ , 1 ) ⊕ λ 1 2 ∈ M 2 ( C ) ⊕ M 2 ( C ) , with τ = T r ⊕ T r \tau=\mathop{\mathrm{Tr}}\oplus\mathop{\mathrm{Tr}} τ = Tr ⊕ Tr , we have κ Z ( ρ λ ) = κ \kappa_\mathcal{Z}(\rho_\lambda)=\kappa κ Z ( ρ λ ) = κ for all λ > 0 \lambda>0 λ > 0 . As λ → 0 \lambda\to0 λ → 0 both constants tend to their values for d i a g ( κ , 1 ) \mathop{\mathrm{diag}}(\kappa,1) diag ( κ , 1 ) in M 2 ( C ) M_2(\mathbb{C}) M 2 ( C ) , which exceed b 2 ( κ ) b_2(\kappa) b 2 ( κ ) and 1 1 1 respectively. As λ → ∞ \lambda\to\infty λ → ∞ , c exp ( ρ λ ) → b 2 ( κ ) c_{\exp}(\rho_\lambda)\to b_2(\kappa) c e x p ( ρ λ ) → b 2 ( κ ) and c exp p r ( ρ λ ) → 1 c^{\rm pr}_{\exp}(\rho_\lambda)\to1 c e x p pr ( ρ λ ) → 1 .
Let c T c_T c T denote c exp c_{\exp} c e x p with Q Q Q restricted by 0 ≤ Q ≤ T 0\le Q\le T 0 ≤ Q ≤ T . Then c T = 1 + γ exp T + O ( T 2 ) c_T=1+\gamma_{\exp}T+O(T^2) c T = 1 + γ e x p T + O ( T 2 ) as T → 0 T\to0 T → 0 , where γ exp ( ρ ) = sup P τ ( ( P ρ P − ρ ) + ) / τ ( ρ ) \gamma_{\exp}(\rho)=\sup_P\tau((P\rho P-\rho)_+)/\tau(\rho) γ e x p ( ρ ) = sup P τ (( P ρP − ρ ) + ) / τ ( ρ ) , and γ exp > 0 \gamma_{\exp}>0 γ e x p > 0 for ρ ∉ Z \rho\notin\mathcal{Z} ρ ∈ / Z . For ρ λ \rho_\lambda ρ λ as in be:c , γ exp ( ρ λ ) = γ exp ( d i a g ( κ , 1 ) ) ( κ + 1 ) / ( κ + 1 + 2 λ ) \gamma_{\exp}(\rho_\lambda)=\gamma_{\exp}(\mathop{\mathrm{diag}}(\kappa,1))\,(\kappa+1)/(\kappa+1+2\lambda) γ e x p ( ρ λ ) = γ e x p ( diag ( κ , 1 )) ( κ + 1 ) / ( κ + 1 + 2 λ ) .
Proof. be:a Let E E E be a central projection. Lemma 5.46 in E M E\mathcal{M} E M , with the trace τ ( E ⋅ ) \tau(E\,\cdot) τ ( E ⋅ ) , gives τ ( E P e P Q P ) ≤ τ ( E e Q ) \tau(EPe^{PQP})\le\tau(Ee^Q) τ ( E P e P QP ) ≤ τ ( E e Q ) , and τ ( E e P Q P ) ≤ τ ( E e Q ) \tau(Ee^{PQP})\le\tau(Ee^Q) τ ( E e P QP ) ≤ τ ( E e Q ) if Q ≥ 0 Q\ge0 Q ≥ 0 . Approximating u = ⌈ ρ ⌉ u=\lceil \rho\rceil u = ⌈ ρ ⌉ by positive combinations of central projections, the same holds with u u u in place of E E E . Since P e P Q P = P e P Q P P ≥ 0 Pe^{PQP}=Pe^{PQP}P\ge0 P e P QP = P e P QP P ≥ 0 , (3) gives φ ( P e P Q P ) ≤ τ ( u P e P Q P ) ≤ τ ( u e Q ) ≤ κ Z τ ( l e Q ) ≤ κ Z φ ( e Q ) \varphi(Pe^{PQP})\le\tau(uPe^{PQP})\le\tau(ue^Q)\le\kappa_\mathcal{Z}\tau(le^Q)\le\kappa_\mathcal{Z}\varphi(e^Q) φ ( P e P QP ) ≤ τ ( u P e P QP ) ≤ τ ( u e Q ) ≤ κ Z τ ( l e Q ) ≤ κ Z φ ( e Q ) , and the same argument applies to e P Q P e^{PQP} e P QP . For s ≥ ∥ Q ∥ s\ge\|Q\| s ≥ ∥ Q ∥ we have Q + s ≥ 0 Q+s\ge0 Q + s ≥ 0 and e P ( Q + s ) P = ( 1 − P ) + e s P e P Q P e^{P(Q+s)P}=(1-P)+e^sPe^{PQP} e P ( Q + s ) P = ( 1 − P ) + e s P e P QP , so
φ ( e P ( Q + s ) P ) φ ( e Q + s ) = φ ( P e P Q P ) φ ( e Q ) + e − s φ ( 1 − P ) φ ( e Q ) , \begin{equation}
\frac{\varphi(e^{P(Q+s)P})}{\varphi(e^{Q+s})}=\frac{\varphi(Pe^{PQP})}{\varphi(e^Q)}+e^{-s}\,\frac{\varphi(1-P)}{\varphi(e^Q)}, \tag*{(10)}
\end{equation} φ ( e Q + s ) φ ( e P ( Q + s ) P ) = φ ( e Q ) φ ( P e P QP ) + e − s φ ( e Q ) φ ( 1 − P ) , (10) which gives b ≤ c exp b\le c_{\exp} b ≤ c e x p . Finally, for Q ≥ 0 Q\ge0 Q ≥ 0 we have e Q ≥ 1 e^Q\ge1 e Q ≥ 1 , so φ ( e P Q P ) = φ ( 1 − P ) + φ ( P e P Q P ) ≤ φ ( 1 ) + b φ ( e Q ) ≤ ( 1 + b ) φ ( e Q ) \varphi(e^{PQP})=\varphi(1-P)+\varphi(Pe^{PQP})\le\varphi(1)+b\,\varphi(e^Q)\le(1+b)\varphi(e^Q) φ ( e P QP ) = φ ( 1 − P ) + φ ( P e P QP ) ≤ φ ( 1 ) + b φ ( e Q ) ≤ ( 1 + b ) φ ( e Q ) .
be:b Since z z z commutes with P P P and Q Q Q , P ( Q + ln z ) P = P Q P + ( ln z ) P P(Q+\ln z)P=PQP+(\ln z)P P ( Q + ln z ) P = P QP + ( ln z ) P with commuting summands. Hence P e P ( Q + ln z ) P = z P e P Q P Pe^{P(Q+\ln z)P}=zPe^{PQP} P e P ( Q + l n z ) P = z P e P QP ; also e Q + ln z = z e Q e^{Q+\ln z}=ze^Q e Q + l n z = z e Q . So (R) holds with σ z ( P , Q ) = ( P , Q + ln z ) \sigma_z(P,Q)=(P,Q+\ln z) σ z ( P , Q ) = ( P , Q + ln z ) . For the lower bound let π \pi π , E π = π ( 1 ) E_\pi=\pi(1) E π = π ( 1 ) , t t t and ε \varepsilon ε be as in the proof of Theorem 4.15 t:low , so that φ ∘ π = t T r ( R ε ⋅ ) \varphi\circ\pi=t\mathop{\mathrm{Tr}}(R_\varepsilon\,\cdot) φ ∘ π = t Tr ( R ε ⋅ ) . For a projection p ∈ M 2 ( C ) p\in M_2(\mathbb{C}) p ∈ M 2 ( C ) and q ∈ M 2 ( C ) s a q\in M_2(\mathbb{C})_{sa} q ∈ M 2 ( C ) s a put P = π ( p ) P=\pi(p) P = π ( p ) and Q L = π ( q ) − L ( 1 − E π ) Q_L=\pi(q)-L(1-E_\pi) Q L = π ( q ) − L ( 1 − E π ) . Then P e P Q L P = π ( p e p q p ) Pe^{PQ_LP}=\pi(pe^{pqp}) P e P Q L P = π ( p e pq p ) and e Q L = π ( e q ) + e − L ( 1 − E π ) e^{Q_L}=\pi(e^q)+e^{-L}(1-E_\pi) e Q L = π ( e q ) + e − L ( 1 − E π ) ; let L → ∞ L\to\infty L → ∞ and then ε → 0 \varepsilon\to0 ε → 0 .
To see that b 2 ( κ ) > 1 b_2(\kappa)>1 b 2 ( κ ) > 1 , let ρ = d i a g ( κ , 1 ) \rho=\mathop{\mathrm{diag}}(\kappa,1) ρ = diag ( κ , 1 ) . Every pair ( P , Q ) (P,Q) ( P , Q ) in M 2 ( C ) M_2(\mathbb{C}) M 2 ( C ) with P P P of rank one is, up to a shift of Q Q Q , of the form P = ∣ w ⟩ ⟨ w ∣ P=|w\rangle\langle w| P = ∣ w ⟩ ⟨ w ∣ and Q = − μ ∣ v 2 ⟩ ⟨ v 2 ∣ Q=-\mu|v_2\rangle\langle v_2| Q = − μ ∣ v 2 ⟩ ⟨ v 2 ∣ , where μ ∈ R \mu\in\mathbb{R} μ ∈ R and v 1 , v 2 v_1,v_2 v 1 , v 2 is an orthonormal basis. Put β i = ⟨ ρ v i , v i ⟩ \beta_i=\langle\rho v_i,v_i\rangle β i = ⟨ ρ v i , v i ⟩ . With x = ∣ ⟨ w , v 2 ⟩ ∣ 2 x=|\langle w,v_2\rangle|^2 x = ∣ ⟨ w , v 2 ⟩ ∣ 2 the ratio is e − μ x ⟨ ρ w , w ⟩ / ( β 1 + e − μ β 2 ) e^{-\mu x}\langle\rho w,w\rangle/(\beta_1+e^{-\mu}\beta_2) e − μx ⟨ ρw , w ⟩ / ( β 1 + e − μ β 2 ) . Its supremum over μ \mu μ is
R ( w , v 1 ) = ⟨ ρ w , w ⟩ x x ( 1 − x ) 1 − x β 1 x − 1 β 2 − x , \begin{equation}
R(w,v_1)=\langle\rho w,w\rangle\,x^x(1-x)^{1-x}\beta_1^{x-1}\beta_2^{-x}, \tag*{(11)}
\end{equation} R ( w , v 1 ) = ⟨ ρw , w ⟩ x x ( 1 − x ) 1 − x β 1 x − 1 β 2 − x , (11) attained at e − μ = x β 1 / ( ( 1 − x ) β 2 ) e^{-\mu}=x\beta_1/((1-x)\beta_2) e − μ = x β 1 / (( 1 − x ) β 2 ) if 0 < x < 1 0<x<1 0 < x < 1 ; for x ∈ { 0 , 1 } x\in\{0,1\} x ∈ { 0 , 1 } , R = 1 R=1 R = 1 . Now choose v 1 v_1 v 1 not an eigenvector of ρ \rho ρ and w = 1 − x v 1 + x e i χ v 2 w=\sqrt{1-x}\,v_1+\sqrt x\,e^{i\chi}v_2 w = 1 − x v 1 + x e i χ v 2 with a suitable χ \chi χ . Then ⟨ ρ w , w ⟩ = β 1 + 2 x ( 1 − x ) ∣ ⟨ ρ v 1 , v 2 ⟩ ∣ + O ( x ) \langle\rho w,w\rangle=\beta_1+2\sqrt{x(1-x)}\,|\langle\rho v_1,v_2\rangle|+O(x) ⟨ ρw , w ⟩ = β 1 + 2 x ( 1 − x ) ∣ ⟨ ρ v 1 , v 2 ⟩ ∣ + O ( x ) , so R = 1 + 2 β 1 − 1 ∣ ⟨ ρ v 1 , v 2 ⟩ ∣ x + O ( x ∣ ln x ∣ ) > 1 R=1+2\beta_1^{-1}|\langle\rho v_1,v_2\rangle|\sqrt x+O(x|\ln x|)>1 R = 1 + 2 β 1 − 1 ∣ ⟨ ρ v 1 , v 2 ⟩ ∣ x + O ( x ∣ ln x ∣ ) > 1 for small x > 0 x>0 x > 0 .
To see that c exp p r ( ρ ) > 1 c^{\rm pr}_{\exp}(\rho)>1 c e x p pr ( ρ ) > 1 for ρ ∉ Z \rho\notin\mathcal{Z} ρ ∈ / Z , let first D ∈ M 2 ( C ) + + D\in M_2(\mathbb{C})_{++} D ∈ M 2 ( C ) ++ be non-scalar. Choose a unit vector v v v that is not an eigenvector of D D D , a unit vector v ′ ⊥ v v'\perp v v ′ ⊥ v , and put η 0 = ∣ ⟨ D v , v ′ ⟩ ∣ > 0 \eta_0=|\langle Dv,v'\rangle|>0 η 0 = ∣ ⟨ D v , v ′ ⟩ ∣ > 0 . For x ∈ ( 0 , 1 ) x\in(0,1) x ∈ ( 0 , 1 ) let w = 1 − x v + x e i χ v ′ w=\sqrt{1-x}\,v+\sqrt x\,e^{i\chi}v' w = 1 − x v + x e i χ v ′ with R e ( e i χ ⟨ D v ′ , v ⟩ ) = η 0 \mathop{\mathrm{Re}}(e^{i\chi}\langle Dv',v\rangle)=\eta_0 Re ( e i χ ⟨ D v ′ , v ⟩) = η 0 . Take p = ∣ w ⟩ ⟨ w ∣ p=|w\rangle\langle w| p = ∣ w ⟩ ⟨ w ∣ and q = ∣ v ⟩ ⟨ v ∣ q=|v\rangle\langle v| q = ∣ v ⟩ ⟨ v ∣ . Then p q p = ( 1 − x ) p pqp=(1-x)p pq p = ( 1 − x ) p , and
T r ( D e p q p ) − T r ( D e q ) = ( e 1 − x − 1 ) ⟨ D w , w ⟩ − ( e − 1 ) ⟨ D v , v ⟩ = 2 ( e − 1 ) η 0 x + O ( x ) , \mathop{\mathrm{Tr}}(De^{pqp})-\mathop{\mathrm{Tr}}(De^q)=(e^{1-x}-1)\langle Dw,w\rangle-(e-1)\langle Dv,v\rangle=2(e-1)\eta_0\sqrt x+O(x), Tr ( D e pq p ) − Tr ( D e q ) = ( e 1 − x − 1 ) ⟨ D w , w ⟩ − ( e − 1 ) ⟨ D v , v ⟩ = 2 ( e − 1 ) η 0 x + O ( x ) , which is > 0 >0 > 0 for small x x x . Apply this to D = d i a g ( α , β ′ ) D=\mathop{\mathrm{diag}}(\alpha,\beta') D = diag ( α , β ′ ) , where α = ⌈ ρ ⌉ ( ω 0 ) \alpha=\lceil \rho\rceil(\omega_0) α = ⌈ ρ ⌉ ( ω 0 ) and β ′ = ⌊ ρ ⌋ ( ω 0 ) \beta'=\lfloor \rho\rfloor(\omega_0) β ′ = ⌊ ρ ⌋ ( ω 0 ) are as in Lemma 4.14 , and fix such p , q p,q p , q . By continuity T r ( R ε e p q p ) > T r ( R ε e q ) \mathop{\mathrm{Tr}}(R_\varepsilon e^{pqp})>\mathop{\mathrm{Tr}}(R_\varepsilon e^q) Tr ( R ε e pq p ) > Tr ( R ε e q ) for small ε \varepsilon ε . For P = π ( p ) P=\pi(p) P = π ( p ) and Q = π ( q ) Q=\pi(q) Q = π ( q ) we have e P Q P = π ( e p q p ) + ( 1 − E π ) e^{PQP}=\pi(e^{pqp})+(1-E_\pi) e P QP = π ( e pq p ) + ( 1 − E π ) and e Q = π ( e q ) + ( 1 − E π ) e^Q=\pi(e^q)+(1-E_\pi) e Q = π ( e q ) + ( 1 − E π ) , so
φ ( e P Q P ) φ ( e Q ) = t T r ( R ε e p q p ) + φ ( 1 − E π ) t T r ( R ε e q ) + φ ( 1 − E π ) > 1. \frac{\varphi(e^{PQP})}{\varphi(e^Q)}=\frac{t\mathop{\mathrm{Tr}}(R_\varepsilon e^{pqp})+\varphi(1-E_\pi)}{t\mathop{\mathrm{Tr}}(R_\varepsilon e^q)+\varphi(1-E_\pi)}>1. φ ( e Q ) φ ( e P QP ) = t Tr ( R ε e q ) + φ ( 1 − E π ) t Tr ( R ε e pq p ) + φ ( 1 − E π ) > 1. be:c κ Z ( ρ λ ) = κ \kappa_\mathcal{Z}(\rho_\lambda)=\kappa κ Z ( ρ λ ) = κ by Proposition 4.5 ch:prop , and φ = ( T r ⊕ T r ) ( ρ λ ⋅ ) \varphi=(\mathop{\mathrm{Tr}}\oplus\mathop{\mathrm{Tr}})(\rho_\lambda\,\cdot) φ = ( Tr ⊕ Tr ) ( ρ λ ⋅ ) . For P = P 1 ⊕ P 2 P=P_1\oplus P_2 P = P 1 ⊕ P 2 and Q = Q 1 ⊕ Q 2 Q=Q_1\oplus Q_2 Q = Q 1 ⊕ Q 2 write the ratio as ( N 1 + λ N 2 ) / ( D 1 + λ D 2 ) (N_1+\lambda N_2)/(D_1+\lambda D_2) ( N 1 + λ N 2 ) / ( D 1 + λ D 2 ) . Then N 2 ≤ D 2 N_2\le D_2 N 2 ≤ D 2 by Lemma 5.46 , and D 2 = T r e Q 2 ≥ 2 D_2=\mathop{\mathrm{Tr}}e^{Q_2}\ge2 D 2 = Tr e Q 2 ≥ 2 . The ratio is therefore at most max { N 1 / D 1 , 1 } \max\{N_1/D_1,1\} max { N 1 / D 1 , 1 } . On the other hand, taking P 2 = Q 2 = 0 P_2=Q_2=0 P 2 = Q 2 = 0 gives ( N 1 + 2 λ ) / ( D 1 + 2 λ ) → N 1 / D 1 (N_1+2\lambda)/(D_1+2\lambda)\to N_1/D_1 ( N 1 + 2 λ ) / ( D 1 + 2 λ ) → N 1 / D 1 as λ → 0 \lambda\to0 λ → 0 . This proves the limits as λ → 0 \lambda\to0 λ → 0 . As λ → ∞ \lambda\to\infty λ → ∞ , for Q ≥ 0 Q\ge0 Q ≥ 0 we have N 1 ≤ φ 1 ( 1 ) + b 2 ( κ ) D 1 N_1\le\varphi_1(1)+b_2(\kappa)D_1 N 1 ≤ φ 1 ( 1 ) + b 2 ( κ ) D 1 by be:a in M 2 ( C ) M_2(\mathbb{C}) M 2 ( C ) . Since b 2 ( κ ) > 1 b_2(\kappa)>1 b 2 ( κ ) > 1 , the ratio is at most b 2 ( κ ) + ( κ + 1 ) / ( 2 λ ) b_2(\kappa)+(\kappa+1)/(2\lambda) b 2 ( κ ) + ( κ + 1 ) / ( 2 λ ) ; and c exp ( ρ λ ) ≥ b ( ρ λ ) ≥ b 2 ( κ ) c_{\exp}(\rho_\lambda)\ge b(\rho_\lambda)\ge b_2(\kappa) c e x p ( ρ λ ) ≥ b ( ρ λ ) ≥ b 2 ( κ ) by be:a and be:b . For projections N 1 ≤ e φ 1 ( 1 ) N_1\le e\,\varphi_1(1) N 1 ≤ e φ 1 ( 1 ) , so the ratio is at most 1 + e ( κ + 1 ) / ( 2 λ ) 1+e(\kappa+1)/(2\lambda) 1 + e ( κ + 1 ) / ( 2 λ ) , while P = 1 P=1 P = 1 gives the value 1 1 1 . It remains to compare the limits. First, c exp p r ( d i a g ( κ , 1 ) ) > 1 c^{\rm pr}_{\exp}(\mathop{\mathrm{diag}}(\kappa,1))>1 c e x p pr ( diag ( κ , 1 )) > 1 by be:b . Second, R R R in (11) is continuous on the compact set of pairs of unit vectors ( w , v 1 ) (w,v_1) ( w , v 1 ) , and rank-two P P P gives the value 1 < b 2 ( κ ) 1<b_2(\kappa) 1 < b 2 ( κ ) . So b 2 ( κ ) b_2(\kappa) b 2 ( κ ) is attained at some ( P , Q ) (P,Q) ( P , Q ) with 0 ≠ P ≠ 1 0\ne P\ne1 0 = P = 1 , and then (10) gives c exp ( d i a g ( κ , 1 ) ) > b 2 ( κ ) c_{\exp}(\mathop{\mathrm{diag}}(\kappa,1))>b_2(\kappa) c e x p ( diag ( κ , 1 )) > b 2 ( κ ) .
be:d For 0 ≤ Q ≤ T 0\le Q\le T 0 ≤ Q ≤ T , e P Q P = 1 + P Q P + O ( T 2 ) e^{PQP}=1+PQP+O(T^2) e P QP = 1 + P QP + O ( T 2 ) and e Q = 1 + Q + O ( T 2 ) e^Q=1+Q+O(T^2) e Q = 1 + Q + O ( T 2 ) uniformly in P P P . So the ratio is 1 + τ ( Q ( P ρ P − ρ ) ) / τ ( ρ ) + O ( T 2 ) 1+\tau(Q(P\rho P-\rho))/\tau(\rho)+O(T^2) 1 + τ ( Q ( P ρP − ρ )) / τ ( ρ ) + O ( T 2 ) , and sup 0 ≤ Q ≤ T τ ( Q ( P ρ P − ρ ) ) = T τ ( ( P ρ P − ρ ) + ) \sup_{0\le Q\le T}\tau(Q(P\rho P-\rho))=T\,\tau((P\rho P-\rho)_+) sup 0 ≤ Q ≤ T τ ( Q ( P ρP − ρ )) = T τ (( P ρP − ρ ) + ) . If ρ ∉ Z \rho\notin\mathcal{Z} ρ ∈ / Z , then P ρ P ≰ ρ P\rho P\not\le\rho P ρP ≤ ρ for some P P P by Theorem 7.2 , so γ exp > 0 \gamma_{\exp}>0 γ e x p > 0 . For ρ λ \rho_\lambda ρ λ and P = P 1 ⊕ P 2 P=P_1\oplus P_2 P = P 1 ⊕ P 2 we have P ρ λ P − ρ λ = ( P 1 ρ 1 P 1 − ρ 1 ) ⊕ λ ( P 2 − 1 ) P\rho_\lambda P-\rho_\lambda=(P_1\rho_1P_1-\rho_1)\oplus\lambda(P_2-1) P ρ λ P − ρ λ = ( P 1 ρ 1 P 1 − ρ 1 ) ⊕ λ ( P 2 − 1 ) with ρ 1 = d i a g ( κ , 1 ) \rho_1=\mathop{\mathrm{diag}}(\kappa,1) ρ 1 = diag ( κ , 1 ) . The second summand is ≤ 0 \le0 ≤ 0 , and τ ( ρ λ ) = κ + 1 + 2 λ \tau(\rho_\lambda)=\kappa+1+2\lambda τ ( ρ λ ) = κ + 1 + 2 λ , which gives the formula for γ exp ( ρ λ ) \gamma_{\exp}(\rho_\lambda) γ e x p ( ρ λ ) . ◻
By be:d , even the slope γ exp \gamma_{\exp} γ e x p depends on the weights of the central summands, not only on κ Z \kappa_\mathcal{Z} κ Z .
The Peierls–Bogoliubov and Araki–Lieb–Thirring inequalities
Bikchentaev [98 ] characterised tracial functionals by two inequalities. The first is the Peierls–Bogoliubov inequality
φ ( e A + B ) ≥ φ ( e A ) exp ( φ ( B e A ) / φ ( e A ) ) \varphi(e^{A+B})\ge\varphi(e^A)\exp\bigl(\varphi(Be^A)/\varphi(e^A)\bigr) φ ( e A + B ) ≥ φ ( e A ) exp ( φ ( B e A ) / φ ( e A ) )
for self-adjoint A , B A,B A , B . The second is the Araki–Lieb–Thirring inequality
φ ( ( B 1 / 2 A B 1 / 2 ) r q ) ≤ φ ( ( B r / 2 A r B r / 2 ) q ) \varphi((B^{1/2}AB^{1/2})^{rq})\le\varphi((B^{r/2}A^rB^{r/2})^q) φ (( B 1/2 A B 1/2 ) r q ) ≤ φ (( B r /2 A r B r /2 ) q )
for A , B ≥ 0 A,B\ge0 A , B ≥ 0 , with exponents r > 1 r>1 r > 1 , q > 0 q>0 q > 0 (see [98 ] for the precise quantifiers). For non-tracial φ \varphi φ the number φ ( B e A ) \varphi(Be^A) φ ( B e A ) need not be real; we use its real part. With ∣ φ ( B e A ) ∣ |\varphi(Be^A)| ∣ φ ( B e A ) ∣ instead, pb:a below remains true, since ∣ φ ( B e A ) ∣ ≥ R e φ ( B e A ) |\varphi(Be^A)|\ge\mathop{\mathrm{Re}}\varphi(Be^A) ∣ φ ( B e A ) ∣ ≥ Re φ ( B e A ) . We also consider the symmetric form with ω A ( B ) = φ ( e A / 2 B e A / 2 ) / φ ( e A ) \omega_A(B)=\varphi(e^{A/2}Be^{A/2})/\varphi(e^A) ω A ( B ) = φ ( e A /2 B e A /2 ) / φ ( e A ) , which coincides with the first form for traces. Put
c P B ( ρ ) = sup A , B ∈ M s a φ ( e A ) exp ( ω A ( B ) ) φ ( e A + B ) , c r , q ( ρ ) = sup A , B ∈ M + φ ( ( B 1 / 2 A B 1 / 2 ) r q ) φ ( ( B r / 2 A r B r / 2 ) q ) . c_{\rm PB}(\rho)=\sup_{A,B\in\mathcal{M}_{sa}}\frac{\varphi(e^A)\exp(\omega_A(B))}{\varphi(e^{A+B})},\qquad
c_{r,q}(\rho)=\sup_{A,B\in\mathcal{M}_+}\frac{\varphi((B^{1/2}AB^{1/2})^{rq})}{\varphi((B^{r/2}A^rB^{r/2})^q)}. c PB ( ρ ) = A , B ∈ M s a sup φ ( e A + B ) φ ( e A ) exp ( ω A ( B )) , c r , q ( ρ ) = A , B ∈ M + sup φ (( B r /2 A r B r /2 ) q ) φ (( B 1/2 A B 1/2 ) r q ) .
Proposition 5.50 (Items 16 and 37).
For ρ ∈ Z \rho\in\mathcal{Z} ρ ∈ Z the supremum below (and c P B c_{\rm PB} c PB ) equals 1 1 1 : φ \varphi φ is then a trace, so the Peierls–Bogoliubov inequality holds [98 ] (and φ ( B e A ) \varphi(Be^A) φ ( B e A ) is real), with equality at B = 0 B=0 B = 0 . If ρ ∉ Z \rho\notin\mathcal{Z} ρ ∈ / Z , then sup φ ( e A ) exp ( R e φ ( B e A ) / φ ( e A ) ) / φ ( e A + B ) = + ∞ \sup\varphi(e^A)\exp(\mathop{\mathrm{Re}}\varphi(Be^A)/\varphi(e^A))/\varphi(e^{A+B})=+\infty sup φ ( e A ) exp ( Re φ ( B e A ) / φ ( e A )) / φ ( e A + B ) = + ∞ , already over A ∈ M s a A\in\mathcal{M}_{sa} A ∈ M s a and B ≤ 0 B\le0 B ≤ 0 .
c P B ( z ρ ) = c P B ( ρ ) c_{\rm PB}(z\rho)=c_{\rm PB}(\rho) c PB ( z ρ ) = c PB ( ρ ) for z ∈ Z + + z\in\mathcal{Z}_{++} z ∈ Z ++ , and b ( ρ ) ≤ c P B ( ρ ) ≤ κ Z ( ρ ) b(\rho)\le c_{\rm PB}(\rho)\le\kappa_\mathcal{Z}(\rho) b ( ρ ) ≤ c PB ( ρ ) ≤ κ Z ( ρ ) . In particular c P B ( ρ ) > 1 c_{\rm PB}(\rho)>1 c PB ( ρ ) > 1 for ρ ∉ Z \rho\notin\mathcal{Z} ρ ∈ / Z .
Proof. pb:a For Y ∈ M + + Y\in\mathcal{M}_{++} Y ∈ M ++ and B ∈ M s a B\in\mathcal{M}_{sa} B ∈ M s a , R e φ ( B Y ) = τ ( B ( ρ ∘ Y ) ) \mathop{\mathrm{Re}}\varphi(BY)=\tau(B(\rho\circ Y)) Re φ ( B Y ) = τ ( B ( ρ ∘ Y )) with ρ ∘ Y = 1 2 ( ρ Y + Y ρ ) \rho\circ Y=\frac12(\rho Y+Y\rho) ρ ∘ Y = 2 1 ( ρ Y + Y ρ ) . We first find Y ∈ M + + Y\in\mathcal{M}_{++} Y ∈ M ++ with ρ ∘ Y ≱ 0 \rho\circ Y\not\ge0 ρ ∘ Y ≥ 0 . Take η 1 , η 2 \eta_1,\eta_2 η 1 , η 2 and π \pi π from Lemma 4.14 c:b for a = ρ a=\rho a = ρ . The compression of ρ \rho ρ to K = span { η 1 , η 2 } \mathcal K=\operatorname{span}\{\eta_1,\eta_2\} K = span { η 1 , η 2 } is D = d i a g ( a 1 , a 2 ) D=\mathop{\mathrm{diag}}(a_1,a_2) D = diag ( a 1 , a 2 ) with a 1 ≠ a 2 a_1\ne a_2 a 1 = a 2 . Let g = ( 1 , 1 ) / 2 g=(1,1)/\sqrt2 g = ( 1 , 1 ) / 2 . Since g g g and D g Dg D g are linearly independent, there is c ∈ C 2 c\in\mathbb{C}^2 c ∈ C 2 with ⟨ c , g ⟩ = 1 \langle c,g\rangle=1 ⟨ c , g ⟩ = 1 and ⟨ c , D g ⟩ = − 1 \langle c,Dg\rangle=-1 ⟨ c , D g ⟩ = − 1 . For ε > 0 \varepsilon>0 ε > 0 put Y = π ( ∣ g ⟩ ⟨ g ∣ + ε 1 ) + ε ( 1 − E π ) ∈ M + + Y=\pi(|g\rangle\langle g|+\varepsilon1)+\varepsilon(1-E_\pi)\in\mathcal{M}_{++} Y = π ( ∣ g ⟩ ⟨ g ∣ + ε 1 ) + ε ( 1 − E π ) ∈ M ++ , where E π = π ( 1 ) E_\pi=\pi(1) E π = π ( 1 ) , and ξ = c 1 η 1 + c 2 η 2 \xi=c_1\eta_1+c_2\eta_2 ξ = c 1 η 1 + c 2 η 2 . Since Y ξ ∈ K Y\xi\in\mathcal K Y ξ ∈ K , ⟨ ( ρ ∘ Y ) ξ , ξ ⟩ = R e ⟨ Y ξ , ρ ξ ⟩ = R e ⟨ y c , D c ⟩ \langle(\rho\circ Y)\xi,\xi\rangle=\mathop{\mathrm{Re}}\langle Y\xi,\rho\xi\rangle=\mathop{\mathrm{Re}}\langle yc,Dc\rangle ⟨( ρ ∘ Y ) ξ , ξ ⟩ = Re ⟨ Y ξ , ρ ξ ⟩ = Re ⟨ y c , D c ⟩ with y = ∣ g ⟩ ⟨ g ∣ + ε 1 y=|g\rangle\langle g|+\varepsilon1 y = ∣ g ⟩ ⟨ g ∣ + ε 1 , and this equals − 1 + O ( ε ) < 0 -1+O(\varepsilon)<0 − 1 + O ( ε ) < 0 . Now let F ≠ 0 F\ne0 F = 0 be the spectral projection of ρ ∘ Y \rho\circ Y ρ ∘ Y for ( − ∞ , 0 ) (-\infty,0) ( − ∞ , 0 ) , ϑ = − τ ( F ( ρ ∘ Y ) ) > 0 \vartheta=-\tau(F(\rho\circ Y))>0 ϑ = − τ ( F ( ρ ∘ Y )) > 0 , A = ln Y A=\ln Y A = ln Y and B s = − s F B_s=-sF B s = − s F with s > 0 s>0 s > 0 . Then R e φ ( B s Y ) = s ϑ \mathop{\mathrm{Re}}\varphi(B_sY)=s\vartheta Re φ ( B s Y ) = s ϑ . Since A + B s ≤ A A+B_s\le A A + B s ≤ A , the monotonicity of τ ∘ exp \tau\circ\exp τ ∘ exp [121 ] gives φ ( e A + B s ) ≤ ∥ ρ ∥ τ ( e A + B s ) ≤ ∥ ρ ∥ τ ( e A ) \varphi(e^{A+B_s})\le\|\rho\|\tau(e^{A+B_s})\le\|\rho\|\tau(e^A) φ ( e A + B s ) ≤ ∥ ρ ∥ τ ( e A + B s ) ≤ ∥ ρ ∥ τ ( e A ) . Hence the ratio is at least φ ( Y ) e s ϑ / φ ( Y ) / ( ∥ ρ ∥ τ ( Y ) ) → ∞ \varphi(Y)e^{s\vartheta/\varphi(Y)}/(\|\rho\|\tau(Y))\to\infty φ ( Y ) e s ϑ / φ ( Y ) / ( ∥ ρ ∥ τ ( Y )) → ∞ as s → ∞ s\to\infty s → ∞ .
pb:b (R) holds with σ z ( A , B ) = ( A + ln z , B ) \sigma_z(A,B)=(A+\ln z,B) σ z ( A , B ) = ( A + ln z , B ) , since
z e A = e A + ln z , z e A / 2 B e A / 2 = e ( A + ln z ) / 2 B e ( A + ln z ) / 2 , z e A + B = e A + ln z + B . ze^A=e^{A+\ln z},\qquad ze^{A/2}Be^{A/2}=e^{(A+\ln z)/2}Be^{(A+\ln z)/2},\qquad ze^{A+B}=e^{A+\ln z+B}. z e A = e A + l n z , z e A /2 B e A /2 = e ( A + l n z ) /2 B e ( A + l n z ) /2 , z e A + B = e A + l n z + B . For the upper bound note that the ratio is unchanged under A ↦ A + t A\mapsto A+t A ↦ A + t , so we may assume φ ( e A ) = 1 \varphi(e^A)=1 φ ( e A ) = 1 . Put Y = e A Y=e^A Y = e A , σ = Y 1 / 2 ρ Y 1 / 2 \sigma=Y^{1/2}\rho Y^{1/2} σ = Y 1/2 ρ Y 1/2 and H = A + B H=A+B H = A + B . Then τ ( σ ) = 1 \tau(\sigma)=1 τ ( σ ) = 1 and ω A ( B ) = τ ( σ B ) = τ ( σ H ) − τ ( σ ln Y ) \omega_A(B)=\tau(\sigma B)=\tau(\sigma H)-\tau(\sigma\ln Y) ω A ( B ) = τ ( σ B ) = τ ( σ H ) − τ ( σ ln Y ) . Also σ ≤ ∥ ρ ∥ Y \sigma\le\|\rho\|Y σ ≤ ∥ ρ ∥ Y , so ln σ ≤ ln ∥ ρ ∥ + ln Y \ln\sigma\le\ln\|\rho\|+\ln Y ln σ ≤ ln ∥ ρ ∥ + ln Y by the operator monotonicity of ln \ln ln . The Peierls–Bogoliubov inequality for τ \tau τ [98 ] , applied to ln σ \ln\sigma ln σ and H − ln σ H-\ln\sigma H − ln σ , gives ln τ ( e H ) ≥ τ ( σ H ) − τ ( σ ln σ ) \ln\tau(e^H)\ge\tau(\sigma H)-\tau(\sigma\ln\sigma) ln τ ( e H ) ≥ τ ( σ H ) − τ ( σ ln σ ) . Moreover φ ( e H ) ≥ ∥ ρ − 1 ∥ − 1 τ ( e H ) \varphi(e^H)\ge\|\rho^{-1}\|^{-1}\tau(e^H) φ ( e H ) ≥ ∥ ρ − 1 ∥ − 1 τ ( e H ) . Hence the logarithm of the ratio is at most τ ( σ ln σ ) − τ ( σ ln Y ) + ln ∥ ρ − 1 ∥ ≤ ln κ ( ρ ) \tau(\sigma\ln\sigma)-\tau(\sigma\ln Y)+\ln\|\rho^{-1}\|\le\ln\kappa(\rho) τ ( σ ln σ ) − τ ( σ ln Y ) + ln ∥ ρ − 1 ∥ ≤ ln κ ( ρ ) , and Lemma 4.12 replaces κ \kappa κ by κ Z \kappa_\mathcal{Z} κ Z . For the lower bound let P ≠ 0 P\ne0 P = 0 be a projection, Q ∈ M s a Q\in\mathcal{M}_{sa} Q ∈ M s a , A L = P Q P − L ( 1 − P ) A_L=PQP-L(1-P) A L = P QP − L ( 1 − P ) and B L = Q − A L B_L=Q-A_L B L = Q − A L . Then e A L + B L = e Q e^{A_L+B_L}=e^Q e A L + B L = e Q and e A L = P e P Q P + e − L ( 1 − P ) e^{A_L}=Pe^{PQP}+e^{-L}(1-P) e A L = P e P QP + e − L ( 1 − P ) . Also ω A L ( B L ) → 0 \omega_{A_L}(B_L)\to0 ω A L ( B L ) → 0 as L → ∞ L\to\infty L → ∞ : in e A L / 2 ( Q − P Q P + L ( 1 − P ) ) e A L / 2 e^{A_L/2}(Q-PQP+L(1-P))e^{A_L/2} e A L /2 ( Q − P QP + L ( 1 − P )) e A L /2 the corner P ⋅ P P\,\cdot\,P P ⋅ P vanishes and the other terms are O ( L e − L / 2 ) O(Le^{-L/2}) O ( L e − L /2 ) . So the ratio tends to φ ( P e P Q P ) / φ ( e Q ) \varphi(Pe^{PQP})/\varphi(e^Q) φ ( P e P QP ) / φ ( e Q ) . Finally b > 1 b>1 b > 1 for ρ ∉ Z \rho\notin\mathcal{Z} ρ ∈ / Z by Proposition 5.47 be:b . ◻
Proposition 5.51 (Items 7 and 38). Let r ≥ 1 r\ge1 r ≥ 1 and q > 0 q>0 q > 0 . Then c r , q ( z ρ ) = c r , q ( ρ ) c_{r,q}(z\rho)=c_{r,q}(\rho) c r , q ( z ρ ) = c r , q ( ρ ) for z ∈ Z + + z\in\mathcal{Z}_{++} z ∈ Z ++ , and c r , q ( ρ ) ≤ κ Z c_{r,q}(\rho)\le\kappa_\mathcal{Z} c r , q ( ρ ) ≤ κ Z . If r > 1 r>1 r > 1 and ρ ∉ Z \rho\notin\mathcal{Z} ρ ∈ / Z , then c r , q ( ρ ) > 1 c_{r,q}(\rho)>1 c r , q ( ρ ) > 1 for every q > 0 q>0 q > 0 , and sup q > 0 c r , q ( ρ ) = κ Z \sup_{q>0}c_{r,q}(\rho)=\kappa_\mathcal{Z} sup q > 0 c r , q ( ρ ) = κ Z .
Proof. (R) holds with σ z ( A , B ) = ( z 1 / ( r q ) A , B ) \sigma_z(A,B)=(z^{1/(rq)}A,B) σ z ( A , B ) = ( z 1/ ( r q ) A , B ) . The upper bound κ \kappa κ follows from Kosaki’s inequality τ ( ( B 1 / 2 A B 1 / 2 ) r q ) ≤ τ ( ( B r / 2 A r B r / 2 ) q ) \tau((B^{1/2}AB^{1/2})^{rq})\le\tau((B^{r/2}A^rB^{r/2})^q) τ (( B 1/2 A B 1/2 ) r q ) ≤ τ (( B r /2 A r B r /2 ) q ) , valid for all r ≥ 1 r\ge1 r ≥ 1 and q > 0 q>0 q > 0 [142 ] , together with ∥ ρ − 1 ∥ − 1 τ ≤ φ ≤ ∥ ρ ∥ τ \|\rho^{-1}\|^{-1}\tau\le\varphi\le\|\rho\|\tau ∥ ρ − 1 ∥ − 1 τ ≤ φ ≤ ∥ ρ ∥ τ on M + \mathcal{M}_+ M + ; then apply Lemma 4.12 . All expressions are admissible, so Theorem 4.15 t:low reduces the remaining claims to M 2 ( C ) M_2(\mathbb{C}) M 2 ( C ) with ρ = d i a g ( κ , 1 ) \rho=\mathop{\mathrm{diag}}(\kappa,1) ρ = diag ( κ , 1 ) , κ > 1 \kappa>1 κ > 1 . For a unit vector a a a , A = ∣ a ⟩ ⟨ a ∣ A=|a\rangle\langle a| A = ∣ a ⟩ ⟨ a ∣ and B > 0 B>0 B > 0 , we have ( B 1 / 2 A B 1 / 2 ) r q = ⟨ B a , a ⟩ r q P u (B^{1/2}AB^{1/2})^{rq}=\langle Ba,a\rangle^{rq}P_u ( B 1/2 A B 1/2 ) r q = ⟨ B a , a ⟩ r q P u and ( B r / 2 A r B r / 2 ) q = ⟨ B r a , a ⟩ q P v (B^{r/2}A^rB^{r/2})^q=\langle B^ra,a\rangle^qP_v ( B r /2 A r B r /2 ) q = ⟨ B r a , a ⟩ q P v , where P u , P v P_u,P_v P u , P v are the projections onto u = B 1 / 2 a u=B^{1/2}a u = B 1/2 a and v = B r / 2 a v=B^{r/2}a v = B r /2 a . So, with g ( x ) = ⟨ ρ x , x ⟩ / ∥ x ∥ 2 g(x)=\langle\rho x,x\rangle/\|x\|^2 g ( x ) = ⟨ ρ x , x ⟩ /∥ x ∥ 2 , the ratio is ( ⟨ B a , a ⟩ r / ⟨ B r a , a ⟩ ) q g ( u ) / g ( v ) (\langle Ba,a\rangle^r/\langle B^ra,a\rangle)^q\,g(u)/g(v) (⟨ B a , a ⟩ r / ⟨ B r a , a ⟩ ) q g ( u ) / g ( v ) . To show c r , q > 1 c_{r,q}>1 c r , q > 1 , let a a a be a unit vector that is not an eigenvector of ρ \rho ρ , d = ( ρ − g ( a ) ) a ≠ 0 d=(\rho-g(a))a\ne0 d = ( ρ − g ( a )) a = 0 (so d ⊥ a d\perp a d ⊥ a ), H = − ( ∣ d ⟩ ⟨ a ∣ + ∣ a ⟩ ⟨ d ∣ ) H=-(|d\rangle\langle a|+|a\rangle\langle d|) H = − ( ∣ d ⟩ ⟨ a ∣ + ∣ a ⟩ ⟨ d ∣ ) and B = 1 + ϵ H B=1+\epsilon H B = 1 + ϵH with small ϵ > 0 \epsilon>0 ϵ > 0 . Since H a = − d ⊥ a Ha=-d\perp a H a = − d ⊥ a , ⟨ B a , a ⟩ = 1 \langle Ba,a\rangle=1 ⟨ B a , a ⟩ = 1 and ⟨ B r a , a ⟩ = 1 + O ( ϵ 2 ) \langle B^ra,a\rangle=1+O(\epsilon^2) ⟨ B r a , a ⟩ = 1 + O ( ϵ 2 ) . Moreover u = a − ϵ 2 d + O ( ϵ 2 ) u=a-\frac\epsilon2d+O(\epsilon^2) u = a − 2 ϵ d + O ( ϵ 2 ) and v = a − r ϵ 2 d + O ( ϵ 2 ) v=a-\frac{r\epsilon}2d+O(\epsilon^2) v = a − 2 r ϵ d + O ( ϵ 2 ) , so g ( u ) = g ( a ) − ϵ ∥ d ∥ 2 + O ( ϵ 2 ) g(u)=g(a)-\epsilon\|d\|^2+O(\epsilon^2) g ( u ) = g ( a ) − ϵ ∥ d ∥ 2 + O ( ϵ 2 ) and g ( v ) = g ( a ) − r ϵ ∥ d ∥ 2 + O ( ϵ 2 ) g(v)=g(a)-r\epsilon\|d\|^2+O(\epsilon^2) g ( v ) = g ( a ) − r ϵ ∥ d ∥ 2 + O ( ϵ 2 ) . Hence the ratio is 1 + ( r − 1 ) ϵ ∥ d ∥ 2 / g ( a ) + O ( ϵ 2 ) > 1 1+(r-1)\epsilon\|d\|^2/g(a)+O(\epsilon^2)>1 1 + ( r − 1 ) ϵ ∥ d ∥ 2 / g ( a ) + O ( ϵ 2 ) > 1 for small ϵ \epsilon ϵ . For the supremum over q q q let A = ∣ a ⟩ ⟨ a ∣ A=|a\rangle\langle a| A = ∣ a ⟩ ⟨ a ∣ with ∥ a ∥ = 1 \|a\|=1 ∥ a ∥ = 1 , and B = d i a g ( 1 , γ ′ ) B=\mathop{\mathrm{diag}}(1,\gamma') B = diag ( 1 , γ ′ ) with γ ′ > 0 \gamma'>0 γ ′ > 0 . As q → 0 q\to0 q → 0 the ratio tends to g ( u ) / g ( v ) g(u)/g(v) g ( u ) / g ( v ) . For a ∝ ( 1 , μ γ ′ − 1 / 2 ) a\propto(1,\mu\gamma'^{-1/2}) a ∝ ( 1 , μ γ ′ − 1/2 ) we get u ∝ ( 1 , μ ) u\propto(1,\mu) u ∝ ( 1 , μ ) and v ∝ ( 1 , μ γ ′ ( r − 1 ) / 2 ) v\propto(1,\mu\gamma'^{(r-1)/2}) v ∝ ( 1 , μ γ ′ ( r − 1 ) /2 ) . Letting μ → 0 \mu\to0 μ → 0 and γ ′ → ∞ \gamma'\to\infty γ ′ → ∞ with μ γ ′ ( r − 1 ) / 2 → ∞ \mu\gamma'^{(r-1)/2}\to\infty μ γ ′ ( r − 1 ) /2 → ∞ , g ( u ) / g ( v ) g(u)/g(v) g ( u ) / g ( v ) tends to κ \kappa κ . ◻
Item 16 is the symmetric form, with the normalisation N = R \mathcal{N}=\mathcal{R} N = R , so V ( ρ ) = c P B ( ρ ) − 1 V(\rho)=c_{\rm PB}(\rho)-1 V ( ρ ) = c PB ( ρ ) − 1 . On M 2 M_2 M 2 the computed envelope of item 16 (Table 5 ) agrees with b 2 ( κ ) − 1 b_2(\kappa)-1 b 2 ( κ ) − 1 to within 3 ⋅ 10 − 3 3\cdot10^{-3} 3 ⋅ 1 0 − 3 at δ = 0.1 \delta=0.1 δ = 0.1 , 0.5 0.5 0.5 and 0.9 0.9 0.9 and lies below it; together with Remark 5.52 this supports c P B = b = b 2 ( κ Z ) c_{\rm PB}=b=b_2(\kappa_\mathcal{Z}) c PB = b = b 2 ( κ Z ) . In the search the best points have large ∥ A ∥ \|A\| ∥ A ∥ and ∥ B ∥ \|B\| ∥ B ∥ , as in the lower bound construction of Proposition 5.50 pb:b , where L → ∞ L\to\infty L → ∞ .
Proposition (Item 7). G Z ( ρ ) 2 ≤ c 2 , 1 ( ρ ) ≤ κ Z ( ρ ) G_\mathcal{Z}(\rho)^2\le c_{2,1}(\rho)\le\kappa_\mathcal{Z}(\rho) G Z ( ρ ) 2 ≤ c 2 , 1 ( ρ ) ≤ κ Z ( ρ ) in every finite von Neumann algebra. In M n ( C ) M_n(\mathbb{C}) M n ( C ) , G 2 − 1 ≤ V n ≤ κ − 1 G^2-1\le V_n\le\kappa-1 G 2 − 1 ≤ V n ≤ κ − 1 for item 7, and the value G 2 G^2 G 2 of the quotient is approached by the explicit family in the proof.
Proof. The upper bound is Proposition 5.51 . For the lower bound let ρ = d i a g ( a , b ) \rho=\mathop{\mathrm{diag}}(a,b) ρ = diag ( a , b ) in M 2 M_2 M 2 and, in the basis f 1 = e + f_1=e_+ f 1 = e + , f 2 = e − f_2=e_- f 2 = e − , write ρ 11 = ρ 22 = a + b 2 \rho_{11}=\rho_{22}=\frac{a+b}2 ρ 11 = ρ 22 = 2 a + b , ρ 12 = a − b 2 \rho_{12}=\frac{a-b}2 ρ 12 = 2 a − b for its entries. Take B = η f 1 f 1 ∗ + f 2 f 2 ∗ B=\eta f_1f_1^*+f_2f_2^* B = η f 1 f 1 ∗ + f 2 f 2 ∗ and A = u u ∗ A=uu^* A = u u ∗ with u = f 1 + s η f 2 u=f_1+s\eta f_2 u = f 1 + sη f 2 . Then B 1 / 2 A B 1 / 2 = w w ∗ B^{1/2}AB^{1/2}=ww^* B 1/2 A B 1/2 = w w ∗ with w = η f 1 + s η f 2 w=\sqrt\eta f_1+s\eta f_2 w = η f 1 + sη f 2 , so L = ∥ w ∥ 2 w ∗ ρ w = η 2 ρ 11 + O ( η 5 / 2 ) \mathcal{L}=\|w\|^2\,w^*\rho w=\eta^2\rho_{11}+O(\eta^{5/2}) L = ∥ w ∥ 2 w ∗ ρw = η 2 ρ 11 + O ( η 5/2 ) , and B u = η ( f 1 + s f 2 ) Bu=\eta(f_1+sf_2) B u = η ( f 1 + s f 2 ) , so R = ∥ u ∥ 2 η 2 ( ρ 11 + 2 s ρ 12 + s 2 ρ 22 ) \mathcal{R}=\|u\|^2\eta^2(\rho_{11}+2s\rho_{12}+s^2\rho_{22}) R = ∥ u ∥ 2 η 2 ( ρ 11 + 2 s ρ 12 + s 2 ρ 22 ) . As η → 0 \eta\to0 η → 0 , L / R → ρ 11 / ( ρ 11 + 2 s ρ 12 + s 2 ρ 22 ) \mathcal{L}/\mathcal{R}\to \rho_{11}/(\rho_{11}+2s\rho_{12}+s^2\rho_{22}) L / R → ρ 11 / ( ρ 11 + 2 s ρ 12 + s 2 ρ 22 ) , which at s = − ρ 12 / ρ 22 s=-\rho_{12}/\rho_{22} s = − ρ 12 / ρ 22 equals ρ 11 ρ 22 / ( ρ 11 ρ 22 − ρ 12 2 ) = ( a + b ) 2 / ( 4 a b ) = G ( ρ ) 2 \rho_{11}\rho_{22}/(\rho_{11}\rho_{22}-\rho_{12}^2)=(a+b)^2/(4ab)=G(\rho)^2 ρ 11 ρ 22 / ( ρ 11 ρ 22 − ρ 12 2 ) = ( a + b ) 2 / ( 4 ab ) = G ( ρ ) 2 . So c 2 , 1 M 2 ( d i a g ( κ , 1 ) ) ≥ G 2 c_{2,1}^{M_2}(\mathop{\mathrm{diag}}(\kappa,1))\ge G^2 c 2 , 1 M 2 ( diag ( κ , 1 )) ≥ G 2 , and Theorem 4.15 t:low (all expressions are admissible, as in Proposition 5.51 ) gives c 2 , 1 ( ρ ) ≥ G Z 2 c_{2,1}(\rho)\ge G_\mathcal{Z}^2 c 2 , 1 ( ρ ) ≥ G Z 2 for ρ ∉ Z \rho\notin\mathcal{Z} ρ ∈ / Z ; for central ρ \rho ρ both sides are 1 1 1 . ◻
For κ = 10 \kappa=10 κ = 10 the bound is G 2 = 121 / 40 = 3.025 G^2=121/40=3.025 G 2 = 121/40 = 3.025 ; it exceeds the value 2.99 2.99 2.99 found by the multistart search of Remark 5.52 . On M n M_n M n we conjecture that G 2 G^2 G 2 is the constant (Conjecture 11.1 ).
Gardner’s inequality with two contrasts
This section works on matrix algebras and their direct sums. With the operator free, Gardner’s constant is G Z G_\mathcal{Z} G Z in every finite von Neumann algebra (Theorem 5.1 ); with the contrast of ∣ X ∣ |X| ∣ X ∣ prescribed as well, the proofs below use the spectral decomposition of finite matrices and Lemma 3.4 , and we do not know whether Theorem A holds in finite von Neumann algebras with δ \delta δ replaced by the noncentrality (the rescaling X ↦ z X X\mapsto zX X ↦ z X used for condition (R) changes the contrast of ∣ X ∣ |X| ∣ X ∣ ). Theorem C concerns the unnormalised quantities of [85 ] on direct sums of matrix algebras with the trace T r \mathop{\mathrm{Tr}} Tr , for which no central version is needed.
The normalised violation
For Gardner’s inequality both contrasts can be prescribed and the envelope found exactly. It depends on the two contrasts only through their product.
Theorem 6.1. Let ρ > 0 \rho>0 ρ > 0 and X X X be in M n M_n M n , n ≥ 2 n\ge2 n ≥ 2 , with O M C ( ρ ) = δ \mathop{\mathrm{OMC}}(\rho)=\delta OMC ( ρ ) = δ and O M C ( ∣ X ∣ ) = δ ′ \mathop{\mathrm{OMC}}(|X|)=\delta' OMC ( ∣ X ∣ ) = δ ′ , δ , δ ′ ∈ [ 0 , 1 ] \delta,\delta'\in[0,1] δ , δ ′ ∈ [ 0 , 1 ] , δ δ ′ < 1 \delta\delta'<1 δ δ ′ < 1 . Then ∣ φ ρ ( X ) ∣ ≤ 1 1 − δ 2 δ ′ 2 φ ρ ( ∣ X ∣ ) , |\varphi_\rho(X)|\le\frac{1}{\sqrt{1-\delta^2\delta'^2}}\,\varphi_\rho(|X|), ∣ φ ρ ( X ) ∣ ≤ 1 − δ 2 δ ′2 1 φ ρ ( ∣ X ∣ ) , and for even n n n the constant is attained: V n ( δ , δ ′ ) = G ( δ δ ′ ) − 1 = ( 1 − δ 2 δ ′ 2 ) − 1 / 2 − 1 V_n(\delta,\delta')=G(\delta\delta')-1=(1-\delta^2\delta'^2)^{-1/2}-1 V n ( δ , δ ′ ) = G ( δ δ ′ ) − 1 = ( 1 − δ 2 δ ′2 ) − 1/2 − 1 for Gardner’s inequality. For odd n = 2 k + 1 n=2k+1 n = 2 k + 1 , V n ( δ , δ ′ ) = max 0 ≤ γ ≤ 1 k S 2 + γ p q + ℓ ℓ ′ k ( S + γ p ) + ℓ ℓ ′ − 1 , S = ℓ + ℓ ′ , q = ( 1 + ℓ ) ( 1 + ℓ ′ ) , p = ( 1 − ℓ ) ( 1 − ℓ ′ ) , V_n(\delta,\delta')=\max_{0\le\gamma\le1}\frac{k\sqrt{S^2+\gamma pq}+\ell\ell'}{k(S+\gamma p)+\ell\ell'}-1,\qquad
S=\ell+\ell',\ q=(1+\ell)(1+\ell'),\ p=(1-\ell)(1-\ell'), V n ( δ , δ ′ ) = max 0 ≤ γ ≤ 1 k ( S + γ p ) + ℓ ℓ ′ k S 2 + γ pq + ℓ ℓ ′ − 1 , S = ℓ + ℓ ′ , q = ( 1 + ℓ ) ( 1 + ℓ ′ ) , p = ( 1 − ℓ ) ( 1 − ℓ ′ ) , with ℓ = 1 − δ 1 + δ \ell=\frac{1-\delta}{1+\delta} ℓ = 1 + δ 1 − δ , ℓ ′ = 1 − δ ′ 1 + δ ′ \ell'=\frac{1-\delta'}{1+\delta'} ℓ ′ = 1 + δ ′ 1 − δ ′ ; it is smaller than G ( δ δ ′ ) − 1 G(\delta\delta')-1 G ( δ δ ′ ) − 1 for δ ′ < 1 \delta'<1 δ ′ < 1 and tends to it as n → ∞ n\to\infty n → ∞ . For δ ′ = 1 \delta'=1 δ ′ = 1 both give G − 1 G-1 G − 1 , the value of Theorem 5.1 . The upper bound G ( δ δ ′ ) G(\delta\delta') G ( δ δ ′ ) holds also for items 2 and 3, with the contrasts of ∣ X ∣ , ∣ Y ∣ |X|,|Y| ∣ X ∣ , ∣ Y ∣ and of ∣ A ∗ ∣ |A^*| ∣ A ∗ ∣ .
Proof. Write X = U T X=UT X = U T with T = ∣ X ∣ T=|X| T = ∣ X ∣ . The supremum of ∣ T r ( ρ U T ) ∣ = ∣ T r ( U T ρ ) ∣ |\mathop{\mathrm{Tr}}(\rho UT)|=|\mathop{\mathrm{Tr}}(U\,T\rho)| ∣ Tr ( ρ U T ) ∣ = ∣ Tr ( U T ρ ) ∣ over unitaries U U U is ∥ T ρ ∥ 1 \|T\rho\|_1 ∥ T ρ ∥ 1 , so we must bound ∥ T ρ ∥ 1 / T r ( T ρ ) \|T\rho\|_1/\mathop{\mathrm{Tr}}(T\rho) ∥ T ρ ∥ 1 / Tr ( T ρ ) . For fixed eigenbases, ∥ T ρ ∥ 1 − c T r ( T ρ ) \|T\rho\|_1-c\mathop{\mathrm{Tr}}(T\rho) ∥ T ρ ∥ 1 − c Tr ( T ρ ) is a convex function of the eigenvalues of T T T (a norm of a linear function minus a linear one), and likewise of the eigenvalues of ρ \rho ρ . By Lemma 3.6 , applied to T T T and then to ρ \rho ρ , we may assume ρ = a ( ℓ + ( 1 − ℓ ) P ) \rho=a(\ell+(1-\ell)P) ρ = a ( ℓ + ( 1 − ℓ ) P ) and T = t ( ℓ ′ + ( 1 − ℓ ′ ) Q ) T=t(\ell'+(1-\ell')Q) T = t ( ℓ ′ + ( 1 − ℓ ′ ) Q ) with projections P , Q P,Q P , Q and ℓ = 1 − δ 1 + δ \ell=\frac{1-\delta}{1+\delta} ℓ = 1 + δ 1 − δ , ℓ ′ = 1 − δ ′ 1 + δ ′ \ell'=\frac{1-\delta'}{1+\delta'} ℓ ′ = 1 + δ ′ 1 − δ ′ . By Lemma 3.4 , C n \mathbb{C}^n C n splits into subspaces of dimension 1 1 1 or 2 2 2 invariant under P P P and Q Q Q , hence under ρ \rho ρ and T T T , and ∥ T ρ ∥ 1 \|T\rho\|_1 ∥ T ρ ∥ 1 and T r ( T ρ ) \mathop{\mathrm{Tr}}(T\rho) Tr ( T ρ ) are sums over them. On a one-dimensional summand the two contributions are equal. On a two-dimensional one, ρ \rho ρ and T T T have the eigenvalues x 1 , x 2 x_1,x_2 x 1 , x 2 and y 1 , y 2 y_1,y_2 y 1 , y 2 , namely a , a ℓ a,a\ell a , a ℓ and t , t ℓ ′ t,t\ell' t , t ℓ ′ ; with S = x 1 y 2 + x 2 y 1 S=x_1y_2+x_2y_1 S = x 1 y 2 + x 2 y 1 , q = ( x 1 + x 2 ) ( y 1 + y 2 ) q=(x_1+x_2)(y_1+y_2) q = ( x 1 + x 2 ) ( y 1 + y 2 ) , p = ( x 1 − x 2 ) ( y 1 − y 2 ) p=(x_1-x_2)(y_1-y_2) p = ( x 1 − x 2 ) ( y 1 − y 2 ) and γ \gamma γ the squared overlap of the leading eigenvectors, the proof of Proposition 6.3 gives ∥ T ρ ∥ 1 2 = S 2 + γ p q \|T\rho\|_1^2=S^2+\gamma pq ∥ T ρ ∥ 1 2 = S 2 + γ pq and T r ( T ρ ) = S + γ p \mathop{\mathrm{Tr}}(T\rho)=S+\gamma p Tr ( T ρ ) = S + γ p . The function γ ↦ ( S 2 + γ p q ) / ( S + γ p ) 2 \gamma\mapsto(S^2+\gamma pq)/(S+\gamma p)^2 γ ↦ ( S 2 + γ pq ) / ( S + γ p ) 2 is maximal at γ = S / q ∈ [ 0 , 1 ] \gamma=S/q\in[0,1] γ = S / q ∈ [ 0 , 1 ] , with the value
q 2 4 S ( S + p ) = ( x 1 + x 2 ) 2 ( y 1 + y 2 ) 2 4 ( x 1 y 2 + x 2 y 1 ) ( x 1 y 1 + x 2 y 2 ) = 1 1 − δ 2 δ ′ 2 , \frac{q^2}{4S(S+p)}=\frac{(x_1+x_2)^2(y_1+y_2)^2}{4(x_1y_2+x_2y_1)(x_1y_1+x_2y_2)}=\frac1{1-\delta^2\delta'^2}, 4 S ( S + p ) q 2 = 4 ( x 1 y 2 + x 2 y 1 ) ( x 1 y 1 + x 2 y 2 ) ( x 1 + x 2 ) 2 ( y 1 + y 2 ) 2 = 1 − δ 2 δ ′2 1 , the last step by inserting x 2 / x 1 = ℓ x_2/x_1=\ell x 2 / x 1 = ℓ , y 2 / y 1 = ℓ ′ y_2/y_1=\ell' y 2 / y 1 = ℓ ′ . A quotient of sums is at most the largest quotient of the summands. Equality for even n n n : n / 2 n/2 n /2 copies of the 2 × 2 2\times2 2 × 2 configuration ρ = d i a g ( 1 , ℓ ) \rho=\mathrm{diag}(1,\ell) ρ = diag ( 1 , ℓ ) , T T T with the eigenvalues 1 , ℓ ′ 1,\ell' 1 , ℓ ′ at γ = S / q \gamma=S/q γ = S / q , and U U U the adjoint polar factor of T ρ T\rho T ρ . For odd n n n at least one summand is one-dimensional and contributes the same amount μ ≥ ℓ ℓ ′ \mu\ge\ell\ell' μ ≥ ℓ ℓ ′ (normalising the largest eigenvalues of ρ \rho ρ and T T T to 1 1 1 ) to ∥ T ρ ∥ 1 \|T\rho\|_1 ∥ T ρ ∥ 1 and to T r ( T ρ ) \mathop{\mathrm{Tr}}(T\rho) Tr ( T ρ ) ; the quotient ( ∑ j s j + μ ) / ( ∑ j t j + μ ) (\sum_js_j+\mu)/(\sum_jt_j+\mu) ( ∑ j s j + μ ) / ( ∑ j t j + μ ) , where s j ≥ t j s_j\ge t_j s j ≥ t j are the contributions of at most k k k two-dimensional summands, is largest for k k k summands and μ = ℓ ℓ ′ \mu=\ell\ell' μ = ℓ ℓ ′ , and, by the argument of Lemma 3.6 applied to s − c t s-ct s − c t , for the same γ \gamma γ in every summand, which gives the stated maximum; it is attained by ρ = d i a g ( 1 , ℓ ) ⊕ k ⊕ ℓ \rho=\mathrm{diag}(1,\ell)^{\oplus k}\oplus\ell ρ = diag ( 1 , ℓ ) ⊕ k ⊕ ℓ and T T T correspondingly. Items 2 and 3 reduce to the first inequality as in the proofs of Theorem 5.4 and Corollary 5.3 , since ∣ U ∗ X ∣ = ∣ X ∣ |U^*X|=|X| ∣ U ∗ X ∣ = ∣ X ∣ for unitary U U U . ◻
Thus the normalised violation of Gardner’s inequality is 1 2 δ 2 δ ′ 2 + O ( δ 4 δ ′ 4 ) \frac12\delta^2\delta'^2+O(\delta^4\delta'^4) 2 1 δ 2 δ ′2 + O ( δ 4 δ ′4 ) : it vanishes to second order in each contrast and is symmetric in the density and the operator. For a Gibbs state at inverse temperature β \beta β and an operator ∣ X ∣ = e − β ′ K |X|=e^{-\beta' K} ∣ X ∣ = e − β ′ K , the constant is ( 1 − tanh 2 ( β Δ E / 2 ) tanh 2 ( β ′ Δ E ′ / 2 ) ) − 1 / 2 (1-\tanh^2(\beta\Delta E/2)\tanh^2(\beta'\Delta
E'/2))^{-1/2} ( 1 − tanh 2 ( β Δ E /2 ) tanh 2 ( β ′ Δ E ′ /2 ) ) − 1/2 , where Δ E \Delta E Δ E and Δ E ′ \Delta E' Δ E ′ are the spectral spreads of H H H and K K K . The two-dimensional computation in the proof has a transparent form, which explains the dependence on the product.
The unnormalised violations of [85 ]
We return to the unnormalised quantity of [85 ] , in which both operators are normalised to norm one and both contrasts are prescribed. Let ν = ∑ c ⌊ k c / 2 ⌋ \nu=\sum_c\lfloor k_c/2\rfloor ν = ∑ c ⌊ k c /2 ⌋ be the largest number of mutually orthogonal two-dimensional subspaces that fit into the blocks of M \mathcal{M} M .
Theorem C 1. Let M = ⨁ c = 1 C M k c ( K ) \mathcal{M}=\bigoplus_{c=1}^CM_{k_c}(\mathbb K) M = ⨁ c = 1 C M k c ( K ) , K ∈ { R , C } \mathbb K\in\{\mathbb{R},\mathbb{C}\} K ∈ { R , C } , with ν ≥ 1 \nu\ge1 ν ≥ 1 . For X ≥ 0 X\ge0 X ≥ 0 and Z Z Z in M \mathcal{M} M with ∥ X ∥ = ∥ Z ∥ = 1 \|X\|=\|Z\|=1 ∥ X ∥ = ∥ Z ∥ = 1 , O M C ( X ) = δ < 1 \mathop{\mathrm{OMC}}(X)=\delta<1 OMC ( X ) = δ < 1 and O M C ( ∣ Z ∣ ) = δ ′ \mathop{\mathrm{OMC}}(|Z|)=\delta' OMC ( ∣ Z ∣ ) = δ ′ , sup ( ∣ T r ( X Z ) ∣ − T r ( X ∣ Z ∣ ) ) = ν δ 2 δ ′ 2 ( 1 + δ ) ( 1 + δ ′ ) , \sup\bigl(|\mathop{\mathrm{Tr}}(XZ)|-\mathop{\mathrm{Tr}}(X|Z|)\bigr)=\frac{\nu\,\delta^2\delta'^2}{(1+\delta)(1+\delta')}, sup ( ∣ Tr ( X Z ) ∣ − Tr ( X ∣ Z ∣ ) ) = ( 1 + δ ) ( 1 + δ ′ ) ν δ 2 δ ′2 , and the supremum is attained (Theorem 6.4 ). The two other functionals of [85 ] , T r ∣ X Y ∣ − T r ( ∣ X ∣ ∣ Y ∣ ) \mathop{\mathrm{Tr}}|XY|-\mathop{\mathrm{Tr}}(|X||Y|) Tr ∣ X Y ∣ − Tr ( ∣ X ∣∣ Y ∣ ) and T r ( Y ∣ X ∣ Y ) − T r ∣ Y X Y ∣ \mathop{\mathrm{Tr}}(Y|X|Y)-\mathop{\mathrm{Tr}}|YXY| Tr ( Y ∣ X ∣ Y ) − Tr ∣ Y X Y ∣ , have the sharp bounds of Proposition 6.6 and Theorem 6.9 .
The unnormalised violation behaves like the normalised one.
In [85 Sect. 5] the violation of Gardner’s inequality was studied without normalisation: for X ≥ 0 X\ge0 X ≥ 0 with ∥ X ∥ = 1 \|X\|=1 ∥ X ∥ = 1 (the density) and Z Z Z with ∥ Z ∥ = 1 \|Z\|=1 ∥ Z ∥ = 1 , the quantity
z ( X , Z ) = ∣ T r ( X Z ) ∣ − T r ( X ∣ Z ∣ ) z(X,Z)=|\mathop{\mathrm{Tr}}(XZ)|-\mathop{\mathrm{Tr}}(X|Z|) z ( X , Z ) = ∣ Tr ( X Z ) ∣ − Tr ( X ∣ Z ∣ )
was sampled against O M C ( X ) \mathop{\mathrm{OMC}}(X) OMC ( X ) and O M C ( ∣ Z ∣ ) \mathop{\mathrm{OMC}}(|Z|) OMC ( ∣ Z ∣ ) . Writing Z = U ∣ Z ∣ Z=U|Z| Z = U ∣ Z ∣ with a unitary (orthogonal, in the real case) U U U , z = ∣ T r ( X U Y ) ∣ − T r ( X Y ) z=|\mathop{\mathrm{Tr}}(XUY)|-\mathop{\mathrm{Tr}}(XY) z = ∣ Tr ( X U Y ) ∣ − Tr ( X Y ) with Y = ∣ Z ∣ ≥ 0 Y=|Z|\ge0 Y = ∣ Z ∣ ≥ 0 , and sup U z = ∥ Y X ∥ 1 − T r ( X Y ) \sup_Uz=\|YX\|_1-\mathop{\mathrm{Tr}}(XY) sup U z = ∥ Y X ∥ 1 − Tr ( X Y ) by the duality of the trace norm, the supremum being attained at the polar unitary of Y X YX Y X , which is real if X , Y X,Y X , Y are. We first solve the 2 × 2 2\times2 2 × 2 case.
Proposition 6.3. Let X , Y ≥ 0 X,Y\ge0 X , Y ≥ 0 in M 2 ( C ) M_2(\mathbb{C}) M 2 ( C ) have eigenvalues x 1 ≥ x 2 ≥ 0 x_1\ge x_2\ge0 x 1 ≥ x 2 ≥ 0 and y 1 ≥ y 2 ≥ 0 y_1\ge y_2\ge0 y 1 ≥ y 2 ≥ 0 , x 1 , y 1 > 0 x_1,y_1>0 x 1 , y 1 > 0 . For every unitary U U U , ∣ T r ( X U Y ) ∣ − T r ( X Y ) ≤ ( x 1 − x 2 ) 2 ( y 1 − y 2 ) 2 4 ( x 1 + x 2 ) ( y 1 + y 2 ) . |\mathop{\mathrm{Tr}}(XUY)|-\mathop{\mathrm{Tr}}(XY)\ \le\ \frac{(x_1-x_2)^2(y_1-y_2)^2}{4(x_1+x_2)(y_1+y_2)} . ∣ Tr ( X U Y ) ∣ − Tr ( X Y ) ≤ 4 ( x 1 + x 2 ) ( y 1 + y 2 ) ( x 1 − x 2 ) 2 ( y 1 − y 2 ) 2 . Let S = x 1 y 2 + x 2 y 1 S=x_1y_2+x_2y_1 S = x 1 y 2 + x 2 y 1 , q = ( x 1 + x 2 ) ( y 1 + y 2 ) q=(x_1+x_2)(y_1+y_2) q = ( x 1 + x 2 ) ( y 1 + y 2 ) , and γ = ∣ ⟨ e 1 , f 1 ⟩ ∣ 2 \gamma=|\langle e_1,f_1\rangle|^2 γ = ∣ ⟨ e 1 , f 1 ⟩ ∣ 2 for unit eigenvectors e 1 e_1 e 1 of X X X for x 1 x_1 x 1 and f 1 f_1 f 1 of Y Y Y for y 1 y_1 y 1 . If x 1 > x 2 x_1>x_2 x 1 > x 2 and y 1 > y 2 y_1>y_2 y 1 > y 2 , the bound is attained iff ∣ T r ( X U Y ) ∣ = ∥ Y X ∥ 1 |\mathop{\mathrm{Tr}}(XUY)|=\|YX\|_1 ∣ Tr ( X U Y ) ∣ = ∥ Y X ∥ 1 and γ = 1 4 + S 2 q \gamma=\frac14+\frac S{2q} γ = 4 1 + 2 q S ; it is attained by real X , Y , U X,Y,U X , Y , U .
Proof. Let p = ( x 1 − x 2 ) ( y 1 − y 2 ) p=(x_1-x_2)(y_1-y_2) p = ( x 1 − x 2 ) ( y 1 − y 2 ) , so that q − 2 S = p q-2S=p q − 2 S = p . First, ∣ T r ( X U Y ) ∣ = ∣ T r ( Y X U ) ∣ ≤ ∥ Y X ∥ 1 |\mathop{\mathrm{Tr}}(XUY)|=|\mathop{\mathrm{Tr}}(YXU)|\le\|YX\|_1 ∣ Tr ( X U Y ) ∣ = ∣ Tr ( Y X U ) ∣ ≤ ∥ Y X ∥ 1 . Expanding in the eigenbases, T r ( X Y ) = ∑ i j x i y j ∣ ⟨ e i , f j ⟩ ∣ 2 = S + γ p \mathop{\mathrm{Tr}}(XY)=\sum_{ij}x_iy_j|\langle e_i,f_j\rangle|^2=S+\gamma p Tr ( X Y ) = ∑ ij x i y j ∣ ⟨ e i , f j ⟩ ∣ 2 = S + γ p , since the matrix ( ∣ ⟨ e i , f j ⟩ ∣ 2 ) (|\langle e_i,f_j\rangle|^2) ( ∣ ⟨ e i , f j ⟩ ∣ 2 ) is doubly stochastic. The singular values of Y X YX Y X satisfy s 1 2 + s 2 2 = T r ( X 2 Y 2 ) s_1^2+s_2^2=\mathop{\mathrm{Tr}}(X^2Y^2) s 1 2 + s 2 2 = Tr ( X 2 Y 2 ) and s 1 s 2 = ∣ det Y X ∣ = x 1 x 2 y 1 y 2 s_1s_2=|\det YX|=x_1x_2y_1y_2 s 1 s 2 = ∣ det Y X ∣ = x 1 x 2 y 1 y 2 , so ∥ Y X ∥ 1 2 = T r ( X 2 Y 2 ) + 2 x 1 x 2 y 1 y 2 = S 2 + γ p q \|YX\|_1^2=\mathop{\mathrm{Tr}}(X^2Y^2)+2x_1x_2y_1y_2=S^2+\gamma pq ∥ Y X ∥ 1 2 = Tr ( X 2 Y 2 ) + 2 x 1 x 2 y 1 y 2 = S 2 + γ pq . With r ≤ r / q + q / 4 \sqrt r\le r/q+q/4 r ≤ r / q + q /4 (equality iff r = q / 2 \sqrt r=q/2 r = q /2 ),
∥ Y X ∥ 1 − T r ( X Y ) ≤ S 2 + γ p q q + q 4 − S − γ p = ( q / 2 − S ) 2 q = p 2 4 q , \|YX\|_1-\mathop{\mathrm{Tr}}(XY)\le\frac{S^2+\gamma pq}q+\frac q4-S-\gamma p=\frac{(q/2-S)^2}q=\frac{p^2}{4q}, ∥ Y X ∥ 1 − Tr ( X Y ) ≤ q S 2 + γ pq + 4 q − S − γ p = q ( q /2 − S ) 2 = 4 q p 2 , with equality iff γ p q = q 2 / 4 − S 2 = p 2 ( q 2 + S ) \gamma pq=q^2/4-S^2=\frac p2(\frac q2+S) γ pq = q 2 /4 − S 2 = 2 p ( 2 q + S ) . Every γ ∈ [ 0 , 1 ] \gamma\in[0,1] γ ∈ [ 0 , 1 ] is realised by a real rotation, and the polar unitary of a real Y X YX Y X is real. ◻
Theorem 6.4. Let M = ⨁ c = 1 C M k c ( K ) \mathcal{M}=\bigoplus_{c=1}^CM_{k_c}(\mathbb K) M = ⨁ c = 1 C M k c ( K ) , K ∈ { R , C } \mathbb K\in\{\mathbb{R},\mathbb{C}\} K ∈ { R , C } , and ν = ∑ c ⌊ k c / 2 ⌋ ≥ 1 \nu=\sum_c\lfloor k_c/2\rfloor\ge1 ν = ∑ c ⌊ k c /2 ⌋ ≥ 1 . For X ≥ 0 X\ge0 X ≥ 0 and Y ≥ 0 Y\ge0 Y ≥ 0 in M \mathcal{M} M with ∥ X ∥ = ∥ Y ∥ = 1 \|X\|=\|Y\|=1 ∥ X ∥ = ∥ Y ∥ = 1 , O M C ( X ) = δ ∈ [ 0 , 1 ) \mathop{\mathrm{OMC}}(X)=\delta\in[0,1) OMC ( X ) = δ ∈ [ 0 , 1 ) and O M C ( Y ) = δ ′ ∈ [ 0 , 1 ] \mathop{\mathrm{OMC}}(Y)=\delta'\in[0,1] OMC ( Y ) = δ ′ ∈ [ 0 , 1 ] , and every unitary (orthogonal) U ∈ M U\in\mathcal{M} U ∈ M , ∣ T r ( X U Y ) ∣ − T r ( X Y ) ≤ ν δ 2 δ ′ 2 ( 1 + δ ) ( 1 + δ ′ ) , |\mathop{\mathrm{Tr}}(XUY)|-\mathop{\mathrm{Tr}}(XY)\le\frac{\nu\,\delta^2\delta'^2}{(1+\delta)(1+\delta')}, ∣ Tr ( X U Y ) ∣ − Tr ( X Y ) ≤ ( 1 + δ ) ( 1 + δ ′ ) ν δ 2 δ ′2 , and the bound is attained. Equivalently, sup { ∣ T r ( X Z ) ∣ − T r ( X ∣ Z ∣ ) : ∥ X ∥ = ∥ Z ∥ = 1 , O M C ( X ) = δ , O M C ( ∣ Z ∣ ) = δ ′ } \sup\{|\mathop{\mathrm{Tr}}(XZ)|-\mathop{\mathrm{Tr}}(X|Z|):\|X\|=\|Z\|=1,\ \mathop{\mathrm{OMC}}(X)=\delta,\ \mathop{\mathrm{OMC}}(|Z|)=\delta'\} sup { ∣ Tr ( X Z ) ∣ − Tr ( X ∣ Z ∣ ) : ∥ X ∥ = ∥ Z ∥ = 1 , OMC ( X ) = δ , OMC ( ∣ Z ∣ ) = δ ′ } has this value.
Proof. Let ℓ = ( 1 − δ ) / ( 1 + δ ) \ell=(1-\delta)/(1+\delta) ℓ = ( 1 − δ ) / ( 1 + δ ) and ℓ ′ = ( 1 − δ ′ ) / ( 1 + δ ′ ) \ell'=(1-\delta')/(1+\delta') ℓ ′ = ( 1 − δ ′ ) / ( 1 + δ ′ ) be the smallest eigenvalues. It suffices to prove the bound in the complex algebra, since the maximisers below are real. By the remark above, sup U z = F ( X , Y ) : = ∥ Y X ∥ 1 − T r ( X Y ) \sup_Uz=F(X,Y):=\|YX\|_1-\mathop{\mathrm{Tr}}(XY) sup U z = F ( X , Y ) := ∥ Y X ∥ 1 − Tr ( X Y ) , and the polar unitary lies in M \mathcal{M} M . For a fixed eigenbasis of X X X , F F F is a convex function of the eigenvalues of X X X (a norm of a linear function of them minus a linear function). The admissible eigenvalues form the union of the faces { x i = 1 , x j = ℓ } \{x_i=1,\ x_j=\ell\} { x i = 1 , x j = ℓ } of the cube [ ℓ , 1 ] K [\ell,1]^K [ ℓ , 1 ] K , K = ∑ c k c K=\sum_ck_c K = ∑ c k c , and a convex function on a face attains its maximum at a vertex; so F ( X , Y ) ≤ F ( X ′ , Y ) F(X,Y)\le F(X',Y) F ( X , Y ) ≤ F ( X ′ , Y ) for some X ′ = ℓ + ( 1 − ℓ ) P X'=\ell+(1-\ell)P X ′ = ℓ + ( 1 − ℓ ) P with a projection P ∈ M P\in\mathcal{M} P ∈ M , P ≠ 0 , 1 P\ne0,1 P = 0 , 1 , and likewise F ( X ′ , Y ) ≤ F ( X ′ , Y ′ ) F(X',Y)\le F(X',Y') F ( X ′ , Y ) ≤ F ( X ′ , Y ′ ) with Y ′ = ℓ ′ + ( 1 − ℓ ′ ) Q Y'=\ell'+(1-\ell')Q Y ′ = ℓ ′ + ( 1 − ℓ ′ ) Q . By Lemma 3.4 , applied in every block, X ′ X' X ′ and Y ′ Y' Y ′ are direct sums over subspaces of dimension 1 1 1 or 2 2 2 , and so are ∥ Y ′ X ′ ∥ 1 \|Y'X'\|_1 ∥ Y ′ X ′ ∥ 1 and T r ( X ′ Y ′ ) \mathop{\mathrm{Tr}}(X'Y') Tr ( X ′ Y ′ ) . One-dimensional summands contribute 0 0 0 . On a two-dimensional one, X ′ = d i a g ( 1 , ℓ ) X'=\mathrm{diag}(1,\ell) X ′ = diag ( 1 , ℓ ) and Y ′ Y' Y ′ has the eigenvalues 1 , ℓ ′ 1,\ell' 1 , ℓ ′ , so by Proposition 6.3 it contributes at most ( 1 − ℓ ) 2 ( 1 − ℓ ′ ) 2 / ( 4 ( 1 + ℓ ) ( 1 + ℓ ′ ) ) = δ 2 δ ′ 2 / ( ( 1 + δ ) ( 1 + δ ′ ) ) (1-\ell)^2(1-\ell')^2/(4(1+\ell)(1+\ell'))=\delta^2\delta'^2/((1+\delta)(1+\delta')) ( 1 − ℓ ) 2 ( 1 − ℓ ′ ) 2 / ( 4 ( 1 + ℓ ) ( 1 + ℓ ′ )) = δ 2 δ ′2 / (( 1 + δ ) ( 1 + δ ′ )) . A block of size k c k_c k c holds at most ⌊ k c / 2 ⌋ \lfloor k_c/2\rfloor ⌊ k c /2 ⌋ two-dimensional summands. Equality: ν \nu ν blocks d i a g ( 1 , ℓ ) \mathrm{diag}(1,\ell) diag ( 1 , ℓ ) and Y Y Y with spectrum { 1 , ℓ ′ } \{1,\ell'\} { 1 , ℓ ′ } at the angle of Proposition 6.3 , X = Y = 1 X=Y=1 X = Y = 1 on the remaining one-dimensional blocks, and the polar unitary, which makes all block traces positive. ◻
Figure 3. The unnormalised Gardner violation z = ∣ T r ( X U Y ) ∣ − T r ( X Y ) z=|\mathop{\mathrm{Tr}}(XUY)|-\mathop{\mathrm{Tr}}(XY) z = ∣ Tr ( X U Y ) ∣ − Tr ( X Y ) for ∥ X ∥ = ∥ Y ∥ = 1 \|X\|=\|Y\|=1 ∥ X ∥ = ∥ Y ∥ = 1 in M 2 ( R ) M_2(\mathbb{R}) M 2 ( R ) (top, ν = 1 \nu=1 ν = 1 ) and M 16 ( R ) M_{16}(\mathbb{R}) M 16 ( R ) (bottom, ν = 8 \nu=8 ν = 8 ), against O M C ( X ) \mathop{\mathrm{OMC}}(X) OMC ( X ) (left) and O M C ( Y ) \mathop{\mathrm{OMC}}(Y) OMC ( Y ) (right). Grey: 100000 100000 100000 random samples with Haar-orthogonal eigenbases and U U U and uniformly drawn contrasts, as in [85 Fig. 1] . Solid: the bound of Theorem 6.4 with O M C ( Y ) \mathop{\mathrm{OMC}}(Y) OMC ( Y ) free, ν δ 2 / ( 2 ( 1 + δ ) ) \nu\delta^2/(2(1+\delta)) ν δ 2 / ( 2 ( 1 + δ )) ; dots: z z z at the explicit maximisers of its proof, which agree with it to 3 × 10 − 15 3\times10^{-15} 3 × 1 0 − 15 (printed by trace_envelopes_ figures.py). Dashed: the supremum along O M C ( Y ) \mathop{\mathrm{OMC}}(Y) OMC ( Y ) , approached as O M C ( X ) → 1 \mathop{\mathrm{OMC}}(X)\to1 OMC ( X ) → 1 . In M 16 M_{16} M 16 no sample violates Gardner’s inequality (z > 0 z>0 z > 0 ), although violations up to almost 2 2 2 exist: the maximisers, two-level spectra on aligned pairs of eigenvectors, form a set of negligible Haar measure. In the figure Δ ( ⋅ ) = O M C ( ⋅ ) \Delta(\cdot)=\mathop{\mathrm{OMC}}(\cdot) Δ ( ⋅ ) = OMC ( ⋅ ) .
Figure 4. M 2 ( R ) M_2(\mathbb{R}) M 2 ( R ) , both contrasts prescribed: the supremum (left) and infimum (right) of z = ∣ T r ( X U Y ) ∣ − T r ( X Y ) z=|\mathop{\mathrm{Tr}}(XUY)|-\mathop{\mathrm{Tr}}(XY) z = ∣ Tr ( X U Y ) ∣ − Tr ( X Y ) found by constrained optimisation on a grid of O M C ( Y ) \mathop{\mathrm{OMC}}(Y) OMC ( Y ) (horizontal) and O M C ( X ) \mathop{\mathrm{OMC}}(X) OMC ( X ) (vertical). The left panel agrees with δ 2 δ ′ 2 / ( ( 1 + δ ) ( 1 + δ ′ ) ) \delta^2\delta'^2/((1+\delta)(1+\delta')) δ 2 δ ′2 / (( 1 + δ ) ( 1 + δ ′ )) (Theorem 6.4 ) to 5 × 10 − 10 5\times 10^{-10} 5 × 1 0 − 10 at every grid point: a violation needs both operators to be far from scalar. In the figure Δ ( ⋅ ) = O M C ( ⋅ ) \Delta(\cdot)=\mathop{\mathrm{OMC}}(\cdot) Δ ( ⋅ ) = OMC ( ⋅ ) .
The other two functionals of [85 ]
Besides Gardner’s inequality, [85 Sect. 5] sampled two further unnormalised trace functionals,
z 2 = T r ∣ X Y ∣ − T r ( ∣ X ∣ ∣ Y ∣ ) , z 3 = T r ( Y ∣ X ∣ Y ) − T r ∣ Y X Y ∣ ( X = X ∗ , Y ≥ 0 ) , z_2=\mathop{\mathrm{Tr}}|XY|-\mathop{\mathrm{Tr}}(|X|\,|Y|),\qquad z_3=\mathop{\mathrm{Tr}}(Y|X|Y)-\mathop{\mathrm{Tr}}|YXY|\quad(X=X^*,\ Y\ge0), z 2 = Tr ∣ X Y ∣ − Tr ( ∣ X ∣ ∣ Y ∣ ) , z 3 = Tr ( Y ∣ X ∣ Y ) − Tr ∣ Y X Y ∣ ( X = X ∗ , Y ≥ 0 ) ,
with ∣ X ∣ |X| ∣ X ∣ and ∣ Y ∣ |Y| ∣ Y ∣ normalised to λ max = 1 \lambda_{\max}=1 λ m a x = 1 and their contrasts as parameters. Both are determined here in M = ⨁ c M k c \mathcal{M}=\bigoplus_cM_{k_c} M = ⨁ c M k c , with ν = ∑ c ⌊ k c / 2 ⌋ \nu=\sum_c\lfloor k_c/2\rfloor ν = ∑ c ⌊ k c /2 ⌋ and ℓ = 1 − δ 1 + δ \ell=\frac{1-\delta}{1+\delta} ℓ = 1 + δ 1 − δ , ℓ ′ = 1 − δ ′ 1 + δ ′ \ell'=\frac{1-\delta'}{1+\delta'} ℓ ′ = 1 + δ ′ 1 − δ ′ as before; note 1 − ℓ = 2 δ 1 + δ 1-\ell=\frac{2\delta}{1+\delta} 1 − ℓ = 1 + δ 2 δ .
Proposition 6.6. Let X = U X 0 X=UX_0 X = U X 0 , Y = V Y 0 Y=VY_0 Y = V Y 0 with X 0 , Y 0 ≥ 0 X_0,Y_0\ge0 X 0 , Y 0 ≥ 0 and unitaries U , V ∈ M U,V\in\mathcal{M} U , V ∈ M , so that ∣ X ∣ = X 0 |X|=X_0 ∣ X ∣ = X 0 , ∣ Y ∣ = Y 0 |Y|=Y_0 ∣ Y ∣ = Y 0 and z 2 = ∥ X 0 V Y 0 ∥ 1 − T r ( X 0 Y 0 ) z_2=\|X_0VY_0\|_1-\mathop{\mathrm{Tr}}(X_0Y_0) z 2 = ∥ X 0 V Y 0 ∥ 1 − Tr ( X 0 Y 0 ) . For given spectra, sup z 2 = − inf z 2 = ∑ c ∑ i x c , i ↓ ( y c , i ↓ − y c , i ↑ ) , \sup z_2=-\inf z_2=\sum_c\sum_ix_{c,i}^\downarrow\bigl(y_{c,i}^\downarrow-y_{c,i}^\uparrow\bigr), sup z 2 = − inf z 2 = ∑ c ∑ i x c , i ↓ ( y c , i ↓ − y c , i ↑ ) , with x c , i ↓ x_{c,i}^\downarrow x c , i ↓ the eigenvalues of X 0 , c X_{0,c} X 0 , c in decreasing order and y c , i ↓ y_{c,i}^\downarrow y c , i ↓ , y c , i ↑ y_{c,i}^\uparrow y c , i ↑ those of Y 0 , c Y_{0,c} Y 0 , c in decreasing and increasing order. With O M C ( ∣ X ∣ ) = δ \mathop{\mathrm{OMC}}(|X|)=\delta OMC ( ∣ X ∣ ) = δ and O M C ( ∣ Y ∣ ) = δ ′ \mathop{\mathrm{OMC}}(|Y|)=\delta' OMC ( ∣ Y ∣ ) = δ ′ prescribed, sup z 2 = − inf z 2 = ν ( 1 − ℓ ) ( 1 − ℓ ′ ) = 4 ν δ δ ′ ( 1 + δ ) ( 1 + δ ′ ) \sup z_2=-\inf z_2=\nu(1-\ell)(1-\ell')=\frac{4\nu\delta\delta'}{(1+\delta)(1+\delta')} sup z 2 = − inf z 2 = ν ( 1 − ℓ ) ( 1 − ℓ ′ ) = ( 1 + δ ) ( 1 + δ ′ ) 4 ν δ δ ′ , attained by two-level spectra.
Proof. V V V and the eigenbases act sector by sector, and von Neumann’s trace inequality ∥ A B ∥ 1 ≤ ∑ i s i ( A ) s i ( B ) \|AB\|_1\le\sum_is_i(A)s_i(B) ∥ A B ∥ 1 ≤ ∑ i s i ( A ) s i ( B ) , its lower counterpart ∥ A B ∥ 1 ≥ ∑ i s i ( A ) s k + 1 − i ( B ) \|AB\|_1\ge\sum_is_i(A)s_{k+1-i}(B) ∥ A B ∥ 1 ≥ ∑ i s i ( A ) s k + 1 − i ( B ) and the rearrangement inequality for T r ( X 0 Y 0 ) \mathop{\mathrm{Tr}}(X_0Y_0) Tr ( X 0 Y 0 ) give the first formula; both extremes are attained by aligning or reversing the eigenbases. Pairing i i i with k + 1 − i k+1-i k + 1 − i , ∑ i x c , i ↓ ( y c , i ↓ − y c , i ↑ ) = ∑ i ≤ k / 2 ( x c , i ↓ − x c , k + 1 − i ↓ ) ( y c , i ↓ − y c , k + 1 − i ↓ ) ≤ ⌊ k / 2 ⌋ ( 1 − ℓ ) ( 1 − ℓ ′ ) \sum_ix_{c,i}^\downarrow(y_{c,i}^\downarrow-y_{c,i}^\uparrow)=\sum_{i\le k/2}(x_{c,i}^\downarrow-x_{c,k+1-i}^\downarrow)
(y_{c,i}^\downarrow-y_{c,k+1-i}^\downarrow)\le\lfloor k/2\rfloor(1-\ell)(1-\ell') ∑ i x c , i ↓ ( y c , i ↓ − y c , i ↑ ) = ∑ i ≤ k /2 ( x c , i ↓ − x c , k + 1 − i ↓ ) ( y c , i ↓ − y c , k + 1 − i ↓ ) ≤ ⌊ k /2 ⌋ ( 1 − ℓ ) ( 1 − ℓ ′ ) , with equality for two-level spectra. ◻
So z 2 z_2 z 2 is linear in each contrast, unlike Gardner’s violation, which is quadratic in each. The third functional is never negative:
Proposition 6.7. For X = X ∗ X=X^* X = X ∗ and Y ≥ 0 Y\ge0 Y ≥ 0 , T r ∣ Y X Y ∣ ≤ T r ( Y ∣ X ∣ Y ) \mathop{\mathrm{Tr}}|YXY|\le\mathop{\mathrm{Tr}}(Y|X|Y) Tr ∣ Y X Y ∣ ≤ Tr ( Y ∣ X ∣ Y ) , with equality if and only if Y X + Y 2 X − Y = 0 YX_+Y^2X_-Y=0 Y X + Y 2 X − Y = 0 , in particular if Y Y Y commutes with X X X .
Proof. With X = X + − X − X=X_+-X_- X = X + − X − and ∣ X ∣ = X + + X − |X|=X_++X_- ∣ X ∣ = X + + X − , the triangle inequality for the trace norm gives ∥ Y X + Y − Y X − Y ∥ 1 ≤ T r ( Y X + Y ) + T r ( Y X − Y ) \|YX_+Y-YX_-Y\|_1\le\mathop{\mathrm{Tr}}(YX_+Y)+\mathop{\mathrm{Tr}}(YX_-Y) ∥ Y X + Y − Y X − Y ∥ 1 ≤ Tr ( Y X + Y ) + Tr ( Y X − Y ) . For positive A , B A,B A , B , ∥ A − B ∥ 1 = T r A + T r B \|A-B\|_1=\mathop{\mathrm{Tr}}A+\mathop{\mathrm{Tr}}B ∥ A − B ∥ 1 = Tr A + Tr B if and only if A B = 0 AB=0 A B = 0 : with P P P the projection onto the positive part of A − B A-B A − B , ∥ A − B ∥ 1 = T r A + T r B − 2 T r ( ( 1 − P ) A ) − 2 T r ( P B ) \|A-B\|_1=\mathop{\mathrm{Tr}}A+\mathop{\mathrm{Tr}}B-2\mathop{\mathrm{Tr}}((1-P)A)-2\mathop{\mathrm{Tr}}(PB) ∥ A − B ∥ 1 = Tr A + Tr B − 2 Tr (( 1 − P ) A ) − 2 Tr ( P B ) , and equality forces ( 1 − P ) A = 0 = P B (1-P)A=0=PB ( 1 − P ) A = 0 = P B , so A B = A P ( 1 − P ) B = 0 AB=AP(1-P)B=0 A B = A P ( 1 − P ) B = 0 ; conversely, if A B = 0 AB=0 A B = 0 the supports are orthogonal and ∣ A − B ∣ = A + B |A-B|=A+B ∣ A − B ∣ = A + B . Apply this to A = Y X + Y A=YX_+Y A = Y X + Y , B = Y X − Y B=YX_-Y B = Y X − Y . If X Y = Y X XY=YX X Y = Y X , then Y X + Y 2 X − Y = Y 4 X + X − = 0 YX_+Y^2X_-Y=Y^4X_+X_-=0 Y X + Y 2 X − Y = Y 4 X + X − = 0 . ◻
Proposition 6.8. Let X = X ∗ X=X^* X = X ∗ in M 2 ( C ) M_2(\mathbb{C}) M 2 ( C ) , let ∣ X ∣ |X| ∣ X ∣ have eigenvalues x 1 ≥ x 2 ≥ 0 x_1\ge x_2\ge0 x 1 ≥ x 2 ≥ 0 , x 1 > 0 x_1>0 x 1 > 0 , and let Y ≥ 0 Y\ge0 Y ≥ 0 have eigenvalues y 1 ≥ y 2 ≥ 0 y_1\ge y_2\ge0 y 1 ≥ y 2 ≥ 0 . Then T r ( Y ∣ X ∣ Y ) − T r ∣ Y X Y ∣ ≤ 2 x 1 x 2 x 1 + x 2 ( y 1 − y 2 ) 2 . \mathop{\mathrm{Tr}}(Y|X|Y)-\mathop{\mathrm{Tr}}|YXY|\le\frac{2x_1x_2}{x_1+x_2}\,(y_1-y_2)^2 . Tr ( Y ∣ X ∣ Y ) − Tr ∣ Y X Y ∣ ≤ x 1 + x 2 2 x 1 x 2 ( y 1 − y 2 ) 2 . With S = x 1 y 2 + x 2 y 1 S=x_1y_2+x_2y_1 S = x 1 y 2 + x 2 y 1 , q = ( x 1 + x 2 ) ( y 1 + y 2 ) q=(x_1+x_2)(y_1+y_2) q = ( x 1 + x 2 ) ( y 1 + y 2 ) and γ = ∣ ⟨ e 1 , f 1 ⟩ ∣ 2 \gamma=|\langle e_1,f_1\rangle|^2 γ = ∣ ⟨ e 1 , f 1 ⟩ ∣ 2 for unit eigenvectors e 1 e_1 e 1 of X X X for its eigenvalue of modulus x 1 x_1 x 1 and f 1 f_1 f 1 of Y Y Y for y 1 y_1 y 1 : if x 2 > 0 x_2>0 x 2 > 0 and y 1 > y 2 y_1>y_2 y 1 > y 2 , the bound is attained if and only if X X X is indefinite and γ = S / q \gamma=S/q γ = S / q ; it is attained by real matrices.
Proof. z 3 z_3 z 3 is invariant under X ↦ − X X\mapsto-X X ↦ − X and vanishes for semidefinite X X X . Otherwise let X e 1 = x 1 e 1 Xe_1=x_1e_1 X e 1 = x 1 e 1 , X e 2 = − x 2 e 2 Xe_2=-x_2e_2 X e 2 = − x 2 e 2 , u = ⟨ e 1 , Y 2 e 1 ⟩ = γ y 1 2 + ( 1 − γ ) y 2 2 u=\langle e_1,Y^2e_1\rangle=\gamma y_1^2+(1-\gamma)y_2^2 u = ⟨ e 1 , Y 2 e 1 ⟩ = γ y 1 2 + ( 1 − γ ) y 2 2 and s = y 1 2 + y 2 2 s=y_1^2+y_2^2 s = y 1 2 + y 2 2 . Then T r ( Y ∣ X ∣ Y ) = T r ( ∣ X ∣ Y 2 ) = A : = ( x 1 − x 2 ) u + x 2 s \mathop{\mathrm{Tr}}(Y|X|Y)=\mathop{\mathrm{Tr}}(|X|Y^2)=A:=(x_1-x_2)u+x_2s Tr ( Y ∣ X ∣ Y ) = Tr ( ∣ X ∣ Y 2 ) = A := ( x 1 − x 2 ) u + x 2 s , and since det ( Y X Y ) = − x 1 x 2 y 1 2 y 2 2 ≤ 0 \det(YXY)=-x_1x_2y_1^2y_2^2\le0 det ( Y X Y ) = − x 1 x 2 y 1 2 y 2 2 ≤ 0 , T r ∣ Y X Y ∣ = T r ( Y X Y ) 2 − 4 det ( Y X Y ) = L 2 + K 2 \mathop{\mathrm{Tr}}|YXY|=\sqrt{\mathop{\mathrm{Tr}}(YXY)^2-4\det(YXY)}=\sqrt{L^2+K^2} Tr ∣ Y X Y ∣ = Tr ( Y X Y ) 2 − 4 det ( Y X Y ) = L 2 + K 2 with L = T r ( X Y 2 ) = ( x 1 + x 2 ) u − x 2 s L=\mathop{\mathrm{Tr}}(XY^2)=(x_1+x_2)u-x_2s L = Tr ( X Y 2 ) = ( x 1 + x 2 ) u − x 2 s and K = 2 x 1 x 2 y 1 y 2 K=2\sqrt{x_1x_2}\,y_1y_2 K = 2 x 1 x 2 y 1 y 2 . For the unit vector ( α , β ) = ( x 1 − x 2 , 2 x 1 x 2 ) / ( x 1 + x 2 ) (\alpha,\beta)=(x_1-x_2,\,2\sqrt{x_1x_2})/(x_1+x_2) ( α , β ) = ( x 1 − x 2 , 2 x 1 x 2 ) / ( x 1 + x 2 ) the Cauchy–Schwarz inequality gives L 2 + K 2 ≥ α L + β K \sqrt{L^2+K^2}\ge\alpha L+\beta K L 2 + K 2 ≥ α L + β K , and A − α L − β K = 2 x 1 x 2 x 1 + x 2 ( y 1 − y 2 ) 2 A-\alpha L-\beta K=\frac{2x_1x_2}{x_1+x_2}(y_1-y_2)^2 A − α L − β K = x 1 + x 2 2 x 1 x 2 ( y 1 − y 2 ) 2 for every u u u ; equality holds iff β L = α K \beta L=\alpha K β L = α K , i.e. u − y 2 2 = ( y 1 − y 2 ) S / ( x 1 + x 2 ) u-y_2^2=(y_1-y_2)S/(x_1+x_2) u − y 2 2 = ( y 1 − y 2 ) S / ( x 1 + x 2 ) , i.e. γ = S / q \gamma=S/q γ = S / q . Real maximisers: diagonal X X X and Y Y Y rotated by arccos γ \arccos\sqrt\gamma arccos γ . ◻
Theorem 6.9. Let M = ⨁ c = 1 C M k c ( K ) \mathcal{M}=\bigoplus_{c=1}^CM_{k_c}(\mathbb K) M = ⨁ c = 1 C M k c ( K ) , K ∈ { R , C } \mathbb K\in\{\mathbb{R},\mathbb{C}\} K ∈ { R , C } , with ν ≥ 1 \nu\ge1 ν ≥ 1 , and let ∣ X ∣ |X| ∣ X ∣ and Y Y Y be normalised to λ max = 1 \lambda_{\max}=1 λ m a x = 1 , with O M C ( ∣ X ∣ ) = δ ∈ [ 0 , 1 ) \mathop{\mathrm{OMC}}(|X|)=\delta\in[0,1) OMC ( ∣ X ∣ ) = δ ∈ [ 0 , 1 ) and O M C ( Y ) = δ ′ ∈ [ 0 , 1 ] \mathop{\mathrm{OMC}}(Y)=\delta'\in[0,1] OMC ( Y ) = δ ′ ∈ [ 0 , 1 ] . Then sup ( T r ( Y ∣ X ∣ Y ) − T r ∣ Y X Y ∣ ) = { ( ν − δ ) ( 2 δ ′ 1 + δ ′ ) 2 if every k c is even , ν ( 2 δ ′ 1 + δ ′ ) 2 if some k c is odd , \sup\bigl(\mathop{\mathrm{Tr}}(Y|X|Y)-\mathop{\mathrm{Tr}}|YXY|\bigr)=\begin{cases}(\nu-\delta)\bigl(\frac{2\delta'}{1+\delta'}\bigr)^2&\text{if every }k_c\text{ is even},\\[2pt]
\nu\bigl(\frac{2\delta'}{1+\delta'}\bigr)^2&\text{if some }k_c\text{ is odd},\end{cases} sup ( Tr ( Y ∣ X ∣ Y ) − Tr ∣ Y X Y ∣ ) = { ( ν − δ ) ( 1 + δ ′ 2 δ ′ ) 2 ν ( 1 + δ ′ 2 δ ′ ) 2 if every k c is even , if some k c is odd , and the supremum is attained by real direct sums of the 2 × 2 2\times2 2 × 2 maximisers of Proposition 6.8 and 1 × 1 1\times1 1 × 1 blocks.
For odd blocks the eigenvalue ℓ \ell ℓ of ∣ X ∣ |X| ∣ X ∣ can sit on a one-dimensional summand, where it costs nothing, so the supremum does not depend on δ \delta δ ; for k = 3 k=3 k = 3 , C = 1 C=1 C = 1 : ( 2 δ ′ / ( 1 + δ ′ ) ) 2 > ( 1 − δ ) ( 2 δ ′ / ( 1 + δ ′ ) ) 2 (2\delta'/(1+\delta'))^2>(1-\delta)(2\delta'/(1+\delta'))^2 ( 2 δ ′ / ( 1 + δ ′ ) ) 2 > ( 1 − δ ) ( 2 δ ′ / ( 1 + δ ′ ) ) 2 . The proof reduces to two-level spectra and, with Lemma 3.4 , to 2 × 2 2\times2 2 × 2 blocks, except in one case: the eigenvector of ∣ X ∣ |X| ∣ X ∣ for ℓ \ell ℓ may be spread over all two-dimensional blocks of its sector, so that X X X is not block diagonal. This case is settled by the following sharp bound, whose proof linearises the problem.
Proposition 6.10. Let X = X + ⊕ ( − 1 ) X=X_+\oplus(-1) X = X + ⊕ ( − 1 ) on C m ⊕ C m \mathbb{C}^m\oplus\mathbb{C}^m C m ⊕ C m , where 0 < X + ≤ 1 0<X_+\le1 0 < X + ≤ 1 has eigenvalues x 1 , … , x m x_1,\dots,x_m x 1 , … , x m ; let c θ = d i a g ( cos θ j ) c_\theta=\mathrm{diag}(\cos\theta_j) c θ = diag ( cos θ j ) , s θ = d i a g ( sin θ j ) s_\theta=\mathrm{diag}(\sin\theta_j) s θ = diag ( sin θ j ) , Q Q Q the projection onto the range of ( c θ s θ ) : C m → C 2 m \binom{c_\theta}{s_\theta}:\mathbb{C}^m\to\mathbb{C}^{2m} ( s θ c θ ) : C m → C 2 m , b ∈ [ 0 , 1 ) b\in[0,1) b ∈ [ 0 , 1 ) and Y = b + ( 1 − b ) Q Y=b+(1-b)Q Y = b + ( 1 − b ) Q (a two-level operator). Then T r ( Y ∣ X ∣ Y ) − T r ∣ Y X Y ∣ ≤ ( 1 − b ) 2 ∑ i = 1 m 2 x i 1 + x i , \mathop{\mathrm{Tr}}(Y|X|Y)-\mathop{\mathrm{Tr}}|YXY|\ \le\ (1-b)^2\sum_{i=1}^m\frac{2x_i}{1+x_i}, Tr ( Y ∣ X ∣ Y ) − Tr ∣ Y X Y ∣ ≤ ( 1 − b ) 2 ∑ i = 1 m 1 + x i 2 x i , with equality for X + = d i a g ( x 1 , … , x m ) X_+=\mathrm{diag}(x_1,\dots,x_m) X + = diag ( x 1 , … , x m ) and cos 2 θ i = ( 1 + b x i ) / ( ( 1 + x i ) ( 1 + b ) ) \cos^2\theta_i=(1+bx_i)/((1+x_i)(1+b)) cos 2 θ i = ( 1 + b x i ) / (( 1 + x i ) ( 1 + b )) .
Proof. Let γ = 1 − b 2 \gamma=1-b^2 γ = 1 − b 2 , K = 1 + X + K=1+X_+ K = 1 + X + , A = c θ 2 A=c_\theta^2 A = c θ 2 (so s θ 2 = 1 − A s_\theta^2=1-A s θ 2 = 1 − A ). In the decomposition C m ⊕ C m \mathbb{C}^m\oplus\mathbb{C}^m C m ⊕ C m , Y 2 = b 2 + γ Q Y^2=b^2+\gamma Q Y 2 = b 2 + γ Q has the blocks b 2 + γ A b^2+\gamma A b 2 + γ A , γ c θ s θ \gamma c_\theta s_\theta γ c θ s θ , γ c θ s θ \gamma c_\theta s_\theta γ c θ s θ , b 2 + γ s θ 2 b^2+\gamma s_\theta^2 b 2 + γ s θ 2 , all diagonal. For λ > 0 \lambda>0 λ > 0 the lower right block of λ − X Y 2 \lambda-XY^2 λ − X Y 2 is the invertible diagonal E = λ + 1 − γ A E=\lambda+1-\gamma A E = λ + 1 − γ A , and as ( b 2 + γ A ) E − γ 2 A s θ 2 = λ ( b 2 + γ A ) + b 2 (b^2+\gamma A)E-\gamma^2As_\theta^2=\lambda(b^2+\gamma A)+b^2 ( b 2 + γ A ) E − γ 2 A s θ 2 = λ ( b 2 + γ A ) + b 2 , the Schur complement formula gives
det ( λ − X Y 2 ) = det ( λ E − X + ( λ ( b 2 + γ A ) + b 2 ) ) = det ( ( λ + 1 ) ( λ − b 2 X + ) − λ γ K A ) = p ( λ ) , \det(\lambda-XY^2)=\det\bigl(\lambda E-X_+(\lambda(b^2+\gamma A)+b^2)\bigr)=\det\bigl((\lambda+1)(\lambda-b^2X_+)-\lambda\gamma KA\bigr)
=p(\lambda), det ( λ − X Y 2 ) = det ( λ E − X + ( λ ( b 2 + γ A ) + b 2 ) ) = det ( ( λ + 1 ) ( λ − b 2 X + ) − λγ K A ) = p ( λ ) , p ( λ ) = det ( ( λ + 1 ) ( λ − b 2 X + ) − λ γ K 1 / 2 A K 1 / 2 ) p(\lambda)=\det\bigl((\lambda+1)(\lambda-b^2X_+)-\lambda\gamma K^{1/2}AK^{1/2}\bigr) p ( λ ) = det ( ( λ + 1 ) ( λ − b 2 X + ) − λγ K 1/2 A K 1/2 ) . The Schur complement of the lower right block λ \lambda λ gives det ( λ − Ξ ) = p ( λ ) \det(\lambda-\Xi)=p(\lambda) det ( λ − Ξ ) = p ( λ ) as well, for the self-adjoint matrix
Ξ = ( γ K 1 / 2 A K 1 / 2 + b 2 X + − 1 b X + 1 / 2 b X + 1 / 2 0 ) , \Xi=\begin{pmatrix}\gamma K^{1/2}AK^{1/2}+b^2X_+-1 & bX_+^{1/2}\\ bX_+^{1/2} & 0\end{pmatrix}, Ξ = ( γ K 1/2 A K 1/2 + b 2 X + − 1 b X + 1/2 b X + 1/2 0 ) , which is affine in A A A . Since det ( λ − Y X Y ) = det ( λ − X Y 2 ) \det(\lambda-YXY)=\det(\lambda-XY^2) det ( λ − Y X Y ) = det ( λ − X Y 2 ) , the polynomials agree and Y X Y YXY Y X Y and Ξ \Xi Ξ have the same eigenvalues; hence T r ∣ Y X Y ∣ = ∥ Ξ ∥ 1 ≥ T r ( T Ξ ) \mathop{\mathrm{Tr}}|YXY|=\|\Xi\|_1\ge\mathop{\mathrm{Tr}}(T\Xi) Tr ∣ Y X Y ∣ = ∥Ξ ∥ 1 ≥ Tr ( T Ξ ) for every self-adjoint T T T with ∥ T ∥ ≤ 1 \|T\|\le1 ∥ T ∥ ≤ 1 . Take
T = ( − Λ Γ Γ Λ ) , Λ = 1 − X + 1 + X + , Γ = 2 X + 1 / 2 1 + X + , T=\begin{pmatrix}-\Lambda & \Gamma\\ \Gamma & \Lambda\end{pmatrix},\qquad \Lambda=\frac{1-X_+}{1+X_+},\quad \Gamma=\frac{2X_+^{1/2}}{1+X_+}, T = ( − Λ Γ Γ Λ ) , Λ = 1 + X + 1 − X + , Γ = 1 + X + 2 X + 1/2 , so that T 2 = 1 T^2=1 T 2 = 1 (Λ \Lambda Λ and Γ \Gamma Γ commute and Λ 2 + Γ 2 = 1 \Lambda^2+\Gamma^2=1 Λ 2 + Γ 2 = 1 ). With T r ( Y ∣ X ∣ Y ) = T r ( ∣ X ∣ Y 2 ) = b 2 T r X + + γ T r ( X + A ) + m b 2 + γ T r ( 1 − A ) \mathop{\mathrm{Tr}}(Y|X|Y)=\mathop{\mathrm{Tr}}(|X|Y^2)=b^2\mathop{\mathrm{Tr}}X_++\gamma\mathop{\mathrm{Tr}}(X_+A)+mb^2+\gamma\mathop{\mathrm{Tr}}(1-A) Tr ( Y ∣ X ∣ Y ) = Tr ( ∣ X ∣ Y 2 ) = b 2 Tr X + + γ Tr ( X + A ) + m b 2 + γ Tr ( 1 − A ) and K 1 / 2 Λ K 1 / 2 = 1 − X + K^{1/2}\Lambda K^{1/2}=1-X_+ K 1/2 Λ K 1/2 = 1 − X + , the terms containing A A A cancel in T r ( ∣ X ∣ Y 2 ) − T r ( T Ξ ) \mathop{\mathrm{Tr}}(|X|Y^2)-\mathop{\mathrm{Tr}}(T\Xi) Tr ( ∣ X ∣ Y 2 ) − Tr ( T Ξ ) , and
T r ( Y ∣ X ∣ Y ) − T r ∣ Y X Y ∣ ≤ ∑ i = 1 m ( b 2 x i + 1 + 1 − x i 1 + x i ( b 2 x i − 1 ) − 4 b x i 1 + x i ) = ( 1 − b ) 2 ∑ i = 1 m 2 x i 1 + x i . \mathop{\mathrm{Tr}}(Y|X|Y)-\mathop{\mathrm{Tr}}|YXY|\le\sum_{i=1}^m\Bigl(b^2x_i+1+\frac{1-x_i}{1+x_i}(b^2x_i-1)-\frac{4bx_i}{1+x_i}\Bigr)=(1-b)^2\sum_{i=1}^m\frac{2x_i}{1+x_i}. Tr ( Y ∣ X ∣ Y ) − Tr ∣ Y X Y ∣ ≤ i = 1 ∑ m ( b 2 x i + 1 + 1 + x i 1 − x i ( b 2 x i − 1 ) − 1 + x i 4 b x i ) = ( 1 − b ) 2 i = 1 ∑ m 1 + x i 2 x i . For diagonal X + X_+ X + both operators are direct sums of 2 × 2 2\times2 2 × 2 blocks, on which X = d i a g ( x i , − 1 ) X=\mathrm{diag}(x_i,-1) X = diag ( x i , − 1 ) , and Proposition 6.8 gives the value ( 1 − b ) 2 2 x i / ( 1 + x i ) (1-b)^2\,2x_i/(1+x_i) ( 1 − b ) 2 2 x i / ( 1 + x i ) at the stated angle (there γ = sin 2 θ i \gamma=\sin^2\theta_i γ = sin 2 θ i ). ◻
Lemma 6.11. Let X = X ∗ X=X^* X = X ∗ be invertible, P P P the projection onto the positive part of X X X , S = 2 P − 1 S=2P-1 S = 2 P − 1 , and Y ≥ 0 Y\ge0 Y ≥ 0 . Then (a) if ∣ X ∣ ≤ ∣ X ′ ∣ |X|\le|X'| ∣ X ∣ ≤ ∣ X ′ ∣ , X ′ = S ∣ X ′ ∣ X'=S|X'| X ′ = S ∣ X ′ ∣ and ∣ X ′ ∣ |X'| ∣ X ′ ∣ commutes with P P P , then z 3 ( X , Y ) ≤ z 3 ( X ′ , Y ) z_3(X,Y)\le z_3(X',Y) z 3 ( X , Y ) ≤ z 3 ( X ′ , Y ) ; (b) for fixed X X X , z 3 z_3 z 3 is a convex function of Y 2 Y^2 Y 2 .
Proof. With A = Y X + Y A=YX_+Y A = Y X + Y and B = Y X − Y B=YX_-Y B = Y X − Y , z 3 = T r A + T r B − ∥ A − B ∥ 1 z_3=\mathop{\mathrm{Tr}}A+\mathop{\mathrm{Tr}}B-\|A-B\|_1 z 3 = Tr A + Tr B − ∥ A − B ∥ 1 . (a) A = Y ∣ X ∣ 1 / 2 P ∣ X ∣ 1 / 2 Y A=Y|X|^{1/2}P|X|^{1/2}Y A = Y ∣ X ∣ 1/2 P ∣ X ∣ 1/2 Y is non-decreasing in ∣ X ∣ |X| ∣ X ∣ among operators commuting with P P P , and so is B B B ; by the triangle inequality ∥ A + D − B ∥ 1 ≤ ∥ A − B ∥ 1 + T r D \|A+D-B\|_1\le\|A-B\|_1+\mathop{\mathrm{Tr}}D ∥ A + D − B ∥ 1 ≤ ∥ A − B ∥ 1 + Tr D for D ≥ 0 D\ge0 D ≥ 0 , so z 3 z_3 z 3 does not decrease when A A A or B B B increases. (b) As ∥ H ∥ 1 = max − 1 ≤ T ≤ 1 T r ( T H ) \|H\|_1=\max_{-1\le T\le1}\mathop{\mathrm{Tr}}(TH) ∥ H ∥ 1 = max − 1 ≤ T ≤ 1 Tr ( T H ) for self-adjoint H H H , with T = 1 − 2 E T=1-2E T = 1 − 2 E we get z 3 = 2 min 0 ≤ E ≤ 1 T r ( E A + ( 1 − E ) B ) z_3=2\min_{0\le E\le1}\mathop{\mathrm{Tr}}\bigl(EA+(1-E)B\bigr) z 3 = 2 min 0 ≤ E ≤ 1 Tr ( E A + ( 1 − E ) B ) . For invertible Y Y Y , Z = ∣ X ∣ 1 / 2 Y Z=|X|^{1/2}Y Z = ∣ X ∣ 1/2 Y is invertible and E ↦ F = Z E Z ∗ E\mapsto F=ZEZ^* E ↦ F = Z E Z ∗ maps { 0 ≤ E ≤ 1 } \{0\le E\le1\} { 0 ≤ E ≤ 1 } onto { 0 ≤ F ≤ Γ } \{0\le F\le \Gamma\} { 0 ≤ F ≤ Γ } , Γ = Z Z ∗ = ∣ X ∣ 1 / 2 Y 2 ∣ X ∣ 1 / 2 \Gamma=ZZ^*=|X|^{1/2}Y^2|X|^{1/2} Γ = Z Z ∗ = ∣ X ∣ 1/2 Y 2 ∣ X ∣ 1/2 , with T r ( E A ) = T r ( P F ) \mathop{\mathrm{Tr}}(EA)=\mathop{\mathrm{Tr}}(PF) Tr ( E A ) = Tr ( P F ) and T r ( ( 1 − E ) B ) = T r ( ( 1 − P ) ( Γ − F ) ) \mathop{\mathrm{Tr}}((1-E)B)=\mathop{\mathrm{Tr}}((1-P)(\Gamma-F)) Tr (( 1 − E ) B ) = Tr (( 1 − P ) ( Γ − F )) . So z 3 = 2 min { T r ( P F + ( 1 − P ) ( Γ − F ) ) : 0 ≤ F ≤ Γ } z_3=2\min\{\mathop{\mathrm{Tr}}(PF+(1-P)(\Gamma-F)):0\le F\le \Gamma\} z 3 = 2 min { Tr ( P F + ( 1 − P ) ( Γ − F )) : 0 ≤ F ≤ Γ } is the partial minimum of a linear function over a convex set of pairs ( F , Γ ) (F,\Gamma) ( F , Γ ) , hence convex in Γ \Gamma Γ , which is affine in Y 2 Y^2 Y 2 ; by continuity also at singular Y Y Y . ◻
Proof of Theorem 6.9 . Write a = ℓ a=\ell a = ℓ , b = ℓ ′ b=\ell' b = ℓ ′ ; it suffices to treat the complex algebra, since the maximisers are real. Let δ < 1 \delta<1 δ < 1 , so that X X X is invertible; z 3 ( − X , Y ) = z 3 ( X , Y ) z_3(-X,Y)=z_3(X,Y) z 3 ( − X , Y ) = z 3 ( X , Y ) . Since ∣ X ∣ |X| ∣ X ∣ commutes with P P P , it has a unit eigenvector e e e for its eigenvalue a a a in one sector and in r a n P \mathrm{ran}\,P ran P or in ker P \ker P ker P ; replacing X X X by − X -X − X if necessary, e ∈ r a n P e\in\mathrm{ran}\,P e ∈ ran P . Then ∣ X ∣ ≤ ∣ X ′ ∣ : = 1 − ( 1 − a ) e e ∗ |X|\le|X'|:=1-(1-a)ee^* ∣ X ∣ ≤ ∣ X ′ ∣ := 1 − ( 1 − a ) e e ∗ , and X ′ = S ∣ X ′ ∣ = S − ( 1 − a ) e e ∗ X'=S|X'|=S-(1-a)ee^* X ′ = S ∣ X ′ ∣ = S − ( 1 − a ) e e ∗ has contrast δ \delta δ ; by Lemma 6.11 (a), z 3 ( X , Y ) ≤ z 3 ( X ′ , Y ) z_3(X,Y)\le z_3(X',Y) z 3 ( X , Y ) ≤ z 3 ( X ′ , Y ) . By Lemma 6.11 (b), a convex function of the eigenvalues of Y 2 Y^2 Y 2 on the faces { y i = 1 , y j = b 2 } \{y_i=1,\ y_j=b^2\} { y i = 1 , y j = b 2 } of [ b 2 , 1 ] K [b^2,1]^K [ b 2 , 1 ] K is maximal at a vertex, so z 3 ( X ′ , Y ) ≤ z 3 ( X ′ , Y ′ ) z_3(X',Y)\le z_3(X',Y') z 3 ( X ′ , Y ) ≤ z 3 ( X ′ , Y ′ ) with Y ′ = b + ( 1 − b ) Q Y'=b+(1-b)Q Y ′ = b + ( 1 − b ) Q , Q ≠ 0 , 1 Q\ne0,1 Q = 0 , 1 a projection in M \mathcal{M} M . Apply Lemma 3.4 to P P P and Q Q Q in every sector. In a sector without e e e , X ′ = S X'=S X ′ = S is a direct sum over the subspaces, a one-dimensional summand contributes 0 0 0 , and a two-dimensional one at most ( 1 − b ) 2 (1-b)^2 ( 1 − b ) 2 by Proposition 6.8 (∣ X ∣ = 1 |X|=1 ∣ X ∣ = 1 there): together at most ⌊ k c / 2 ⌋ ( 1 − b ) 2 \lfloor k_c/2\rfloor(1-b)^2 ⌊ k c /2 ⌋ ( 1 − b ) 2 . In the sector of e e e , Lemma 6.11 (a) with ∣ X ′ ∣ ≤ 1 |X'|\le1 ∣ X ′ ∣ ≤ 1 gives the same bound for z 3 z_3 z 3 restricted to it, and, if its size k c k_c k c is odd or it contains fewer than k c / 2 k_c/2 k c /2 two-dimensional subspaces, at most ⌊ k c / 2 ⌋ ( 1 − b ) 2 \lfloor k_c/2\rfloor(1-b)^2 ⌊ k c /2 ⌋ ( 1 − b ) 2 , resp. ( k c / 2 − 1 ) ( 1 − b ) 2 (k_c/2-1)(1-b)^2 ( k c /2 − 1 ) ( 1 − b ) 2 . Otherwise its m = k c / 2 m=k_c/2 m = k c /2 two-dimensional subspaces fill it, with bases f j , f j ′ f_j,f_j' f j , f j ′ , f j ∈ r a n P f_j\in\mathrm{ran}\,P f j ∈ ran P ; the unitary that multiplies f j ′ f_j' f j ′ by e − i χ j e^{-i\chi_j} e − i χ j commutes with X ′ X' X ′ (which is − 1 -1 − 1 on ker P \ker P ker P ) and removes the phases, and in the basis ( f 1 , … , f m , f 1 ′ , … , f m ′ ) (f_1,\dots,f_m,f_1',\dots,f_m') ( f 1 , … , f m , f 1 ′ , … , f m ′ ) we are in the setting of Proposition 6.10 with X + = 1 − ( 1 − a ) e e ∗ X_+=1-(1-a)ee^* X + = 1 − ( 1 − a ) e e ∗ , whose eigenvalues are a , 1 , … , 1 a,1,\dots,1 a , 1 , … , 1 : the sector contributes at most ( 1 − b ) 2 ( 2 a 1 + a + m − 1 ) = ( k c / 2 − δ ) ( 1 − b ) 2 (1-b)^2\bigl(\frac{2a}{1+a}+m-1\bigr)=(k_c/2-\delta)(1-b)^2 ( 1 − b ) 2 ( 1 + a 2 a + m − 1 ) = ( k c /2 − δ ) ( 1 − b ) 2 . Summing over the sectors, z 3 ≤ ( ν − δ ) ( 1 − b ) 2 z_3\le(\nu-\delta)(1-b)^2 z 3 ≤ ( ν − δ ) ( 1 − b ) 2 if every k c k_c k c is even and z 3 ≤ ν ( 1 − b ) 2 z_3\le\nu(1-b)^2 z 3 ≤ ν ( 1 − b ) 2 otherwise. Equality: if every k c k_c k c is even, X = d i a g ( a , − 1 ) ⊕ d i a g ( 1 , − 1 ) ⊕ ( ν − 1 ) X=\mathrm{diag}(a,-1)\oplus\mathrm{diag}(1,-1)^{\oplus(\nu-1)} X = diag ( a , − 1 ) ⊕ diag ( 1 , − 1 ) ⊕ ( ν − 1 ) and Y Y Y with eigenvalues 1 , b 1,b 1 , b in every block at the angle of Proposition 6.8 ; if some k c k_c k c is odd, X = d i a g ( 1 , − 1 ) X=\mathrm{diag}(1,-1) X = diag ( 1 , − 1 ) on ν \nu ν blocks, the eigenvalue a a a on a one-dimensional block of an odd sector, 1 1 1 on the other one-dimensional blocks, and Y Y Y as before (Y = 1 Y=1 Y = 1 on the one-dimensional blocks). ◻
Before this proof was found, the cases k = 4 k=4 k = 4 and k = 6 k=6 k = 6 (C = 1 C=1 C = 1 ) had been verified by a rigorous branch-and-bound computation over the angles of the reduced model (with gaps 10 − 7 10^{-7} 1 0 − 7 and 10 − 5 10^{-5} 1 0 − 5 that include floating-point rounding), and numerical optimisation reproduces all suprema of z 2 z_2 z 2 and z 3 z_3 z 3 . Theorems 6.4 and 6.9 and Proposition 6.6 thus settle the three functionals of [85 Sect. 5] with both contrasts prescribed. Like Gardner’s violation, they are additive over the blocks: they count the mutually orthogonal two-dimensional subspaces that fit into the sectors.
Characterisations of central elements
In this section M \mathcal{M} M is an arbitrary von Neumann algebra, unless finiteness is stated, and a ∈ M + + a\in\mathcal{M}_{++} a ∈ M ++ . Normal functionals have the polar decomposition φ = ∣ φ ∣ ( ⋅ u ) \varphi=|\varphi|(\cdot\,u) φ = ∣ φ ∣ ( ⋅ u ) , i.e. φ ( x ) = ∣ φ ∣ ( x u ) \varphi(x)=|\varphi|(xu) φ ( x ) = ∣ φ ∣ ( xu ) , with ∣ φ ∣ ≥ 0 |\varphi|\ge0 ∣ φ ∣ ≥ 0 and a partial isometry u u u with u ∗ u = s ( ∣ φ ∣ ) u^*u=s(|\varphi|) u ∗ u = s ( ∣ φ ∣ ) , and this decomposition is unique [162 Ch. III, §4] ; then ∣ φ ∣ ( x ) = φ ( x u ∗ ) |\varphi|(x)=\varphi(xu^*) ∣ φ ∣ ( x ) = φ ( x u ∗ ) . Hermitian normal functionals have a unique Jordan decomposition φ = φ + − φ − \varphi=\varphi^+-\varphi^- φ = φ + − φ − with ∥ φ ∥ = ∥ φ + ∥ + ∥ φ − ∥ \|\varphi\|=\|\varphi^+\|+\|\varphi^-\| ∥ φ ∥ = ∥ φ + ∥ + ∥ φ − ∥ , and then ∣ φ ∣ = φ + + φ − |\varphi|=\varphi^++\varphi^- ∣ φ ∣ = φ + + φ − [162 Ch. III, §4] . (The other convention φ = ∣ φ ∣ ( u ⋅ ) \varphi=|\varphi|(u\,\cdot) φ = ∣ φ ∣ ( u ⋅ ) is interchanged with ours by φ ↦ φ ∗ \varphi\mapsto\varphi^* φ ↦ φ ∗ , where φ ∗ ( x ) = φ ( x ∗ ) ‾ \varphi^*(x)=\overline{\varphi(x^*)} φ ∗ ( x ) = φ ( x ∗ ) ; this does not change any of the quantities below.) For z ∈ Z + + z\in\mathcal{Z}_{++} z ∈ Z ++ put φ z = φ ( z ⋅ ) \varphi_z=\varphi(z\,\cdot) φ z = φ ( z ⋅ ) .
Lemma 7.1. For z ∈ Z + + z\in\mathcal{Z}_{++} z ∈ Z ++ , φ ↦ φ z \varphi\mapsto\varphi_z φ ↦ φ z is a bijection of the normal (Hermitian, positive) functionals, ∣ φ z ∣ = ∣ φ ∣ z |\varphi_z|=|\varphi|_z ∣ φ z ∣ = ∣ φ ∣ z , ( φ z ) ± = ( φ ± ) z (\varphi_z)^\pm=(\varphi^\pm)_z ( φ z ) ± = ( φ ± ) z , and ∣ φ z + ω z ∣ = ∣ φ + ω ∣ z |\varphi_z+\omega_z|=|\varphi+\omega|_z ∣ φ z + ω z ∣ = ∣ φ + ω ∣ z .
Proof. φ z ( x ) = ∣ φ ∣ ( z x u ) = ∣ φ ∣ z ( x u ) \varphi_z(x)=|\varphi|(zxu)=|\varphi|_z(xu) φ z ( x ) = ∣ φ ∣ ( z xu ) = ∣ φ ∣ z ( xu ) and s ( ∣ φ ∣ z ) = s ( ∣ φ ∣ ) = u ∗ u s(|\varphi|_z)=s(|\varphi|)=u^*u s ( ∣ φ ∣ z ) = s ( ∣ φ ∣ ) = u ∗ u , since z z z is central and invertible; uniqueness gives ∣ φ z ∣ = ∣ φ ∣ z |\varphi_z|=|\varphi|_z ∣ φ z ∣ = ∣ φ ∣ z . The other claims are similar. ◻
Theorem 7.2 (Item 40). λ ( a ) : = min { λ : p a p ≤ λ a for all projections p ∈ M } = sup p ∥ a 1 / 2 p a − 1 / 2 ∥ 2 = cosh 2 L N C ( a ) \lambda(a):=\min\{\lambda:pap\le\lambda a\text{ for all projections }p\in\mathcal{M}\}=\sup_p\|a^{1/2}pa^{-1/2}\|^2=\cosh^2\mathop{\mathrm{LNC}}(a) λ ( a ) := min { λ : p a p ≤ λa for all projections p ∈ M } = sup p ∥ a 1/2 p a − 1/2 ∥ 2 = cosh 2 LNC ( a ) . For M = M n ( C ) \mathcal{M}=M_n(\mathbb{C}) M = M n ( C ) the upper bound is contained in [113 Cor. 1.5] .
Proof. p a p ≤ λ a pap\le\lambda a p a p ≤ λa iff ∥ a − 1 / 2 p a p a − 1 / 2 ∥ ≤ λ \|a^{-1/2}papa^{-1/2}\|\le\lambda ∥ a − 1/2 p a p a − 1/2 ∥ ≤ λ , and a − 1 / 2 p a p a − 1 / 2 = ( a 1 / 2 p a − 1 / 2 ) ∗ ( a 1 / 2 p a − 1 / 2 ) a^{-1/2}papa^{-1/2}=(a^{1/2}pa^{-1/2})^*(a^{1/2}pa^{-1/2}) a − 1/2 p a p a − 1/2 = ( a 1/2 p a − 1/2 ) ∗ ( a 1/2 p a − 1/2 ) . For z ∈ Z + + z\in\mathcal{Z}_{++} z ∈ Z ++ , p ( z a ) p = z 1 / 2 p a p z 1 / 2 p(za)p=z^{1/2}papz^{1/2} p ( z a ) p = z 1/2 p a p z 1/2 , so λ ( z a ) = λ ( a ) \lambda(za)=\lambda(a) λ ( z a ) = λ ( a ) , and Lemma 3.3 with T = z 0 a T=z_0a T = z 0 a gives λ ( a ) ≤ cosh 2 L N C ( a ) \lambda(a)\le\cosh^2\mathop{\mathrm{LNC}}(a) λ ( a ) ≤ cosh 2 LNC ( a ) . For the lower bound, if a ∉ Z a\notin\mathcal{Z} a ∈ / Z , use Lemma 4.14 c:b : the compression of a a a to K = span { η 1 , η 2 } K=\operatorname{span}\{\eta_1,\eta_2\} K = span { η 1 , η 2 } is D = d i a g ( a 1 , a 2 ) D=\mathop{\mathrm{diag}}(a_1,a_2) D = diag ( a 1 , a 2 ) with a i ∈ I i a_i\in I_i a i ∈ I i , and K K K is invariant under p = π ( ∣ w ⟩ ⟨ w ∣ ) p=\pi(|w\rangle\langle w|) p = π ( ∣ w ⟩ ⟨ w ∣ ) , w = ( 1 , 1 ) / 2 w=(1,1)/\sqrt2 w = ( 1 , 1 ) / 2 . For k = c 1 η 1 + c 2 η 2 k=c_1\eta_1+c_2\eta_2 k = c 1 η 1 + c 2 η 2 , ⟨ p a p k , k ⟩ = ∣ ⟨ c , w ⟩ ∣ 2 ⟨ D w , w ⟩ \langle papk,k\rangle=|\langle c,w\rangle|^2\langle Dw,w\rangle ⟨ p a p k , k ⟩ = ∣ ⟨ c , w ⟩ ∣ 2 ⟨ D w , w ⟩ and ⟨ a k , k ⟩ = ⟨ D c , c ⟩ \langle ak,k\rangle=\langle Dc,c\rangle ⟨ ak , k ⟩ = ⟨ D c , c ⟩ ; the supremum over c c c of the ratio is ⟨ D w , w ⟩ ⟨ D − 1 w , w ⟩ = ( a 1 + a 2 ) 2 / ( 4 a 1 a 2 ) \langle Dw,w\rangle\langle D^{-1}w,w\rangle=(a_1+a_2)^2/(4a_1a_2) ⟨ D w , w ⟩ ⟨ D − 1 w , w ⟩ = ( a 1 + a 2 ) 2 / ( 4 a 1 a 2 ) (the Kantorovich expression, which here attains its maximum G ( D ) 2 G(D)^2 G ( D ) 2 ), which tends to cosh 2 L N C ( a ) \cosh^2\mathop{\mathrm{LNC}}(a) cosh 2 LNC ( a ) as ε → 0 \varepsilon\to0 ε → 0 . ◻
Theorem 7.3 (Item 39). C ( a ) : = sup 0 ≠ φ ∈ M ∗ ∣ φ ( a ) ∣ ∣ φ ∣ ( a ) = cosh L N C ( a ) . C(a):=\sup_{0\ne\varphi\in\mathcal{M}_*}\frac{|\varphi(a)|}{|\varphi|(a)}=\cosh\mathop{\mathrm{LNC}}(a). C ( a ) := sup 0 = φ ∈ M ∗ ∣ φ ∣ ( a ) ∣ φ ( a ) ∣ = cosh LNC ( a ) . More strongly, ∣ φ ( a y ) ∣ ≤ cosh L N C ( a ) ∣ φ ∣ ( a ) |\varphi(ay)|\le\cosh\mathop{\mathrm{LNC}}(a)\,|\varphi|(a) ∣ φ ( a y ) ∣ ≤ cosh LNC ( a ) ∣ φ ∣ ( a ) for every φ ∈ M ∗ \varphi\in\mathcal{M}_* φ ∈ M ∗ and every contraction y ∈ M y\in\mathcal{M} y ∈ M .
Proof. Upper bound. Let ω = ∣ φ ∣ \omega=|\varphi| ω = ∣ φ ∣ and ( π ω , ξ ) (\pi_\omega,\xi) ( π ω , ξ ) its GNS pair. Then
∣ φ ( a y ) ∣ = ∣ ω ( a y u ) ∣ = ∣ ⟨ π ω ( y u ) ξ , π ω ( a ) ξ ⟩ ∣ ≤ ∥ ξ ∥ ∥ π ω ( a ) ξ ∥ ≤ cosh O L C ( a ) ω ( a ) |\varphi(ay)|=|\omega(ayu)|=|\langle\pi_\omega(yu)\xi,\pi_\omega(a)\xi\rangle|\le\|\xi\|\,\|\pi_\omega(a)\xi\|\le\cosh\mathop{\mathrm{OLC}}(a)\,\omega(a) ∣ φ ( a y ) ∣ = ∣ ω ( a y u ) ∣ = ∣ ⟨ π ω ( y u ) ξ , π ω ( a ) ξ ⟩ ∣ ≤ ∥ ξ ∥ ∥ π ω ( a ) ξ ∥ ≤ cosh OLC ( a ) ω ( a ) by Lemma 3.1 , since σ ( π ω ( a ) ) ⊂ σ ( a ) \sigma(\pi_\omega(a))\subset\sigma(a) σ ( π ω ( a )) ⊂ σ ( a ) . With z 0 z_0 z 0 as in Lemma 4.12 and by Lemma 7.1 ,
∣ φ ( a y ) ∣ = ∣ φ z 0 − 1 ( z 0 a y ) ∣ ≤ cosh O L C ( z 0 a ) ∣ φ z 0 − 1 ∣ ( z 0 a ) = cosh L N C ( a ) ∣ φ ∣ ( a ) . |\varphi(ay)|=|\varphi_{z_0^{-1}}(z_0ay)|\le\cosh\mathop{\mathrm{OLC}}(z_0a)\,|\varphi_{z_0^{-1}}|(z_0a)=\cosh\mathop{\mathrm{LNC}}(a)\,|\varphi|(a). ∣ φ ( a y ) ∣ = ∣ φ z 0 − 1 ( z 0 a y ) ∣ ≤ cosh OLC ( z 0 a ) ∣ φ z 0 − 1 ∣ ( z 0 a ) = cosh LNC ( a ) ∣ φ ∣ ( a ) . Lower bound (a ∉ Z a\notin\mathcal{Z} a ∈ / Z ). In Lemma 4.14 c:b let y ∈ C 2 y\in\mathbb{C}^2 y ∈ C 2 be a unit vector with ∣ y 1 ∣ 2 = a 2 / ( a 1 + a 2 ) |y_1|^2=a_2/(a_1+a_2) ∣ y 1 ∣ 2 = a 2 / ( a 1 + a 2 ) , ∣ y 2 ∣ 2 = a 1 / ( a 1 + a 2 ) |y_2|^2=a_1/(a_1+a_2) ∣ y 2 ∣ 2 = a 1 / ( a 1 + a 2 ) , x = D y / ∥ D y ∥ x=Dy/\|Dy\| x = D y /∥ D y ∥ , and w ∈ M 2 ( C ) w\in M_2(\mathbb{C}) w ∈ M 2 ( C ) unitary with w y = x wy=x w y = x . Put ζ = y 1 η 1 + y 2 η 2 \zeta=y_1\eta_1+y_2\eta_2 ζ = y 1 η 1 + y 2 η 2 , ω = ω ζ \omega=\omega_\zeta ω = ω ζ and φ = ω ( ⋅ π ( w ) s ( ω ) ) \varphi=\omega(\cdot\,\pi(w)s(\omega)) φ = ω ( ⋅ π ( w ) s ( ω )) . Since ζ ∈ e H \zeta\in e\mathcal{H} ζ ∈ e H with e = f 1 + f 2 ∈ M e=f_1+f_2\in\mathcal{M} e = f 1 + f 2 ∈ M , s ( ω ) ≤ e s(\omega)\le e s ( ω ) ≤ e , so u = π ( w ) s ( ω ) u=\pi(w)s(\omega) u = π ( w ) s ( ω ) satisfies u ∗ u = s ( ω ) u^*u=s(\omega) u ∗ u = s ( ω ) and ∣ φ ∣ = ω |\varphi|=\omega ∣ φ ∣ = ω . Since s ( ω ) ζ = ζ s(\omega)\zeta=\zeta s ( ω ) ζ = ζ and π ( w ) ζ = x 1 η 1 + x 2 η 2 \pi(w)\zeta=x_1\eta_1+x_2\eta_2 π ( w ) ζ = x 1 η 1 + x 2 η 2 ,
∣ φ ( a ) ∣ ∣ φ ∣ ( a ) = ∣ ⟨ D x , y ⟩ ∣ ⟨ D y , y ⟩ = ∥ D y ∥ ⟨ D y , y ⟩ = a 1 + a 2 2 a 1 a 2 , \frac{|\varphi(a)|}{|\varphi|(a)}=\frac{|\langle Dx,y\rangle|}{\langle Dy,y\rangle}=\frac{\|Dy\|}{\langle Dy,y\rangle}=\frac{a_1+a_2}{2\sqrt{a_1a_2}}, ∣ φ ∣ ( a ) ∣ φ ( a ) ∣ = ⟨ D y , y ⟩ ∣ ⟨ D x , y ⟩ ∣ = ⟨ D y , y ⟩ ∥ D y ∥ = 2 a 1 a 2 a 1 + a 2 , which tends to cosh L N C ( a ) \cosh\mathop{\mathrm{LNC}}(a) cosh LNC ( a ) . ◻
In the finite case, writing φ = τ ( ⋅ h ) \varphi=\tau(\cdot\,h) φ = τ ( ⋅ h ) , Theorem 7.3 is the L 1 L^1 L 1 -version of Theorem 5.1 with the roles reversed: the density a a a is fixed and the operator h h h varies, and the extremal configuration is the same.
Theorem 7.4 (Item 41). For the remaining items of [152 Thm. 1 and Cor. 1] :
sup φ + ( a ) / φ 1 ( a ) \sup\varphi^+(a)/\varphi_1(a) sup φ + ( a ) / φ 1 ( a ) over Hermitian normal φ = φ 1 − φ 2 \varphi=\varphi_1-\varphi_2 φ = φ 1 − φ 2 with φ i ≥ 0 \varphi_i\ge0 φ i ≥ 0 ; sup φ + ( a ) / ω + ( a ) \sup\varphi^+(a)/\omega^+(a) sup φ + ( a ) / ω + ( a ) over Hermitian normal φ ≤ ω \varphi\le\omega φ ≤ ω ; and sup ( φ + ω ) + ( a ) / ( φ + ( a ) + ω + ( a ) ) \sup(\varphi+\omega)^+(a)/(\varphi^+(a)+\omega^+(a)) sup ( φ + ω ) + ( a ) / ( φ + ( a ) + ω + ( a )) over Hermitian normal φ , ω \varphi,\omega φ , ω are all equal to cosh 2 L N C ( a ) \cosh^2\mathop{\mathrm{LNC}}(a) cosh 2 LNC ( a ) ;
sup ∣ φ + ω ∣ ( a ) / ( ∣ φ ∣ ( a ) + ∣ ω ∣ ( a ) ) = cosh L N C ( a ) \sup|\varphi+\omega|(a)/(|\varphi|(a)+|\omega|(a))=\cosh\mathop{\mathrm{LNC}}(a) sup ∣ φ + ω ∣ ( a ) / ( ∣ φ ∣ ( a ) + ∣ ω ∣ ( a )) = cosh LNC ( a ) , both over Hermitian and over all normal φ , ω \varphi,\omega φ , ω .
Proof. r:a Upper bounds: let λ = cosh 2 L N C ( a ) \lambda=\cosh^2\mathop{\mathrm{LNC}}(a) λ = cosh 2 LNC ( a ) and e = s ( φ + ) e=s(\varphi^+) e = s ( φ + ) . Then φ + ( x ) = φ ( e x e ) \varphi^+(x)=\varphi(exe) φ + ( x ) = φ ( e x e ) , and e a e ≤ λ a eae\le\lambda a e a e ≤ λa by Theorem 7.2 . Hence φ + ( a ) = φ ( e a e ) ≤ φ 1 ( e a e ) ≤ λ φ 1 ( a ) \varphi^+(a)=\varphi(eae)\le\varphi_1(eae)\le\lambda\varphi_1(a) φ + ( a ) = φ ( e a e ) ≤ φ 1 ( e a e ) ≤ λ φ 1 ( a ) ; for φ ≤ ω \varphi\le\omega φ ≤ ω , φ + ( a ) ≤ ω ( e a e ) ≤ ω + ( e a e ) ≤ λ ω + ( a ) \varphi^+(a)\le\omega(eae)\le\omega^+(eae)\le\lambda\omega^+(a) φ + ( a ) ≤ ω ( e a e ) ≤ ω + ( e a e ) ≤ λ ω + ( a ) ; and with e ′ = s ( ( φ + ω ) + ) e'=s((\varphi+\omega)^+) e ′ = s (( φ + ω ) + ) , ( φ + ω ) + ( a ) ≤ ( φ + + ω + ) ( e ′ a e ′ ) ≤ λ ( φ + ( a ) + ω + ( a ) ) (\varphi+\omega)^+(a)\le(\varphi^++\omega^+)(e'ae')\le\lambda(\varphi^+(a)+\omega^+(a)) ( φ + ω ) + ( a ) ≤ ( φ + + ω + ) ( e ′ a e ′ ) ≤ λ ( φ + ( a ) + ω + ( a )) . Lower bound: the first supremum is a special case of the other two (ω = φ 1 \omega=\varphi_1 ω = φ 1 , resp. φ 1 \varphi_1 φ 1 and − φ 2 -\varphi_2 − φ 2 ), so it suffices to treat it. With the notation of Lemma 4.14 c:b , for a Hermitian 2 × 2 2\times2 2 × 2 matrix h h h let χ h ( x ) = ∑ i , j h i j ⟨ x η i , η j ⟩ = T r ( h C x ) \chi_h(x)=\sum_{i,j}h_{ij}\langle x\eta_i,\eta_j\rangle=\mathop{\mathrm{Tr}}(hC_x) χ h ( x ) = ∑ i , j h ij ⟨ x η i , η j ⟩ = Tr ( h C x ) , where C x C_x C x is the matrix of the compression of x x x to K K K ; thus χ h ( a ) = T r ( h D ) \chi_h(a)=\mathop{\mathrm{Tr}}(hD) χ h ( a ) = Tr ( h D ) and χ ∣ g ⟩ ⟨ g ∣ = ω g 1 η 1 + g 2 η 2 \chi_{|g\rangle\langle g|}=\omega_{g_1\eta_1+g_2\eta_2} χ ∣ g ⟩ ⟨ g ∣ = ω g 1 η 1 + g 2 η 2 . If h = μ + ∣ g + ⟩ ⟨ g + ∣ − μ − ∣ g − ⟩ ⟨ g − ∣ h=\mu_+|g_+\rangle\langle g_+|-\mu_-|g_-\rangle\langle g_-| h = μ + ∣ g + ⟩ ⟨ g + ∣ − μ − ∣ g − ⟩ ⟨ g − ∣ is the spectral decomposition, then χ h = μ + ω g + − μ − ω g − \chi_h=\mu_+\omega_{g_+}-\mu_-\omega_{g_-} χ h = μ + ω g + − μ − ω g − (vectors g ± g_\pm g ± read in K K K ), and the element x = π ( ∣ g + ⟩ ⟨ g + ∣ − ∣ g − ⟩ ⟨ g − ∣ ) x=\pi(|g_+\rangle\langle g_+|-|g_-\rangle\langle g_-|) x = π ( ∣ g + ⟩ ⟨ g + ∣ − ∣ g − ⟩ ⟨ g − ∣ ) of norm 1 1 1 gives χ h ( x ) = μ + + μ − \chi_h(x)=\mu_++\mu_- χ h ( x ) = μ + + μ − , so this is the Jordan decomposition. Take w = ( 1 , 1 ) / 2 w=(1,1)/\sqrt2 w = ( 1 , 1 ) / 2 , r ⊥ w r\perp w r ⊥ w , q = D − 1 w / ∥ D − 1 w ∥ q=D^{-1}w/\|D^{-1}w\| q = D − 1 w /∥ D − 1 w ∥ and h t = ∣ q ⟩ ⟨ q ∣ − t ∣ r ⟩ ⟨ r ∣ h_t=|q\rangle\langle q|-t|r\rangle\langle r| h t = ∣ q ⟩ ⟨ q ∣ − t ∣ r ⟩ ⟨ r ∣ , φ 1 = χ ∣ q ⟩ ⟨ q ∣ \varphi_1=\chi_{|q\rangle\langle q|} φ 1 = χ ∣ q ⟩ ⟨ q ∣ , φ 2 = t χ ∣ r ⟩ ⟨ r ∣ \varphi_2=t\chi_{|r\rangle\langle r|} φ 2 = t χ ∣ r ⟩ ⟨ r ∣ . Then φ + ( a ) / φ 1 ( a ) = T r ( D ( h t ) + ) / ⟨ D q , q ⟩ \varphi^+(a)/\varphi_1(a)=\mathop{\mathrm{Tr}}(D(h_t)_+)/\langle Dq,q\rangle φ + ( a ) / φ 1 ( a ) = Tr ( D ( h t ) + ) / ⟨ D q , q ⟩ . As t → ∞ t\to\infty t → ∞ , ( h t ) + → ∣ ⟨ q , w ⟩ ∣ 2 ∣ w ⟩ ⟨ w ∣ (h_t)_+\to|\langle q,w\rangle|^2|w\rangle\langle w| ( h t ) + → ∣ ⟨ q , w ⟩ ∣ 2 ∣ w ⟩ ⟨ w ∣ (in the basis w , r w,r w , r the positive eigenvalue of h t h_t h t tends to ∣ ⟨ q , w ⟩ ∣ 2 |\langle q,w\rangle|^2 ∣ ⟨ q , w ⟩ ∣ 2 and its eigenvector to w w w ), and the ratio tends to the Kantorovich expression ⟨ D − 1 w , w ⟩ ⟨ D w , w ⟩ = ( a 1 + a 2 ) 2 / ( 4 a 1 a 2 ) \langle D^{-1}w,w\rangle\langle Dw,w\rangle=(a_1+a_2)^2/(4a_1a_2) ⟨ D − 1 w , w ⟩ ⟨ D w , w ⟩ = ( a 1 + a 2 ) 2 / ( 4 a 1 a 2 ) .
r:b Upper bound: if χ = φ + ω = ∣ χ ∣ ( ⋅ u ) \chi=\varphi+\omega=|\chi|(\cdot\,u) χ = φ + ω = ∣ χ ∣ ( ⋅ u ) , then ∣ χ ∣ ( a ) = φ ( a u ∗ ) + ω ( a u ∗ ) ≤ cosh L N C ( a ) ( ∣ φ ∣ ( a ) + ∣ ω ∣ ( a ) ) |\chi|(a)=\varphi(au^*)+\omega(au^*)\le\cosh\mathop{\mathrm{LNC}}(a)(|\varphi|(a)+|\omega|(a)) ∣ χ ∣ ( a ) = φ ( a u ∗ ) + ω ( a u ∗ ) ≤ cosh LNC ( a ) ( ∣ φ ∣ ( a ) + ∣ ω ∣ ( a )) by Theorem 7.3 . Lower bound: let P ± P_\pm P ± be the projections of the proof of Theorem 5.4 built for D D D in place of d i a g ( κ , 1 ) \mathop{\mathrm{diag}}(\kappa,1) diag ( κ , 1 ) (i.e. s 2 = a 2 / ( a 1 + a 2 ) s^2=a_2/(a_1+a_2) s 2 = a 2 / ( a 1 + a 2 ) , c 2 = a 1 / ( a 1 + a 2 ) c^2=a_1/(a_1+a_2) c 2 = a 1 / ( a 1 + a 2 ) ), φ = χ P + \varphi=\chi_{P_+} φ = χ P + , ω = − χ P − \omega=-\chi_{P_-} ω = − χ P − . Then ∣ φ ∣ ( a ) + ∣ ω ∣ ( a ) = T r ( D ( P + + P − ) ) = 4 a 1 a 2 / ( a 1 + a 2 ) |\varphi|(a)+|\omega|(a)=\mathop{\mathrm{Tr}}(D(P_++P_-))=4a_1a_2/(a_1+a_2) ∣ φ ∣ ( a ) + ∣ ω ∣ ( a ) = Tr ( D ( P + + P − )) = 4 a 1 a 2 / ( a 1 + a 2 ) , and φ + ω = χ P + − P − \varphi+\omega=\chi_{P_+-P_-} φ + ω = χ P + − P − with P + − P − = 2 s c σ P_+-P_-=2sc\,\sigma P + − P − = 2 sc σ , whose Jordan decomposition (as above) gives ∣ φ + ω ∣ = 2 s c ( ω η 1 + ω η 2 ) |\varphi+\omega|=2sc(\omega_{\eta_1}+\omega_{\eta_2}) ∣ φ + ω ∣ = 2 sc ( ω η 1 + ω η 2 ) . The ratio is 2 s c ( a 1 + a 2 ) 2 / ( 4 a 1 a 2 ) = ( a 1 + a 2 ) / ( 2 a 1 a 2 ) 2sc(a_1+a_2)^2/(4a_1a_2)=(a_1+a_2)/(2\sqrt{a_1a_2}) 2 sc ( a 1 + a 2 ) 2 / ( 4 a 1 a 2 ) = ( a 1 + a 2 ) / ( 2 a 1 a 2 ) . ◻
The convexity items (vii) and (viii) of [152 Remark 1] are not treated here.
Theorem (Item 42).
If M \mathcal{M} M is finite and A ∈ M + + A\in\mathcal{M}_{++} A ∈ M ++ , then sup X ≠ 0 τ ( ∣ A X A ∣ ) τ ( A ∣ X ∣ A ) = κ Z ( A ) , inf X ≠ 0 τ ( ∣ A X A ∣ ) τ ( A ∣ X ∣ A ) = κ Z ( A ) − 1 . \sup_{X\ne0}\frac{\tau(|AXA|)}{\tau(A|X|A)}=\kappa_\mathcal{Z}(A),\qquad
\inf_{X\ne0}\frac{\tau(|AXA|)}{\tau(A|X|A)}=\kappa_\mathcal{Z}(A)^{-1}. sup X = 0 τ ( A ∣ X ∣ A ) τ ( ∣ A X A ∣ ) = κ Z ( A ) , inf X = 0 τ ( A ∣ X ∣ A ) τ ( ∣ A X A ∣ ) = κ Z ( A ) − 1 .
For every normal φ ≠ 0 \varphi\ne0 φ = 0 , with a 1 / 2 φ a 1 / 2 = φ ( a 1 / 2 ⋅ a 1 / 2 ) a^{1/2}\varphi a^{1/2}=\varphi(a^{1/2}\cdot a^{1/2}) a 1/2 φ a 1/2 = φ ( a 1/2 ⋅ a 1/2 ) , κ Z ( a ) − 1 / 2 ≤ ∥ a 1 / 2 φ a 1 / 2 ∥ ∣ φ ∣ ( a ) ≤ κ Z ( a ) 1 / 2 , \kappa_\mathcal{Z}(a)^{-1/2}\le\frac{\|a^{1/2}\varphi a^{1/2}\|}{|\varphi|(a)}\le\kappa_\mathcal{Z}(a)^{1/2}, κ Z ( a ) − 1/2 ≤ ∣ φ ∣ ( a ) ∥ a 1/2 φ a 1/2 ∥ ≤ κ Z ( a ) 1/2 , and if M \mathcal{M} M is finite both bounds are sharp.
Proof. l:a Let X = U ∣ X ∣ X=U|X| X = U ∣ X ∣ . By Hölder’s inequality,
τ ( ∣ A X A ∣ ) ≤ ∥ A U ∣ X ∣ 1 / 2 ∥ 2 ∥ ∣ X ∣ 1 / 2 A ∥ 2 = τ ( A 2 ∣ X ∗ ∣ ) 1 / 2 τ ( A 2 ∣ X ∣ ) 1 / 2 , \tau(|AXA|)\le\|AU|X|^{1/2}\|_2\,\||X|^{1/2}A\|_2=\tau(A^2|X^*|)^{1/2}\tau(A^2|X|)^{1/2}, τ ( ∣ A X A ∣ ) ≤ ∥ A U ∣ X ∣ 1/2 ∥ 2 ∥∣ X ∣ 1/2 A ∥ 2 = τ ( A 2 ∣ X ∗ ∣ ) 1/2 τ ( A 2 ∣ X ∣ ) 1/2 , and Theorem (?) for the density A 2 A^2 A 2 gives τ ( A 2 ∣ X ∗ ∣ ) ≤ κ Z ( A 2 ) τ ( A 2 ∣ X ∣ ) \tau(A^2|X^*|)\le\kappa_\mathcal{Z}(A^2)\tau(A^2|X|) τ ( A 2 ∣ X ∗ ∣ ) ≤ κ Z ( A 2 ) τ ( A 2 ∣ X ∣ ) , where κ Z ( A 2 ) = κ Z ( A ) 2 \kappa_\mathcal{Z}(A^2)=\kappa_\mathcal{Z}(A)^2 κ Z ( A 2 ) = κ Z ( A ) 2 . For the infimum, ∣ X ∣ = U ∗ X |X|=U^*X ∣ X ∣ = U ∗ X gives
τ ( A ∣ X ∣ A ) = τ ( ( A U ∗ A − 1 ) ( A X A ) ) ≤ ∥ A U ∗ A − 1 ∥ τ ( ∣ A X A ∣ ) ≤ κ ( A ) τ ( ∣ A X A ∣ ) , \tau(A|X|A)=\tau((AU^*A^{-1})(AXA))\le\|AU^*A^{-1}\|\,\tau(|AXA|)\le\kappa(A)\,\tau(|AXA|), τ ( A ∣ X ∣ A ) = τ (( A U ∗ A − 1 ) ( A X A )) ≤ ∥ A U ∗ A − 1 ∥ τ ( ∣ A X A ∣ ) ≤ κ ( A ) τ ( ∣ A X A ∣ ) , and the ratio at ( z A , X ) (zA,X) ( z A , X ) equals that at ( A , z 2 X ) (A,z^2X) ( A , z 2 X ) for z ∈ Z + + z\in\mathcal{Z}_{++} z ∈ Z ++ , so κ ( A ) \kappa(A) κ ( A ) may be replaced by κ Z ( A ) \kappa_\mathcal{Z}(A) κ Z ( A ) . Sharpness (A ∉ Z A\notin\mathcal{Z} A ∈ / Z ): with Lemma 4.14 c:a ,c:c for A A A , X = v X=v X = v gives τ ( ∣ A v A ∣ ) ≥ τ ( v ∗ A v A ) ≥ ( α − 2 ε ) τ ( f 2 A f 2 ) ≥ ( α − 2 ε ) ( β − ε ) t \tau(|AvA|)\ge\tau(v^*AvA)\ge(\alpha-2\varepsilon)\tau(f_2Af_2)\ge(\alpha-2\varepsilon)(\beta-\varepsilon)t τ ( ∣ A v A ∣ ) ≥ τ ( v ∗ A v A ) ≥ ( α − 2 ε ) τ ( f 2 A f 2 ) ≥ ( α − 2 ε ) ( β − ε ) t and τ ( A ∣ v ∣ A ) = τ ( f 2 A 2 f 2 ) ≤ ( β + 2 ε ) 2 t \tau(A|v|A)=\tau(f_2A^2f_2)\le(\beta+2\varepsilon)^2t τ ( A ∣ v ∣ A ) = τ ( f 2 A 2 f 2 ) ≤ ( β + 2 ε ) 2 t ; X = v ∗ X=v^* X = v ∗ gives, since A v ∗ A = ( A v A ) ∗ Av^*A=(AvA)^* A v ∗ A = ( A v A ) ∗ , τ ( ∣ Y ∗ ∣ ) = τ ( ∣ Y ∣ ) \tau(|Y^*|)=\tau(|Y|) τ ( ∣ Y ∗ ∣ ) = τ ( ∣ Y ∣ ) and A v A = ( f 1 A ) ∗ ( v f 2 A ) AvA=(f_1A)^*(vf_2A) A v A = ( f 1 A ) ∗ ( v f 2 A ) , τ ( ∣ A v ∗ A ∣ ) = τ ( ∣ A v A ∣ ) ≤ ∥ A f 1 ∥ 2 ∥ v f 2 A ∥ 2 ≤ ( α + ε ) ( β + 2 ε ) t \tau(|Av^*A|)=\tau(|AvA|)\le\|Af_1\|_2\|vf_2A\|_2\le(\alpha+\varepsilon)(\beta+2\varepsilon)t τ ( ∣ A v ∗ A ∣ ) = τ ( ∣ A v A ∣ ) ≤ ∥ A f 1 ∥ 2 ∥ v f 2 A ∥ 2 ≤ ( α + ε ) ( β + 2 ε ) t and τ ( A ∣ v ∗ ∣ A ) = τ ( f 1 A 2 f 1 ) ≥ ( α − 2 ε ) 2 t \tau(A|v^*|A)=\tau(f_1A^2f_1)\ge(\alpha-2\varepsilon)^2t τ ( A ∣ v ∗ ∣ A ) = τ ( f 1 A 2 f 1 ) ≥ ( α − 2 ε ) 2 t .
l:b Let χ = a 1 / 2 φ a 1 / 2 \chi=a^{1/2}\varphi a^{1/2} χ = a 1/2 φ a 1/2 and ω = ∣ φ ∣ \omega=|\varphi| ω = ∣ φ ∣ . By the Cauchy–Schwarz inequality,
∣ χ ( x ) ∣ = ∣ ω ( a 1 / 2 x a 1 / 2 u ) ∣ ≤ ω ( a 1 / 2 x x ∗ a 1 / 2 ) 1 / 2 ω ( u ∗ a u ) 1 / 2 ≤ ∥ x ∥ ω ( a ) 1 / 2 ( ∥ a ∥ ω ( 1 ) ) 1 / 2 , |\chi(x)|=|\omega(a^{1/2}xa^{1/2}u)|\le\omega(a^{1/2}xx^*a^{1/2})^{1/2}\omega(u^*au)^{1/2}\le\|x\|\,\omega(a)^{1/2}(\|a\|\omega(1))^{1/2}, ∣ χ ( x ) ∣ = ∣ ω ( a 1/2 x a 1/2 u ) ∣ ≤ ω ( a 1/2 x x ∗ a 1/2 ) 1/2 ω ( u ∗ a u ) 1/2 ≤ ∥ x ∥ ω ( a ) 1/2 ( ∥ a ∥ ω ( 1 ) ) 1/2 , and ω ( 1 ) ≤ ∥ a − 1 ∥ ω ( a ) \omega(1)\le\|a^{-1}\|\omega(a) ω ( 1 ) ≤ ∥ a − 1 ∥ ω ( a ) , so ∥ χ ∥ ≤ κ ( a ) 1 / 2 ∣ φ ∣ ( a ) \|\chi\|\le\kappa(a)^{1/2}|\varphi|(a) ∥ χ ∥ ≤ κ ( a ) 1/2 ∣ φ ∣ ( a ) . Conversely ∣ φ ∣ ( a ) = φ ( a u ∗ ) = χ ( a 1 / 2 u ∗ a − 1 / 2 ) ≤ κ ( a ) 1 / 2 ∥ χ ∥ |\varphi|(a)=\varphi(au^*)=\chi(a^{1/2}u^*a^{-1/2})\le\kappa(a)^{1/2}\|\chi\| ∣ φ ∣ ( a ) = φ ( a u ∗ ) = χ ( a 1/2 u ∗ a − 1/2 ) ≤ κ ( a ) 1/2 ∥ χ ∥ . By Lemma 7.1 the ratio at ( z a , φ ) (za,\varphi) ( z a , φ ) equals the ratio at ( a , φ z ) (a,\varphi_z) ( a , φ z ) , so κ ( a ) \kappa(a) κ ( a ) may be replaced by κ Z ( a ) \kappa_\mathcal{Z}(a) κ Z ( a ) . If M \mathcal{M} M is finite and φ = τ ( ⋅ h ) \varphi=\tau(\cdot\,h) φ = τ ( ⋅ h ) with h = w ∣ h ∣ ∈ M h=w|h|\in\mathcal{M} h = w ∣ h ∣ ∈ M , then ∣ φ ∣ = τ ( ∣ h ∣ ⋅ ) |\varphi|=\tau(|h|\,\cdot) ∣ φ ∣ = τ ( ∣ h ∣ ⋅ ) , ∣ φ ∣ ( a ) = τ ( a 1 / 2 ∣ h ∣ a 1 / 2 ) |\varphi|(a)=\tau(a^{1/2}|h|a^{1/2}) ∣ φ ∣ ( a ) = τ ( a 1/2 ∣ h ∣ a 1/2 ) and ∥ χ ∥ = τ ( ∣ a 1 / 2 h a 1 / 2 ∣ ) \|\chi\|=\tau(|a^{1/2}ha^{1/2}|) ∥ χ ∥ = τ ( ∣ a 1/2 h a 1/2 ∣ ) , so sharpness follows from l:a with A = a 1 / 2 A=a^{1/2} A = a 1/2 and κ Z ( a 1 / 2 ) = κ Z ( a ) 1 / 2 \kappa_\mathcal{Z}(a^{1/2})=\kappa_\mathcal{Z}(a)^{1/2} κ Z ( a 1/2 ) = κ Z ( a ) 1/2 . ◻
Contrast versus noncentrality
Call a violation constant treated in finite von Neumann algebras in § 5 or in § 7 exact if it was shown to be f ( κ Z ) f(\kappa_\mathcal{Z}) f ( κ Z ) for a function f f f given in closed form; these are the constants of Theorems 5.1 , 5.4 , (?) , (?) , (?) (γ \gamma γ ), 5.19 , 5.27 , 5.32 , 7.2 , 7.3 , 7.4 , (?) (for finite M \mathcal{M} M ), of Corollaries 5.3 , 5.6 , 5.20 and of Propositions (?) , 5.16 and 5.33 ; for an infimum (as in Theorems (?) and (?) ) we use its reciprocal, which is a supremum as in Definition 4.11 . In all these cases f f f is continuous and strictly increasing on [ 1 , ∞ ) [1,\infty) [ 1 , ∞ ) with f ( 1 ) = 1 f(1)=1 f ( 1 ) = 1 (f ( 1 ) = 0 f(1)=0 f ( 1 ) = 0 for γ \gamma γ and for item 21). The constant of item 37 (Proposition 5.50 pb:a ) is also exact, equal to 1 1 1 on Z \mathcal{Z} Z and to + ∞ +\infty + ∞ elsewhere; parts cp:a and cp:b below concern the exact constants with finite f f f , and parts cp:c and cp:d hold for item 37 as well.
Theorem 8.1. Let ρ ∈ M + + \rho\in\mathcal{M}_{++} ρ ∈ M ++ (with M \mathcal{M} M finite for the constants of § 5 and of Theorem (?) ).
For every exact constant with finite f f f , c ( ρ ) = f ( κ Z ( ρ ) ) ≤ f ( κ ( ρ ) ) c(\rho)=f(\kappa_\mathcal{Z}(\rho))\le f(\kappa(\rho)) c ( ρ ) = f ( κ Z ( ρ )) ≤ f ( κ ( ρ )) : the bound given by the contrast is valid.
c ( ρ ) < f ( κ ( ρ ) ) c(\rho)<f(\kappa(\rho)) c ( ρ ) < f ( κ ( ρ )) if and only if max Ω ⌈ ρ ⌉ \max_\Omega\lceil \rho\rceil max Ω ⌈ ρ ⌉ and min Ω ⌊ ρ ⌋ \min_\Omega\lfloor \rho\rfloor min Ω ⌊ ρ ⌋ are not attained at a common point of Ω \Omega Ω . In a factor c ( ρ ) = f ( κ ( ρ ) ) c(\rho)=f(\kappa(\rho)) c ( ρ ) = f ( κ ( ρ )) .
The loss is unbounded: in M = M 2 ( C ) ⊕ M 2 ( C ) \mathcal{M}=M_2(\mathbb{C})\oplus M_2(\mathbb{C}) M = M 2 ( C ) ⊕ M 2 ( C ) with any faithful tracial state, ρ λ = d i a g ( 1 , 4 ) ⊕ λ d i a g ( 1 , 4 ) \rho_\lambda=\mathop{\mathrm{diag}}(1,4)\oplus\lambda\mathop{\mathrm{diag}}(1,4) ρ λ = diag ( 1 , 4 ) ⊕ λ diag ( 1 , 4 ) (λ ≥ 1 \lambda\ge1 λ ≥ 1 ) has Gardner constant cosh L N C ( ρ λ ) = 5 / 4 \cosh\mathop{\mathrm{LNC}}(\rho_\lambda)=5/4 cosh LNC ( ρ λ ) = 5/4 , while cosh O L C ( ρ λ ) = ( 1 + 4 λ ) / ( 2 4 λ ) → ∞ \cosh\mathop{\mathrm{OLC}}(\rho_\lambda)=(1+4\lambda)/(2\sqrt{4\lambda})\to\infty cosh OLC ( ρ λ ) = ( 1 + 4 λ ) / ( 2 4 λ ) → ∞ ; and every central ρ \rho ρ has all constants equal to their tracial value, whatever its contrast. (Items 11 and 24, which characterise the canonical trace, are not violation constants in this sense and are violated by every non-scalar central density, Remark 8.3 .)
Let M \mathcal{M} M be neither a factor nor abelian. For every t > 1 t>1 t > 1 there are ρ 1 , ρ 2 ∈ M + + \rho_1,\rho_2\in\mathcal{M}_{++} ρ 1 , ρ 2 ∈ M ++ with κ ( ρ 1 ) = κ ( ρ 2 ) = t \kappa(\rho_1)=\kappa(\rho_2)=t κ ( ρ 1 ) = κ ( ρ 2 ) = t , κ Z ( ρ 1 ) = 1 \kappa_\mathcal{Z}(\rho_1)=1 κ Z ( ρ 1 ) = 1 and κ Z ( ρ 2 ) = t \kappa_\mathcal{Z}(\rho_2)=t κ Z ( ρ 2 ) = t . Consequently, for every inequality treated in finite von Neumann algebras in § 5 and § 7 the violation constant is not a function of the contrast on M \mathcal{M} M , whereas every exact constant is a function of the noncentrality N C \mathop{\mathrm{NC}} NC (equivalently, of κ Z \kappa_\mathcal{Z} κ Z ).
Proof. cp:a κ Z ≤ κ \kappa_\mathcal{Z}\le\kappa κ Z ≤ κ and f f f is increasing. cp:b Let u = ⌈ ρ ⌉ u=\lceil \rho\rceil u = ⌈ ρ ⌉ , l = ⌊ ρ ⌋ l=\lfloor \rho\rfloor l = ⌊ ρ ⌋ . By Corollary 4.3 , κ = max u / min l \kappa=\max u/\min l κ = max u / min l , and κ Z = max ( u / l ) \kappa_\mathcal{Z}=\max(u/l) κ Z = max ( u / l ) . Since u ( ω ) ≤ max u u(\omega)\le\max u u ( ω ) ≤ max u and l ( ω ) ≥ min l l(\omega)\ge\min l l ( ω ) ≥ min l , u ( ω ) / l ( ω ) = κ u(\omega)/l(\omega)=\kappa u ( ω ) / l ( ω ) = κ iff u ( ω ) = max u u(\omega)=\max u u ( ω ) = max u and l ( ω ) = min l l(\omega)=\min l l ( ω ) = min l . As f f f is strictly increasing, c < f ( κ ) c<f(\kappa) c < f ( κ ) iff κ Z < κ \kappa_\mathcal{Z}<\kappa κ Z < κ . In a factor Ω \Omega Ω is a point. cp:c κ Z ( ρ λ ) = 4 \kappa_\mathcal{Z}(\rho_\lambda)=4 κ Z ( ρ λ ) = 4 by Proposition 4.5 ch:prop and κ ( ρ λ ) = 4 λ \kappa(\rho_\lambda)=4\lambda κ ( ρ λ ) = 4 λ ; cosh ( ln 2 ) = 5 / 4 \cosh(\ln2)=5/4 cosh ( ln 2 ) = 5/4 . For central ρ \rho ρ , κ Z = 1 \kappa_\mathcal{Z}=1 κ Z = 1 . cp:d Since M \mathcal{M} M is not a factor, there is a central projection P ≠ 0 , 1 P\ne0,1 P = 0 , 1 ; put ρ 1 = P + t ( 1 − P ) ∈ Z \rho_1=P+t(1-P)\in\mathcal{Z} ρ 1 = P + t ( 1 − P ) ∈ Z . Since M \mathcal{M} M is not abelian, there is a non-central projection e e e ; then c ( e ) c ( 1 − e ) ≠ 0 c(e)c(1-e)\ne0 c ( e ) c ( 1 − e ) = 0 (otherwise 1 − e ≤ 1 − c ( e ) 1-e\le1-c(e) 1 − e ≤ 1 − c ( e ) , i.e. c ( e ) ≤ e c(e)\le e c ( e ) ≤ e and e = c ( e ) e=c(e) e = c ( e ) ). Put ρ 2 = 1 + ( t − 1 ) e \rho_2=1+(t-1)e ρ 2 = 1 + ( t − 1 ) e . Its spectrum is { 1 , t } \{1,t\} { 1 , t } , and by Lemma 4.1 , ⌈ ρ 2 ⌉ = 1 + ( t − 1 ) c ( e ) \lceil \rho_2\rceil=1+(t-1)c(e) ⌈ ρ 2 ⌉ = 1 + ( t − 1 ) c ( e ) and ⌊ ρ 2 ⌋ = 1 + ( t − 1 ) ( 1 − c ( 1 − e ) ) \lfloor \rho_2\rfloor=1+(t-1)(1-c(1-e)) ⌊ ρ 2 ⌋ = 1 + ( t − 1 ) ( 1 − c ( 1 − e )) , so ⌈ ρ 2 ⌉ / ⌊ ρ 2 ⌋ = t \lceil \rho_2\rceil/\lfloor \rho_2\rfloor=t ⌈ ρ 2 ⌉ / ⌊ ρ 2 ⌋ = t on c ( e ) c ( 1 − e ) c(e)c(1-e) c ( e ) c ( 1 − e ) and κ Z ( ρ 2 ) = t \kappa_\mathcal{Z}(\rho_2)=t κ Z ( ρ 2 ) = t . Every such violation constant equals its tracial value (1 1 1 , and 0 0 0 for γ \gamma γ and for item 21) iff κ Z = 1 \kappa_\mathcal{Z}=1 κ Z = 1 , and exceeds it otherwise: for the exact ones this is the formula. For the others, namely Propositions (?) , 5.23 , 5.29 , 5.44 , 5.47 , 5.50 , 5.51 (with r > 1 r>1 r > 1 ), 5.43 , (?) and Theorem 5.40 , the upper bound equals 1 1 1 at κ Z = 1 \kappa_\mathcal{Z}=1 κ Z = 1 and the constant exceeds 1 1 1 for κ Z > 1 \kappa_\mathcal{Z}>1 κ Z > 1 . For the Peierls–Bogoliubov inequality in the form with φ ( B e A ) \varphi(Be^A) φ ( B e A ) the constant is 1 1 1 for central ρ \rho ρ and + ∞ +\infty + ∞ otherwise. Hence c ( ρ 1 ) c(\rho_1) c ( ρ 1 ) is the tracial value and c ( ρ 2 ) c(\rho_2) c ( ρ 2 ) exceeds it, although O M C ( ρ 1 ) = O M C ( ρ 2 ) \mathop{\mathrm{OMC}}(\rho_1)=\mathop{\mathrm{OMC}}(\rho_2) OMC ( ρ 1 ) = OMC ( ρ 2 ) . ◻
Conversely, in a factor the contrast determines every exact constant, and in an abelian algebra every functional is tracial and all constants equal their tracial values. Thus, for the exact constants, the contrast gives the correct value for every ρ \rho ρ precisely in factors (in an abelian algebra the constants equal their tracial value f ( 1 ) f(1) f ( 1 ) , which is < f ( κ ( ρ ) ) <f(\kappa(\rho)) < f ( κ ( ρ )) unless ρ \rho ρ is scalar), while the noncentrality gives it in every von Neumann algebra (every finite one for the constants of § 5 and of Theorem (?) ). For the Petz–Zemánek constants c k c_k c k , k ≥ 3 k\ge3 k ≥ 3 , even the noncentrality does not suffice: by Propositions 5.11 and 5.12 , c k c_k c k is not a function of κ Z = κ \kappa_\mathcal{Z}=\kappa κ Z = κ on M k ( C ) M_k(\mathbb{C}) M k ( C ) . Nor are the constants a a a , c exp c_{\exp} c e x p and c exp p r c^{\rm pr}_{\exp} c e x p pr functions of κ Z \kappa_\mathcal{Z} κ Z (Propositions 5.44 and 5.47 ). For the other constants for which only bounds are proved, whether they are functions of κ Z \kappa_\mathcal{Z} κ Z is open (§ 11 ).
Finite direct sums
Let M = ⨁ c = 1 C M k c \mathcal{M}=\bigoplus_{c=1}^CM_{k_c} M = ⨁ c = 1 C M k c , with the trace T r = ∑ c T r c \mathop{\mathrm{Tr}}=\sum_c\mathop{\mathrm{Tr}}_c Tr = ∑ c Tr c and φ ρ = T r ( ρ ⋅ ) \varphi_\rho=\mathop{\mathrm{Tr}}(\rho\,\cdot\,) φ ρ = Tr ( ρ ⋅ ) , ρ = ⨁ c ρ c ≥ 0 \rho=\bigoplus_c\rho_c\ge0 ρ = ⨁ c ρ c ≥ 0 . The traces on M \mathcal{M} M are the functionals ∑ c t c T r c \sum_ct_c\mathop{\mathrm{Tr}}_c ∑ c t c Tr c , t c ≥ 0 t_c\ge0 t c ≥ 0 ; so φ ρ \varphi_\rho φ ρ is a trace iff ρ \rho ρ is central, whatever its global contrast O M C ( ρ ) \mathop{\mathrm{OMC}}(\rho) OMC ( ρ ) . By Remark 4.6 , N C ( ρ ) = max c O M C ( ρ c ) \mathop{\mathrm{NC}}(\rho)=\max_c\mathop{\mathrm{OMC}}(\rho_c) NC ( ρ ) = max c OMC ( ρ c ) is the contrast inside the blocks, and for every exact constant Theorem 8.1 gives c M ( ρ ) = f ( max c κ ( ρ c ) ) = max c c M k c ( ρ c ) c_\mathcal{M}(\rho)=f(\max_c\kappa(\rho_c))=\max_cc_{M_{k_c}}(\rho_c) c M ( ρ ) = f ( max c κ ( ρ c )) = max c c M k c ( ρ c ) . The following corollary extends this, as an inequality, to the items of Table 3 for which only matrix envelopes are known; its proof is elementary.
Corollary 8.2 (Direct sums). Let ρ = ⨁ c ρ c > 0 \rho=\bigoplus_c\rho_c>0 ρ = ⨁ c ρ c > 0 in M \mathcal{M} M , and let V M ( ρ ) V_\mathcal{M}(\rho) V M ( ρ ) and V M k c ( ρ c ) V_{M_{k_c}}(\rho_c) V M k c ( ρ c ) be the suprema of the normalised violation of one of the items 1–25 of Table 3 , other than 11, 16 and 24, for φ ρ \varphi_\rho φ ρ on M \mathcal{M} M and for φ ρ c \varphi_{\rho_c} φ ρ c on M k c M_{k_c} M k c . Then V M ( ρ ) ≤ max c V M k c ( ρ c ) , V_\mathcal{M}(\rho)\le\max_cV_{M_{k_c}}(\rho_c), V M ( ρ ) ≤ max c V M k c ( ρ c ) , with equality for items 1–10, 12–14, 17, 18, 21–23 and 25. In particular, if every ρ c \rho_c ρ c is scalar, i.e. N C ( ρ ) = 0 \mathop{\mathrm{NC}}(\rho)=0 NC ( ρ ) = 0 , none of the items 1–25 except 11 and 24 is violated, however different the scalars are; and the envelope on M \mathcal{M} M along the noncentrality N C ( ρ ) = max c O M C ( ρ c ) = δ \mathop{\mathrm{NC}}(\rho)=\max_c\mathop{\mathrm{OMC}}(\rho_c)=\delta NC ( ρ ) = max c OMC ( ρ c ) = δ is at most the largest envelope V k c ( δ ) V_{k_c}(\delta) V k c ( δ ) of the blocks.
Proof. All variables of the items are in M \mathcal{M} M , hence block diagonal, and every functional calculus acts blockwise. Let L c \mathcal{L}_c L c , R c \mathcal{R}_c R c , N c \mathcal{N}_c N c be the expressions of the item for the block variables and φ ρ c \varphi_{\rho_c} φ ρ c . For all items except 11, 16 and 24, L ≤ ∑ c L c \mathcal{L}\le\sum_cL_c L ≤ ∑ c L c (with equality unless L \mathcal{L} L contains an absolute value of a functional, as in 1, 4 and 14), N = ∑ c N c \mathcal{N}=\sum_cN_c N = ∑ c N c , and R = ∑ c R c \mathcal{R}=\sum_cR_c R = ∑ c R c , except in items 6 and 14, where R = ∑ c φ c ( A c ) ∑ c φ c ( B c ) ≥ ∑ c R c \mathcal{R}=\sqrt{\sum_c\varphi_c(A_c)\sum_c\varphi_c(B_c)}\ge
\sum_cR_c R = ∑ c φ c ( A c ) ∑ c φ c ( B c ) ≥ ∑ c R c by the Cauchy–Schwarz inequality and R ≥ ∑ c R c \mathcal{R}\ge\sum_cR_c R ≥ ∑ c R c by Hölder’s inequality. Since ρ > 0 \rho>0 ρ > 0 , N c = 0 \mathcal{N}_c=0 N c = 0 forces the variables of block c c c to vanish where they enter L c \mathcal{L}_c L c , so L c ≤ R c \mathcal{L}_c\le \mathcal{R}_c L c ≤ R c there. Hence ( L − R ) / N ≤ ∑ c ( L c − R c ) / ∑ c N c ≤ max c ( L c − R c ) / N c (\mathcal{L}-\mathcal{R})/\mathcal{N}\le\sum_c(\mathcal{L}_c-\mathcal{R}_c)/\sum_cN_c\le\max_c(\mathcal{L}_c-\mathcal{R}_c)/\mathcal{N}_c ( L − R ) / N ≤ ∑ c ( L c − R c ) / ∑ c N c ≤ max c ( L c − R c ) / N c , the maximum over the blocks with N c > 0 \mathcal{N}_c>0 N c > 0 (for N = R \mathcal{N}=\mathcal{R} N = R in items 6 and 14 use L / R ≤ ∑ c L c / ∑ c R c \mathcal{L}/\mathcal{R}\le\sum_cL_c/\sum_cR_c L / R ≤ ∑ c L c / ∑ c R c instead). This proves the inequality. For the equality, in the listed items N \mathcal{N} N is additive and the variables may vanish on all blocks but one without making N \mathcal{N} N vanish; restricting a near-optimal configuration of one block and setting the variables to zero on the other blocks reproduces its quotient (in 19, 20 and 15, N \mathcal{N} N contains φ ( 1 ) \varphi(1) φ ( 1 ) and the other blocks dilute the violation). For item 6, whose variables are invertible, set them equal to ε 1 \varepsilon1 ε 1 on the other blocks instead; their contributions to L \mathcal{L} L , R \mathcal{R} R and N \mathcal{N} N are O ( ε ) O(\varepsilon) O ( ε ) , and the quotient follows as ε → 0 \varepsilon\to0 ε → 0 . If every ρ c \rho_c ρ c is scalar, every V M k c ( ρ c ) V_{M_{k_c}}(\rho_c) V M k c ( ρ c ) vanishes, since φ ρ c \varphi_{\rho_c} φ ρ c is then a trace on M k c M_{k_c} M k c ; for item 16 the same conclusion holds because the Peierls–Bogoliubov inequality holds for every trace on M \mathcal{M} M . ◻
Contrasts of composite operators
The operators of Table 3 enter the inequalities through composites: sums A + B A+B A + B (items 2, 10, 12, 13), congruences B 1 / 2 A B 1 / 2 B^{1/2}AB^{1/2} B 1/2 A B 1/2 (items 6, 7), means (item 6), Jordan products R e P Q = 1 2 ( P Q + Q P ) \mathrm{Re}\,PQ=\frac12(PQ+QP) Re P Q = 2 1 ( P Q + QP ) (items 9, 21, 22), and, behind every characterisation, the commutator A B − B A AB-BA A B − B A , on which a functional vanishes if and only if it is a trace. This section describes how the contrast of a composite depends on the contrasts of its factors, and then uses the answers to explain the thresholds observed in the numerical search. Two reparametrisations of the contrast make the answers additive. For A > 0 A>0 A > 0 with extreme eigenvalues M A ≥ m A > 0 M_A\ge m_A>0 M A ≥ m A > 0 and condition number κ A = M A / m A \kappa_A=M_A/m_A κ A = M A / m A let
η ( A ) = artanh O M C ( A ) = 1 2 log κ A , ψ ( A ) = arcsin O M C ( A ) ∈ [ 0 , π 2 ] , λ ˉ ( A ) = M A + m A 2 . \eta(A)=\operatorname{artanh}\mathop{\mathrm{OMC}}(A)=\tfrac12\log \kappa_A,\qquad
\psi(A)=\arcsin\mathop{\mathrm{OMC}}(A)\in[0,\tfrac\pi2],\qquad \bar\lambda(A)=\frac{M_A+m_A}2 . η ( A ) = artanh OMC ( A ) = 2 1 log κ A , ψ ( A ) = arcsin OMC ( A ) ∈ [ 0 , 2 π ] , λ ˉ ( A ) = 2 M A + m A .
The rapidity η ( A ) = O L C ( A ) \eta(A)=\mathop{\mathrm{OLC}}(A) η ( A ) = OLC ( A ) is half the logarithm of the condition number, that is, half the Hilbert projective distance of A A A to 1 1 1 and the Thompson distance of A A A to the scalars (§ 2 ). The angle ψ ( A ) \psi(A) ψ ( A ) is that of Krein and Gustafson [143 , 127 , 129 ] , the largest angle between x x x and A x Ax A x : cos ψ ( A ) = min x ≠ 0 ⟨ A x , x ⟩ / ( ∥ A x ∥ ∥ x ∥ ) = 2 M A m A / ( M A + m A ) \cos\psi(A)=\min_{x\ne0}\langle Ax,x\rangle/(\|Ax\|\|x\|)=2\sqrt{M_Am_A}/(M_A+m_A) cos ψ ( A ) = min x = 0 ⟨ A x , x ⟩ / ( ∥ A x ∥∥ x ∥ ) = 2 M A m A / ( M A + m A ) , and sin ψ ( A ) = min t > 0 ∥ t A − 1 ∥ = O M C ( A ) \sin\psi(A)=\min_{t>0}\|tA-1\|=\mathop{\mathrm{OMC}}(A) sin ψ ( A ) = min t > 0 ∥ t A − 1∥ = OMC ( A ) , attained at t = 1 / λ ˉ ( A ) t=1/\bar\lambda(A) t = 1/ λ ˉ ( A ) (Figure 1 (a)). For the density, G = cosh η ( ρ ) = sec ψ ( ρ ) G=\cosh\eta(\rho)=\sec\psi(\rho) G = cosh η ( ρ ) = sec ψ ( ρ ) , so the envelope G − 1 G-1 G − 1 of items 1–5, 18 and 22 (Theorem 5.1 ) is sec ψ ( ρ ) − 1 \sec\psi(\rho)-1 sec ψ ( ρ ) − 1 . The angle between vectors is ∠ ( x , y ) = arccos ( R e ⟨ x , y ⟩ / ( ∥ x ∥ ∥ y ∥ ) ) \angle(x,y)=\arccos(\mathrm{Re}\,\langle x,y\rangle/
(\|x\|\|y\|)) ∠ ( x , y ) = arccos ( Re ⟨ x , y ⟩ / ( ∥ x ∥∥ y ∥ )) , the angle of the real inner product R e ⟨ ⋅ , ⋅ ⟩ \mathrm{Re}\,\langle\cdot,\cdot\rangle Re ⟨ ⋅ , ⋅ ⟩ on C n = R 2 n \mathbb{C}^n=\mathbb{R}^{2n} C n = R 2 n ; it satisfies the triangle inequality.
Proposition 9.1 (Multiplicative composites: rapidities add). Let A , B > 0 A,B>0 A , B > 0 in M n M_n M n and η A = η ( A ) \eta_A=\eta(A) η A = η ( A ) , η B = η ( B ) \eta_B=\eta(B) η B = η ( B ) .
O M C ( A p ) = tanh ( ∣ p ∣ η A ) \mathop{\mathrm{OMC}}(A^p)=\tanh(|p|\eta_A) OMC ( A p ) = tanh ( ∣ p ∣ η A ) for real p ≠ 0 p\ne0 p = 0 , O M C ( A ⊗ B ) = tanh ( η A + η B ) \mathop{\mathrm{OMC}}(A\otimes B)=\tanh(\eta_A+\eta_B) OMC ( A ⊗ B ) = tanh ( η A + η B ) , and the compression of A A A to a subspace has contrast at most O M C ( A ) \mathop{\mathrm{OMC}}(A) OMC ( A ) .
tanh ∣ η A − η B ∣ ≤ O M C ( A 1 / 2 B A 1 / 2 ) ≤ tanh ( η A + η B ) \tanh|\eta_A-\eta_B|\le\mathop{\mathrm{OMC}}(A^{1/2}BA^{1/2})\le\tanh(\eta_A+\eta_B) tanh ∣ η A − η B ∣ ≤ OMC ( A 1/2 B A 1/2 ) ≤ tanh ( η A + η B ) ; the same bounds hold for the contrast of ∣ A B ∣ |AB| ∣ A B ∣ , and for the spectrum of A B AB A B .
For the geometric mean A # B = A 1 / 2 ( A − 1 / 2 B A − 1 / 2 ) 1 / 2 A 1 / 2 A\#B=A^{1/2}(A^{-1/2}BA^{-1/2})^{1/2}A^{1/2} A # B = A 1/2 ( A − 1/2 B A − 1/2 ) 1/2 A 1/2 , tanh ∣ η A − η B ∣ 2 ≤ O M C ( A # B ) ≤ tanh η A + η B 2 \tanh\frac{|\eta_A-\eta_B|}2\le\mathop{\mathrm{OMC}}(A\#B)\le\tanh\frac{\eta_A+\eta_B}2 tanh 2 ∣ η A − η B ∣ ≤ OMC ( A # B ) ≤ tanh 2 η A + η B .
O M C ( s A + t B ) ≤ max ( O M C ( A ) , O M C ( B ) ) \mathop{\mathrm{OMC}}(sA+tB)\le\max(\mathop{\mathrm{OMC}}(A),\mathop{\mathrm{OMC}}(B)) OMC ( s A + tB ) ≤ max ( OMC ( A ) , OMC ( B )) for s , t > 0 s,t>0 s , t > 0 , and the same holds for the harmonic mean. If O M C ( A ) , O M C ( B ) > 0 \mathop{\mathrm{OMC}}(A),\mathop{\mathrm{OMC}}(B)>0 OMC ( A ) , OMC ( B ) > 0 , then O M C ( A + t B ) = 0 \mathop{\mathrm{OMC}}(A+tB)=0 OMC ( A + tB ) = 0 for suitable t > 0 t>0 t > 0 and a suitable relative position.
All bounds are attained by commuting operators, with the extreme eigenvectors aligned (upper bounds) or interchanged (lower bounds), in every dimension n ≥ 2 n\ge2 n ≥ 2 .
These facts are classical, and we include the short proof for the attainment statements. The upper bound in (b) is the submultiplicativity of the condition number, κ ( A 1 / 2 B A 1 / 2 ) ≤ κ A κ B \kappa(A^{1/2}BA^{1/2})\le \kappa_A\kappa_B κ ( A 1/2 B A 1/2 ) ≤ κ A κ B . In the language of Hilbert’s projective metric d H d_H d H , which is invariant under congruences and under inversion, (b) reads ∣ d H ( B , 1 ) − d H ( 1 , A − 1 ) ∣ ≤ d H ( B , A − 1 ) = d H ( A 1 / 2 B A 1 / 2 , 1 ) ≤ d H ( B , 1 ) + d H ( 1 , A − 1 ) |d_H(B,1)-d_H(1,A^{-1})|\le d_H(B,A^{-1})=d_H(A^{1/2}BA^{1/2},1)\le d_H(B,1)+d_H(1,A^{-1}) ∣ d H ( B , 1 ) − d H ( 1 , A − 1 ) ∣ ≤ d H ( B , A − 1 ) = d H ( A 1/2 B A 1/2 , 1 ) ≤ d H ( B , 1 ) + d H ( 1 , A − 1 ) , i.e. it is the triangle inequality [115 , 144 ] ; this is why rapidities, the halved distances to the scalars, add.
Proof. (a) Spectral mapping, the eigenvalues of A ⊗ B A\otimes B A ⊗ B are the products of eigenvalues, and Cauchy interlacing. (b) From m B A ≤ A 1 / 2 B A 1 / 2 ≤ M B A m_BA\le A^{1/2}BA^{1/2}\le M_BA m B A ≤ A 1/2 B A 1/2 ≤ M B A and the monotonicity of the eigenvalues [93 ] , λ k ( A ) m B ≤ λ k ( A 1 / 2 B A 1 / 2 ) ≤ λ k ( A ) M B \lambda_k(A)m_B\le\lambda_k(A^{1/2}BA^{1/2})\le\lambda_k(A)M_B λ k ( A ) m B ≤ λ k ( A 1/2 B A 1/2 ) ≤ λ k ( A ) M B for every k k k . Hence the largest eigenvalue lies in [ M A m B , M A M B ] [M_Am_B,M_AM_B] [ M A m B , M A M B ] and the smallest in [ m A m B , m A M B ] [m_Am_B,m_AM_B] [ m A m B , m A M B ] , so the condition number lies in [ κ A / κ B , κ A κ B ] [\kappa_A/\kappa_B,\kappa_A\kappa_B] [ κ A / κ B , κ A κ B ] ; since A 1 / 2 B A 1 / 2 A^{1/2}BA^{1/2} A 1/2 B A 1/2 is similar to B 1 / 2 A B 1 / 2 B^{1/2}AB^{1/2} B 1/2 A B 1/2 , it is also at least κ B / κ A \kappa_B/\kappa_A κ B / κ A . With O M C = tanh ( 1 2 log κ ) \mathop{\mathrm{OMC}}=\tanh(\frac12\log \kappa) OMC = tanh ( 2 1 log κ ) this is (b). For ∣ A B ∣ |AB| ∣ A B ∣ use the singular values, σ k ( A ) σ min ( B ) ≤ σ k ( A B ) ≤ σ k ( A ) σ max ( B ) \sigma_k(A)\sigma_{\min}(B)\le\sigma_k(AB)\le\sigma_k(A)\sigma_{\max}(B) σ k ( A ) σ m i n ( B ) ≤ σ k ( A B ) ≤ σ k ( A ) σ m a x ( B ) ; A B AB A B is similar to A 1 / 2 B A 1 / 2 A^{1/2}BA^{1/2} A 1/2 B A 1/2 . (c) The mean is monotone and homogeneous, so m A m B ≤ A # B ≤ M A M B \sqrt{m_Am_B}\le A\#B\le
\sqrt{M_AM_B} m A m B ≤ A # B ≤ M A M B , which gives the upper bound. From B = ( A # B ) A − 1 ( A # B ) B=(A\#B)A^{-1}(A\#B) B = ( A # B ) A − 1 ( A # B ) , whose spectrum is that of A − 1 / 2 ( A # B ) 2 A − 1 / 2 A^{-1/2}(A\#B)^2A^{-1/2} A − 1/2 ( A # B ) 2 A − 1/2 , (a) and (b) give η B ≤ η A + 2 η ( A # B ) \eta_B\le\eta_A+2\eta(A\#B) η B ≤ η A + 2 η ( A # B ) , and symmetrically. (d) The mediant inequality s M A + t M B s m A + t m B ≤ max ( κ A , κ B ) \frac{sM_A+tM_B}{sm_A+tm_B}\le\max(\kappa_A,\kappa_B) s m A + t m B s M A + t M B ≤ max ( κ A , κ B ) bounds the condition number of s A + t B sA+tB s A + tB ; for the harmonic mean apply it to A − 1 + B − 1 A^{-1}+B^{-1} A − 1 + B − 1 . For the last claim take B = c − A B=c-A B = c − A with c > M A c>M_A c > M A in the eigenbasis of A A A and with the intermediate eigenvalues reflected; O M C ( c − A ) = ( M A − m A ) / ( 2 c − M A − m A ) \mathop{\mathrm{OMC}}(c-A)=(M_A-m_A)/(2c-M_A-m_A) OMC ( c − A ) = ( M A − m A ) / ( 2 c − M A − m A ) takes every value in ( 0 , 1 ) (0,1) ( 0 , 1 ) . Attainment: for commuting A , B A,B A , B with aligned extreme eigenvectors all extreme eigenvalues multiply (add, for sums), and with interchanged ones the quotients do. ◻
The Jordan product.
For A B + B A AB+BA A B + B A the picture changes: it need not be positive, and angles, not rapidities, add. The reason is geometric. A positive operator A A A turns every vector by at most its angle: ∠ ( x , A x ) ≤ ψ A \angle(x,Ax)\le\psi_A ∠ ( x , A x ) ≤ ψ A , where sin ψ A = O M C ( A ) \sin\psi_A=\mathop{\mathrm{OMC}}(A) sin ψ A = OMC ( A ) . Hence A B x ABx A B x and B A x BAx B A x differ from x x x by at most ψ A + ψ B \psi_A+\psi_B ψ A + ψ B , and so does their sum ( A B + B A ) x (AB+BA)x ( A B + B A ) x , as long as ψ A + ψ B < π 2 \psi_A+\psi_B<\frac\pi2 ψ A + ψ B < 2 π . Then ( A B + B A ) x (AB+BA)x ( A B + B A ) x stays in the half-space of x x x , i.e. ⟨ ( A B + B A ) x , x ⟩ > 0 \langle(AB+BA)x,x\rangle>0 ⟨( A B + B A ) x , x ⟩ > 0 . If ψ A + ψ B > π 2 \psi_A+\psi_B>\frac\pi2 ψ A + ψ B > 2 π , a suitable rotation of B B B against A A A turns some x x x by more than a right angle, and A B + B A AB+BA A B + B A has a negative eigenvalue. In terms of contrasts: A B + B A AB+BA A B + B A is positive for every relative position if and only if the point ( O M C ( A ) , O M C ( B ) ) (\mathop{\mathrm{OMC}}(A),\mathop{\mathrm{OMC}}(B)) ( OMC ( A ) , OMC ( B )) lies in the unit disc (Figure 5 ). The sufficiency of this criterion is Gustafson’s [127 ] , whose angle argument we reproduce. Its sharpness is in the spirit of the work on eigenvalues of Jordan products by Strang [161 ] , Alikakos and Bates [87 , 88 ] and Conley, Pucci and Serrin [118 ] ; we include the short proof of both directions, since we need the extremal configuration, and add that the same angles bound the contrast of A B + B A AB+BA A B + B A .
Figure 5. Is A B + B A AB+BA A B + B A positive? 100000 100000 100000 random pairs A , B > 0 A,B>0 A , B > 0 per dimension, with contrasts uniform in [ 0 , 1 ] 2 [0,1]^2 [ 0 , 1 ] 2 and Haar random relative position: positive (grey) or indefinite (orange). No indefinite product occurs inside the quarter circle O M C ( A ) 2 + O M C ( B ) 2 = 1 \mathop{\mathrm{OMC}}(A)^2+\mathop{\mathrm{OMC}}(B)^2=1 OMC ( A ) 2 + OMC ( B ) 2 = 1 (Theorem 9.2 (a)); outside it occurs for some positions, more often the further out. In the figure Δ ( ⋅ ) = O M C ( ⋅ ) \Delta(\cdot)=\mathop{\mathrm{OMC}}(\cdot) Δ ( ⋅ ) = OMC ( ⋅ ) .
Theorem 9.2 (Jordan product). Let A , B > 0 A,B>0 A , B > 0 in M n M_n M n , n ≥ 2 n\ge2 n ≥ 2 , with contrasts δ A , δ B \delta_A,\delta_B δ A , δ B , angles ψ A = arcsin δ A \psi_A=\arcsin\delta_A ψ A = arcsin δ A , ψ B = arcsin δ B \psi_B=\arcsin\delta_B ψ B = arcsin δ B and rapidities η A = artanh δ A \eta_A=\operatorname{artanh}\delta_A η A = artanh δ A , η B = artanh δ B \eta_B=\operatorname{artanh}\delta_B η B = artanh δ B .
A B + B A ≥ 0 AB+BA\ge0 A B + B A ≥ 0 for every relative position of A A A and B B B if and only if δ A 2 + δ B 2 ≤ 1 \delta_A^2+\delta_B^2\le1 δ A 2 + δ B 2 ≤ 1 . If δ A 2 + δ B 2 > 1 \delta_A^2+\delta_B^2>1 δ A 2 + δ B 2 > 1 , A B + B A AB+BA A B + B A is indefinite for some relative position, in every dimension.
If δ A 2 + δ B 2 < 1 \delta_A^2+\delta_B^2<1 δ A 2 + δ B 2 < 1 , then tanh ∣ η A − η B ∣ ≤ O M C ( A B + B A ) ≤ sin ( ψ A + ψ B ) \tanh|\eta_A-\eta_B|\le\mathop{\mathrm{OMC}}(AB+BA)\le\sin(\psi_A+\psi_B) tanh ∣ η A − η B ∣ ≤ OMC ( A B + B A ) ≤ sin ( ψ A + ψ B ) , and the lower bound is attained by commuting A , B A,B A , B whose extreme eigenvectors are interchanged.
Proof. (a) and the upper bound in (b). We use the real angle ∠ ( x , y ) \angle(x,y) ∠ ( x , y ) defined above. By the formula for cos ψ \cos\psi cos ψ , ∠ ( x , B x ) ≤ ψ B \angle(x,Bx)\le\psi_B ∠ ( x , B x ) ≤ ψ B and ∠ ( B x , A B x ) ≤ ψ A \angle(Bx,ABx)\le\psi_A ∠ ( B x , A B x ) ≤ ψ A , so ∠ ( x , A B x ) ≤ ψ A + ψ B \angle(x,ABx)\le\psi_A+\psi_B ∠ ( x , A B x ) ≤ ψ A + ψ B by the triangle inequality; likewise ∠ ( x , B A x ) ≤ ψ A + ψ B \angle(x,BAx)\le\psi_A+\psi_B ∠ ( x , B A x ) ≤ ψ A + ψ B . If ψ A + ψ B < π 2 \psi_A+\psi_B<\frac\pi2 ψ A + ψ B < 2 π , the cone { y : ∠ ( x , y ) ≤ ψ A + ψ B } \{y:\angle(x,y)\le\psi_A+\psi_B\} { y : ∠ ( x , y ) ≤ ψ A + ψ B } is convex and contains ( A B + B A ) x (AB+BA)x ( A B + B A ) x , so ⟨ ( A B + B A ) x , x ⟩ > 0 \langle(AB+BA)x,x\rangle>0 ⟨( A B + B A ) x , x ⟩ > 0 ; moreover the largest angle of A B + B A AB+BA A B + B A is at most ψ A + ψ B \psi_A+\psi_B ψ A + ψ B , and its sine is O M C ( A B + B A ) \mathop{\mathrm{OMC}}(AB+BA) OMC ( A B + B A ) . The case ψ A + ψ B = π 2 \psi_A+\psi_B=\frac\pi2 ψ A + ψ B = 2 π follows by continuity.
Converse in (a). On M 2 M_2 M 2 take A = d i a g ( 1 , ℓ A ) A=\mathrm{diag}(1,\ell_A) A = diag ( 1 , ℓ A ) and B B B with eigenvalues 1 , ℓ B 1,\ell_B 1 , ℓ B and eigenvector ( cos θ , sin θ ) (\cos\theta,\sin\theta) ( cos θ , sin θ ) for 1 1 1 , where ℓ A = 1 − δ A 1 + δ A \ell_A=\frac{1-\delta_A}{1+\delta_A} ℓ A = 1 + δ A 1 − δ A , ℓ B = 1 − δ B 1 + δ B \ell_B=\frac{1-\delta_B}{1+\delta_B} ℓ B = 1 + δ B 1 − δ B . With c = cos 2 θ c=\cos^2\theta c = cos 2 θ and p = ( 1 − ℓ A ) ( 1 − ℓ B ) p=(1-\ell_A)(1-\ell_B) p = ( 1 − ℓ A ) ( 1 − ℓ B ) a direct computation gives
1 2 T r ( A B + B A ) = ℓ A + ℓ B + p c , det ( A B + B A ) = 4 ℓ A ℓ B − p 2 c ( 1 − c ) . \tfrac12\mathop{\mathrm{Tr}}(AB+BA)=\ell_A+\ell_B+pc,\qquad\det(AB+BA)=4\ell_A\ell_B-p^2c(1-c). 2 1 Tr ( A B + B A ) = ℓ A + ℓ B + p c , det ( A B + B A ) = 4 ℓ A ℓ B − p 2 c ( 1 − c ) . At θ = π / 4 \theta=\pi/4 θ = π /4 the determinant is negative if and only if 16 ℓ A ℓ B < p 2 16\ell_A\ell_B<p^2 16 ℓ A ℓ B < p 2 ; with ℓ A = 1 − δ A 2 / ( 1 + δ A ) \sqrt{\ell_A}=\sqrt{1-\delta_A^2}/(1+\delta_A) ℓ A = 1 − δ A 2 / ( 1 + δ A ) and 1 − ℓ A = 2 δ A / ( 1 + δ A ) 1-\ell_A=2\delta_A/(1+\delta_A) 1 − ℓ A = 2 δ A / ( 1 + δ A ) this reads ( 1 − δ A 2 ) ( 1 − δ B 2 ) < δ A 2 δ B 2 (1-\delta_A^2)(1-\delta_B^2)<\delta_A^2\delta_B^2 ( 1 − δ A 2 ) ( 1 − δ B 2 ) < δ A 2 δ B 2 , i.e. δ A 2 + δ B 2 > 1 \delta_A^2+\delta_B^2>1 δ A 2 + δ B 2 > 1 . On M n M_n M n add the direct summands ℓ A 1 n − 2 \ell_A1_{n-2} ℓ A 1 n − 2 and ℓ B 1 n − 2 \ell_B1_{n-2} ℓ B 1 n − 2 .
Lower bound in (b). If x , y x,y x , y are unit eigenvectors of A A A for its extreme eigenvalues M A M_A M A , m A m_A m A , then λ max ( A B + B A ) ≥ 2 R e ⟨ B x , A x ⟩ = 2 M A ⟨ B x , x ⟩ ≥ 2 M A m B \lambda_{\max}(AB+BA)\ge2\mathrm{Re}\,\langle Bx,Ax\rangle=2M_A\langle Bx,x\rangle\ge2M_Am_B λ m a x ( A B + B A ) ≥ 2 Re ⟨ B x , A x ⟩ = 2 M A ⟨ B x , x ⟩ ≥ 2 M A m B and likewise λ min ( A B + B A ) ≤ 2 m A M B \lambda_{\min}(AB+BA)\le2m_AM_B λ m i n ( A B + B A ) ≤ 2 m A M B . So the condition number of A B + B A AB+BA A B + B A is at least κ A / κ B \kappa_A/\kappa_B κ A / κ B , and by symmetry at least κ B / κ A \kappa_B/\kappa_A κ B / κ A . ◻
The angle argument in the proof of (a) works on any Hilbert space, so O M C ( A ) 2 + O M C ( B ) 2 < 1 \mathop{\mathrm{OMC}}(A)^2+\mathop{\mathrm{OMC}}(B)^2<1 OMC ( A ) 2 + OMC ( B ) 2 < 1 implies A B + B A > 0 AB+BA>0 A B + B A > 0 for positive invertible operators on any Hilbert space. Since A B + B A > 0 AB+BA>0 A B + B A > 0 is invariant under ( A , B ) ↦ ( z 1 A , z 2 B ) (A,B)\mapsto(z_1A,z_2B) ( A , B ) ↦ ( z 1 A , z 2 B ) with z 1 , z 2 ∈ Z + + z_1,z_2\in\mathcal{Z}_{++} z 1 , z 2 ∈ Z ++ , rescaling then shows that in every von Neumann algebra N C ( A ) 2 + N C ( B ) 2 < 1 \mathop{\mathrm{NC}}(A)^2+\mathop{\mathrm{NC}}(B)^2<1 NC ( A ) 2 + NC ( B ) 2 < 1 implies A B + B A > 0 AB+BA>0 A B + B A > 0 (Remark 4.10 ).
How large can O M C ( A B + B A ) \mathop{\mathrm{OMC}}(AB+BA) OMC ( A B + B A ) be? The upper bound sin ( ψ A + ψ B ) \sin(\psi_A+\psi_B) sin ( ψ A + ψ B ) is reached only at the boundary of the disc. The true maximum depends on the dimension, because in M 2 M_2 M 2 both extreme eigenvalues of A B + B A AB+BA A B + B A come from the same pair of 2 × 2 2\times2 2 × 2 matrices, while in M 3 M_3 M 3 they can come from different blocks.
Proposition 9.3 (Largest contrast of the Jordan product). In the notation of the proof above, let δ A 2 + δ B 2 < 1 \delta_A^2+\delta_B^2<1 δ A 2 + δ B 2 < 1 .
On M 2 M_2 M 2 , max O M C ( A B + B A ) = max 0 ≤ c ≤ 1 ( 1 − 4 ℓ A ℓ B − p 2 c ( 1 − c ) ( ℓ A + ℓ B + p c ) 2 ) 1 / 2 \displaystyle\max\mathop{\mathrm{OMC}}(AB+BA)=\max_{0\le c\le1}\Bigl(1-\frac{4\ell_A\ell_B-p^2c(1-c)}{(\ell_A+\ell_B+pc)^2}\Bigr)^{1/2} max OMC ( A B + B A ) = 0 ≤ c ≤ 1 max ( 1 − ( ℓ A + ℓ B + p c ) 2 4 ℓ A ℓ B − p 2 c ( 1 − c ) ) 1/2 .
On M n M_n M n , n ≥ 3 n\ge3 n ≥ 3 , max O M C ( A B + B A ) ≥ ( 2 − μ ) / ( 2 + μ ) \max\mathop{\mathrm{OMC}}(AB+BA)\ge(2-\mu)/(2+\mu) max OMC ( A B + B A ) ≥ ( 2 − μ ) / ( 2 + μ ) , where μ = min 0 ≤ c ≤ 1 ( ℓ A + ℓ B + p c − ( ℓ A + ℓ B + p c ) 2 − 4 ℓ A ℓ B + p 2 c ( 1 − c ) ) \mu=\min_{0\le c\le1}\bigl(\ell_A+\ell_B+pc-\sqrt{(\ell_A+\ell_B+pc)^2-4\ell_A\ell_B+p^2c(1-c)}\,\bigr) μ = min 0 ≤ c ≤ 1 ( ℓ A + ℓ B + p c − ( ℓ A + ℓ B + p c ) 2 − 4 ℓ A ℓ B + p 2 c ( 1 − c ) ) is the smallest eigenvalue of A B + B A AB+BA A B + B A over the 2 × 2 2\times2 2 × 2 pairs of (a).
Numerically, both values equal the commuting value tanh ( η A + η B ) \tanh(\eta_A+\eta_B) tanh ( η A + η B ) for small contrasts and exceed it for large ones.
Proof. (a) On M 2 M_2 M 2 every pair is unitarily equivalent to the pair of the proof of Theorem 9.2 for some θ \theta θ , and O M C 2 = 1 − 4 det / T r 2 \mathop{\mathrm{OMC}}^2=1-4\det/\mathop{\mathrm{Tr}}^2 OMC 2 = 1 − 4 det / Tr 2 . (b) Take A = 1 ⊕ A 2 A=1\oplus A_2 A = 1 ⊕ A 2 and B = 1 ⊕ B 2 B=1\oplus B_2 B = 1 ⊕ B 2 (padded with ones) with the 2 × 2 2\times2 2 × 2 pair A 2 , B 2 A_2,B_2 A 2 , B 2 of smallest eigenvalue μ \mu μ . Then A B + B A = 2 ⊕ ( A 2 B 2 + B 2 A 2 ) AB+BA=2\oplus(A_2B_2+B_2A_2) A B + B A = 2 ⊕ ( A 2 B 2 + B 2 A 2 ) has the largest eigenvalue 2 2 2 , since ∥ A 2 B 2 + B 2 A 2 ∥ ≤ 2 \|A_2B_2+B_2A_2\|\le2 ∥ A 2 B 2 + B 2 A 2 ∥ ≤ 2 , and the smallest eigenvalue μ \mu μ . ◻
For δ A = δ B = 0.6 \delta_A=\delta_B=0.6 δ A = δ B = 0.6 : the commuting value is 0.8824 0.8824 0.8824 , the maximum on M 2 M_2 M 2 is 0.8839 0.8839 0.8839 , the value of (b) for n ≥ 3 n\ge3 n ≥ 3 is 0.8963 0.8963 0.8963 , and sin ( ψ A + ψ B ) = 0.9600 \sin(\psi_A+\psi_B)=0.9600 sin ( ψ A + ψ B ) = 0.9600 . A numerical optimisation over all relative positions and intermediate eigenvalues (24 starts per point) finds on M 3 M_3 M 3 and on M 4 M_4 M 4 exactly the value of (b) (0.8963 0.8963 0.8963 and 0.8963 0.8963 0.8963 at δ A = δ B = 0.6 \delta_A=\delta_B=0.6 δ A = δ B = 0.6 , and largest difference 3 × 10 − 7 3\times 10^{-7} 3 × 1 0 − 7 on a grid of δ A \delta_A δ A with δ B = 0.6 \delta_B=0.6 δ B = 0.6 ; Figure 7 ). Equality for every n ≥ 3 n\ge3 n ≥ 3 would follow if the smallest eigenvalue of a Jordan product with spectra in [ ℓ A , 1 ] [\ell_A,1] [ ℓ A , 1 ] and [ ℓ B , 1 ] [\ell_B,1] [ ℓ B , 1 ] were always attained on a two-dimensional subspace. Strang [161 ] bounded this smallest eigenvalue from below, and Alikakos and Bates showed that his bound is best possible [88 ] ; we have not checked whether it coincides with μ \mu μ .
Proposition 9.4 (Commutators: three contrasts multiply). Let A , B > 0 A,B>0 A , B > 0 in M n M_n M n .
∥ A B − B A ∥ ≤ 2 O M C ( A ) O M C ( B ) λ ˉ ( A ) λ ˉ ( B ) \|AB-BA\|\le2\mathop{\mathrm{OMC}}(A)\mathop{\mathrm{OMC}}(B)\,\bar\lambda(A)\bar\lambda(B) ∥ A B − B A ∥ ≤ 2 OMC ( A ) OMC ( B ) λ ˉ ( A ) λ ˉ ( B ) , with equality for suitable A , B A,B A , B of any prescribed contrasts in every M n M_n M n , n ≥ 2 n\ge2 n ≥ 2 . In particular ∥ A B − B A ∥ ≤ 2 O M C ( A ) O M C ( B ) ( 1 + O M C ( A ) ) ( 1 + O M C ( B ) ) ∥ A ∥ ∥ B ∥ ≤ 1 2 ∥ A ∥ ∥ B ∥ \|AB-BA\|\le\frac{2\mathop{\mathrm{OMC}}(A)\mathop{\mathrm{OMC}}(B)}{(1+\mathop{\mathrm{OMC}}(A))(1+\mathop{\mathrm{OMC}}(B))}\|A\|\|B\|\le
\frac12\|A\|\|B\| ∥ A B − B A ∥ ≤ ( 1 + OMC ( A )) ( 1 + OMC ( B )) 2 OMC ( A ) OMC ( B ) ∥ A ∥∥ B ∥ ≤ 2 1 ∥ A ∥∥ B ∥ .
For a state φ ρ \varphi_\rho φ ρ (T r ρ = 1 \mathop{\mathrm{Tr}}\rho=1 Tr ρ = 1 ) with O M C ( ρ ) = δ \mathop{\mathrm{OMC}}(\rho)=\delta OMC ( ρ ) = δ , κ = 1 + δ 1 − δ \kappa=\frac{1+\delta}{1-\delta} κ = 1 − δ 1 + δ , ∣ φ ρ ( A B − B A ) ∣ ≤ 2 c n ( δ ) O M C ( A ) O M C ( B ) λ ˉ ( A ) λ ˉ ( B ) , c n ( δ ) = { δ , n even , k ( κ − 1 ) k ( κ + 1 ) + 1 , n = 2 k + 1 , |\varphi_\rho(AB-BA)|\le2\,c_n(\delta)\,\mathop{\mathrm{OMC}}(A)\mathop{\mathrm{OMC}}(B)\,\bar\lambda(A)\bar\lambda(B),\qquad
c_n(\delta)=\begin{cases}\delta,&n\ \text{even},\\[2pt] \dfrac{k(\kappa-1)}{k(\kappa+1)+1},&n=2k+1,\end{cases} ∣ φ ρ ( A B − B A ) ∣ ≤ 2 c n ( δ ) OMC ( A ) OMC ( B ) λ ˉ ( A ) λ ˉ ( B ) , c n ( δ ) = ⎩ ⎨ ⎧ δ , k ( κ + 1 ) + 1 k ( κ − 1 ) , n even , n = 2 k + 1 , and the bound is attained for every δ \delta δ and all prescribed contrasts of A A A and B B B .
Part (a) is a consequence of Kittaneh’s inequality ∥ X Y − Y X ∥ ≤ ∥ X ∥ ∥ Y ∥ \|XY-YX\|\le\|X\|\|Y\| ∥ X Y − Y X ∥ ≤ ∥ X ∥∥ Y ∥ for X ≥ 0 X\ge0 X ≥ 0 [141 ] , applied to X = A − m A ≥ 0 X=A-m_A\ge0 X = A − m A ≥ 0 and Y = B − λ ˉ ( B ) Y=B-\bar\lambda(B) Y = B − λ ˉ ( B ) , whose norms are 2 O M C ( A ) λ ˉ ( A ) 2\mathop{\mathrm{OMC}}(A)\bar\lambda(A) 2 OMC ( A ) λ ˉ ( A ) and O M C ( B ) λ ˉ ( B ) \mathop{\mathrm{OMC}}(B)\bar\lambda(B) OMC ( B ) λ ˉ ( B ) ; equivalently, of the bound ∥ A B − B A ∥ ≤ 1 2 ∥ A ∥ ∥ B ∥ \|AB-BA\|\le\frac12\|A\|\|B\| ∥ A B − B A ∥ ≤ 2 1 ∥ A ∥∥ B ∥ for positive operators applied to A − m A A-m_A A − m A and B − m B B-m_B B − m B . Our contribution is the sharp trace version (b), the commutator test of traciality with all three contrasts prescribed.
Proof. (a) Write A = λ ˉ ( A ) ( 1 + E ) A=\bar\lambda(A)(1+E) A = λ ˉ ( A ) ( 1 + E ) , B = λ ˉ ( B ) ( 1 + F ) B=\bar\lambda(B)(1+F) B = λ ˉ ( B ) ( 1 + F ) with self-adjoint E , F E,F E , F , ∥ E ∥ = O M C ( A ) \|E\|=\mathop{\mathrm{OMC}}(A) ∥ E ∥ = OMC ( A ) , ∥ F ∥ = O M C ( B ) \|F\|=\mathop{\mathrm{OMC}}(B) ∥ F ∥ = OMC ( B ) . Then A B − B A = λ ˉ ( A ) λ ˉ ( B ) ( E F − F E ) AB-BA=\bar\lambda(A)\bar\lambda(B)(EF-FE) A B − B A = λ ˉ ( A ) λ ˉ ( B ) ( E F − F E ) and ∥ E F − F E ∥ ≤ 2 ∥ E ∥ ∥ F ∥ \|EF-FE\|\le2\|E\|\|F\| ∥ E F − F E ∥ ≤ 2∥ E ∥∥ F ∥ . Equality holds on M 2 M_2 M 2 for E = O M C ( A ) σ z E=\mathop{\mathrm{OMC}}(A)\sigma_z E = OMC ( A ) σ z , F = O M C ( B ) σ x F=\mathop{\mathrm{OMC}}(B)\sigma_x F = OMC ( B ) σ x , since σ z σ x − σ x σ z = 2 i σ y \sigma_z\sigma_x-\sigma_x\sigma_z=2i\sigma_y σ z σ x − σ x σ z = 2 i σ y ; in M n M_n M n add a direct summand λ ˉ ( A ) ⊕ λ ˉ ( B ) \bar\lambda(A)\oplus\bar\lambda(B) λ ˉ ( A ) ⊕ λ ˉ ( B ) . The second form uses ∥ A ∥ = λ ˉ ( A ) ( 1 + O M C ( A ) ) \|A\|=\bar\lambda(A)(1+\mathop{\mathrm{OMC}}(A)) ∥ A ∥ = λ ˉ ( A ) ( 1 + OMC ( A )) . (b) C = i ( A B − B A ) C=i(AB-BA) C = i ( A B − B A ) is self-adjoint with T r C = 0 \mathop{\mathrm{Tr}}C=0 Tr C = 0 and ∥ C ∥ ≤ K \|C\|\le K ∥ C ∥ ≤ K , the bound of (a). By von Neumann’s trace inequality [93 ] , ∣ T r ρ C ∣ |\mathop{\mathrm{Tr}}\rho C| ∣ Tr ρC ∣ is at most the maximum of ∑ i w i c i \sum_iw_ic_i ∑ i w i c i over decreasingly ordered w i w_i w i (the eigenvalues of ρ \rho ρ ) and c i ∈ [ − K , K ] c_i\in[-K,K] c i ∈ [ − K , K ] with ∑ i c i = 0 \sum_ic_i=0 ∑ i c i = 0 , which is K ∑ i ≤ k ( w i − w n + 1 − i ) K\sum_{i\le k}(w_i-w_{n+1-i}) K ∑ i ≤ k ( w i − w n + 1 − i ) , k = ⌊ n / 2 ⌋ k=\lfloor n/2\rfloor k = ⌊ n /2 ⌋ . Let a = w 1 a=w_1 a = w 1 , b = w n = a / κ b=w_n=a/\kappa b = w n = a / κ and S = ∑ i ≤ k w i ≤ k a S=\sum_{i\le k}w_i\le ka S = ∑ i ≤ k w i ≤ k a . Then ∑ i ≤ k ( w i − w n + 1 − i ) ≤ S − k b \sum_{i\le k}(w_i-w_{n+1-i})\le S-kb ∑ i ≤ k ( w i − w n + 1 − i ) ≤ S − k b and 1 = T r ρ ≥ S + ( n − k ) b 1=\mathop{\mathrm{Tr}}\rho\ge S+(n-k)b 1 = Tr ρ ≥ S + ( n − k ) b . Maximising S − k b S-kb S − k b under S ≤ k κ b S\le k\kappa b S ≤ k κb and S ≤ 1 − ( n − k ) b S\le1-(n-k)b S ≤ 1 − ( n − k ) b gives the value at k κ b = 1 − ( n − k ) b k\kappa b=1-(n-k)b k κb = 1 − ( n − k ) b , namely k ( κ − 1 ) / ( k κ + n − k ) k(\kappa-1)/(k\kappa+n-k) k ( κ − 1 ) / ( k κ + n − k ) , which is c n ( δ ) c_n(\delta) c n ( δ ) for even and odd n n n . For equality take k k k copies of the 2 × 2 2\times2 2 × 2 pair of (a), ρ \rho ρ diagonal in the eigenbasis of the k k k blocks of C C C with the eigenvalues a a a on the positive and b b b on the negative eigenvectors of C C C , and, for odd n n n , the eigenvalue b b b of ρ \rho ρ and scalars A A A , B B B on the remaining one-dimensional block. ◻
So the most basic characterisation of the trace, φ ( A B ) = φ ( B A ) \varphi(AB)=\varphi(BA) φ ( A B ) = φ ( B A ) , fails exactly at the order of the product of the three contrasts, 2 O M C ( ρ ) O M C ( A ) O M C ( B ) 2\mathop{\mathrm{OMC}}(\rho)\mathop{\mathrm{OMC}}(A)\mathop{\mathrm{OMC}}(B) 2 OMC ( ρ ) OMC ( A ) OMC ( B ) in even dimension; this is the simplest instance of the principle that a violation needs a functional far from a trace and operators far from scalars. The odd dimensions are slightly weaker, as for Gardner’s inequality (Theorem 6.1 ), and c 2 k + 1 ( δ ) → δ c_{2k+1}(\delta)\to\delta c 2 k + 1 ( δ ) → δ as k → ∞ k\to\infty k → ∞ .
Random samples (100000 100000 100000 per bound and dimension, Haar random relative positions, the intermediate eigenvalues uniform between the extreme ones) never exceed the bounds of this section on M 2 M_2 M 2 , M 3 M_3 M 3 and M 4 M_4 M 4 (the largest value of sample minus bound is − 5 × 10 − 8 -5\times 10^{-8} − 5 × 1 0 − 8 ), and among 100000 100000 100000 random pairs per dimension with contrasts uniform in [ 0 , 1 ] 2 [0,1]^2 [ 0 , 1 ] 2 the number of indefinite Jordan products with δ A 2 + δ B 2 ≤ 1 \delta_A^2+\delta_B^2\le1 δ A 2 + δ B 2 ≤ 1 is 0 0 0 . As in § 10.5 , random positions rarely come close to the extremes in higher dimension: for the commutator test on M 4 M_4 M 4 the samples stay far below the bound, which needs two aligned 2 × 2 2\times2 2 × 2 blocks (Figure 8 ).
Thresholds in the contrast of the operators
The numerical search of § 10.3 shows that several characterisations are not violated at all unless the operators have contrast close to 1 1 1 . The composition rules above explain this for two of them. The simplest example is the analogue of item 21 for positive operators, φ ( R e A B ) ≥ 0 \varphi(\mathrm{Re}\,AB)\ge0 φ ( Re A B ) ≥ 0 : it holds for every positive functional as long as O M C ( A ) 2 + O M C ( B ) 2 ≤ 1 \mathop{\mathrm{OMC}}(A)^2+\mathop{\mathrm{OMC}}(B)^2\le1 OMC ( A ) 2 + OMC ( B ) 2 ≤ 1 , because then R e A B = 1 2 ( A B + B A ) ≥ 0 \mathrm{Re}\,AB=\frac12(AB+BA)\ge0 Re A B = 2 1 ( A B + B A ) ≥ 0 by Theorem 9.2 , and for operators of equal contrast δ ′ \delta' δ ′ it can fail only for δ ′ > 1 / 2 \delta'>1/\sqrt2 δ ′ > 1/ 2 . Projections have the angle π 2 \frac\pi2 2 π , R e P Q \mathrm{Re}\,PQ Re P Q is indefinite in general, and accordingly item 21 is violated by every non-tracial functional, with the envelope 1 2 tan 2 ( ψ / 2 ) \frac12\tan^2(\psi/2) 2 1 tan 2 ( ψ /2 ) of Theorem B.
For subadditivity of the square root (item 12) the same mechanism works one level deeper, through the link between Jordan products and subadditivity found by Moslehian and Najafi [147 ] .
Proposition 9.5 (Sharp contrast threshold for item 12). Let A , B > 0 A,B>0 A , B > 0 with O M C ( A 1 / 2 ) 2 + O M C ( B 1 / 2 ) 2 ≤ 1 \mathop{\mathrm{OMC}}(A^{1/2})^2+\mathop{\mathrm{OMC}}(B^{1/2})^2\le1 OMC ( A 1/2 ) 2 + OMC ( B 1/2 ) 2 ≤ 1 , i.e. ( O M C ( A ) 1 + 1 − O M C ( A ) 2 ) 2 + ( O M C ( B ) 1 + 1 − O M C ( B ) 2 ) 2 ≤ 1. \Bigl(\frac{\mathop{\mathrm{OMC}}(A)}{1+\sqrt{1-\mathop{\mathrm{OMC}}(A)^2}}\Bigr)^2+\Bigl(\frac{\mathop{\mathrm{OMC}}(B)}{1+\sqrt{1-\mathop{\mathrm{OMC}}(B)^2}}\Bigr)^2\le1 . ( 1 + 1 − OMC ( A ) 2 OMC ( A ) ) 2 + ( 1 + 1 − OMC ( B ) 2 OMC ( B ) ) 2 ≤ 1. Then ( A + B ) 1 / 2 ≤ A 1 / 2 + B 1 / 2 (A+B)^{1/2}\le A^{1/2}+B^{1/2} ( A + B ) 1/2 ≤ A 1/2 + B 1/2 , and hence φ ( ( A + B ) 1 / 2 ) ≤ φ ( A 1 / 2 ) + φ ( B 1 / 2 ) \varphi((A+B)^{1/2})\le\varphi(A^{1/2})+\varphi(B^{1/2}) φ (( A + B ) 1/2 ) ≤ φ ( A 1/2 ) + φ ( B 1/2 ) for every positive functional φ \varphi φ . In particular, if O M C ( A ) = O M C ( B ) = δ ′ ≤ 2 2 / 3 ≈ 0.943 \mathop{\mathrm{OMC}}(A)=\mathop{\mathrm{OMC}}(B)=\delta'\le2\sqrt2/3\approx0.943 OMC ( A ) = OMC ( B ) = δ ′ ≤ 2 2 /3 ≈ 0.943 , item 12 is not violated for any density, so V n ( δ , δ ′ ) ≤ 0 V_n(\delta,\delta')\le0 V n ( δ , δ ′ ) ≤ 0 for all δ \delta δ and n n n . The threshold is sharp: if δ ′ > 2 2 / 3 \delta'>2\sqrt2/3 δ ′ > 2 2 /3 , then for every n ≥ 2 n\ge2 n ≥ 2 item 12 is violated by operators of contrast δ ′ \delta' δ ′ and every density of contrast δ \delta δ close enough to 1 1 1 .
Proof. By Proposition 9.1 (a), O M C ( A 1 / 2 ) = tanh ( η A / 2 ) = O M C ( A ) / ( 1 + 1 − O M C ( A ) 2 ) \mathop{\mathrm{OMC}}(A^{1/2})=\tanh(\eta_A/2)=\mathop{\mathrm{OMC}}(A)/(1+\sqrt{1-\mathop{\mathrm{OMC}}(A)^2}) OMC ( A 1/2 ) = tanh ( η A /2 ) = OMC ( A ) / ( 1 + 1 − OMC ( A ) 2 ) , which gives the two forms of the hypothesis. By Theorem 9.2 (a) applied to A 1 / 2 A^{1/2} A 1/2 and B 1 / 2 B^{1/2} B 1/2 , A 1 / 2 B 1 / 2 + B 1 / 2 A 1 / 2 ≥ 0 A^{1/2}B^{1/2}+B^{1/2}A^{1/2}\ge0 A 1/2 B 1/2 + B 1/2 A 1/2 ≥ 0 , hence
( A 1 / 2 + B 1 / 2 ) 2 = A + B + A 1 / 2 B 1 / 2 + B 1 / 2 A 1 / 2 ≥ A + B , (A^{1/2}+B^{1/2})^2=A+B+A^{1/2}B^{1/2}+B^{1/2}A^{1/2}\ge A+B , ( A 1/2 + B 1/2 ) 2 = A + B + A 1/2 B 1/2 + B 1/2 A 1/2 ≥ A + B , and, the square root being operator monotone, A 1 / 2 + B 1 / 2 ≥ ( A + B ) 1 / 2 A^{1/2}+B^{1/2}\ge(A+B)^{1/2} A 1/2 + B 1/2 ≥ ( A + B ) 1/2 . A positive functional preserves this inequality. For O M C ( A ) = O M C ( B ) = δ ′ \mathop{\mathrm{OMC}}(A)=\mathop{\mathrm{OMC}}(B)=\delta' OMC ( A ) = OMC ( B ) = δ ′ the hypothesis reads tanh ( η / 2 ) ≤ 1 / 2 \tanh(\eta/2)\le1/\sqrt2 tanh ( η /2 ) ≤ 1/ 2 with tanh η = δ ′ \tanh\eta=\delta' tanh η = δ ′ , i.e. δ ′ = 2 t / ( 1 + t 2 ) \delta'=2t/(1+t^2) δ ′ = 2 t / ( 1 + t 2 ) at t = 1 / 2 t=1/\sqrt2 t = 1/ 2 , which is 2 2 / 3 2\sqrt2/3 2 2 /3 .
Sharpness. Let δ ′ > 2 2 / 3 \delta'>2\sqrt2/3 δ ′ > 2 2 /3 , so that t = O M C ( A 1 / 2 ) > 1 / 2 t=\mathop{\mathrm{OMC}}(A^{1/2})>1/\sqrt2 t = OMC ( A 1/2 ) > 1/ 2 , and on M 2 M_2 M 2 let A = d i a g ( 1 , ℓ ′ ) A=\mathrm{diag}(1,\ell') A = diag ( 1 , ℓ ′ ) and B = R A R ∗ B=RAR^* B = R A R ∗ with R R R the rotation by π / 4 \pi/4 π /4 . The reflection F F F in the line at angle π / 8 \pi/8 π /8 interchanges A A A and B B B , hence A 1 / 2 A^{1/2} A 1/2 and B 1 / 2 B^{1/2} B 1/2 , so it commutes with A + B A+B A + B , with S = A 1 / 2 + B 1 / 2 S=A^{1/2}+B^{1/2} S = A 1/2 + B 1/2 and with J = A 1 / 2 B 1 / 2 + B 1 / 2 A 1 / 2 J=A^{1/2}B^{1/2}+B^{1/2}A^{1/2} J = A 1/2 B 1/2 + B 1/2 A 1/2 . Since F F F has the simple eigenvalues ± 1 \pm1 ± 1 , these three matrices are diagonal in its eigenbasis u 1 , u 2 u_1,u_2 u 1 , u 2 , and S 2 = A + B + J S^2=A+B+J S 2 = A + B + J gives ⟨ u i , S u i ⟩ 2 = ⟨ u i , ( A + B ) u i ⟩ + ⟨ u i , J u i ⟩ \langle u_i,Su_i\rangle^2=\langle u_i,(A+B)u_i\rangle+\langle u_i,Ju_i\rangle ⟨ u i , S u i ⟩ 2 = ⟨ u i , ( A + B ) u i ⟩ + ⟨ u i , J u i ⟩ . The pair A 1 / 2 A^{1/2} A 1/2 , B 1 / 2 B^{1/2} B 1/2 is the configuration of the converse in Theorem 9.2 (a) with δ A = δ B = t \delta_A=\delta_B=t δ A = δ B = t , so J J J is indefinite since 2 t 2 > 1 2t^2>1 2 t 2 > 1 , and for the u i u_i u i with ⟨ u i , J u i ⟩ < 0 \langle u_i,Ju_i\rangle<0 ⟨ u i , J u i ⟩ < 0 we get ⟨ u i , ( A 1 / 2 + B 1 / 2 ) u i ⟩ < ⟨ u i , ( A + B ) 1 / 2 u i ⟩ \langle u_i,(A^{1/2}+B^{1/2})u_i\rangle<\langle u_i,(A+B)^{1/2}u_i\rangle ⟨ u i , ( A 1/2 + B 1/2 ) u i ⟩ < ⟨ u i , ( A + B ) 1/2 u i ⟩ . On M n M_n M n replace A , B A,B A , B by A ⊕ 1 n − 2 A\oplus1_{n-2} A ⊕ 1 n − 2 and B ⊕ 1 n − 2 B\oplus1_{n-2} B ⊕ 1 n − 2 , which have the same contrast, and let ρ \rho ρ have u i u_i u i as eigenvector for its largest eigenvalue and all other eigenvalues equal. As δ → 1 \delta\to1 δ → 1 , φ ρ → n ⟨ u i , ⋅ u i ⟩ \varphi_\rho\to n\langle u_i,\cdot\,u_i\rangle φ ρ → n ⟨ u i , ⋅ u i ⟩ , so φ ρ ( ( A + B ) 1 / 2 ) > φ ρ ( A 1 / 2 ) + φ ρ ( B 1 / 2 ) \varphi_\rho((A+B)^{1/2})>\varphi_\rho(A^{1/2})+\varphi_\rho(B^{1/2}) φ ρ (( A + B ) 1/2 ) > φ ρ ( A 1/2 ) + φ ρ ( B 1/2 ) for all δ \delta δ close enough to 1 1 1 . ◻
Figure 6. (a) The contrast square: level curves of the Gardner envelope ( 1 − δ 2 δ ′ 2 ) − 1 / 2 − 1 (1-\delta^2\delta'^2)^{-1/2}-1 ( 1 − δ 2 δ ′2 ) − 1/2 − 1 (blue; the values 0.02 0.02 0.02 , 0.1 0.1 0.1 , 0.3 0.3 0.3 , 1 1 1 ), the band in which item 12 cannot be violated (green, Proposition 9.5 ), the exact violation thresholds of item 24 for n = 2 n=2 n = 2 and n = 3 n=3 n = 3 (orange, Proposition 9.6 ), and, in black, the interval of grid values of δ ′ \delta' δ ′ in which the numerical search on M 2 M_2 M 2 first finds a violation of item 12, at δ = 0.3 \delta=0.3 δ = 0.3 , 0.6 0.6 0.6 and 0.9 0.9 0.9 . (b) The plane of the two operator contrasts: A B + B A ≥ 0 AB+BA\ge0 A B + B A ≥ 0 for all relative positions inside the quarter disc (blue, Theorem 9.2 ), and ( A + B ) 1 / 2 ≤ A 1 / 2 + B 1 / 2 (A+B)^{1/2}\le A^{1/2}+B^{1/2} ( A + B ) 1/2 ≤ A 1/2 + B 1/2 in the larger green region (Proposition 9.5 ). In the figure Δ ( ⋅ ) = O M C ( ⋅ ) \Delta(\cdot)=\mathop{\mathrm{OMC}}(\cdot) Δ ( ⋅ ) = OMC ( ⋅ ) , δ \delta δ is the contrast of the density and δ ′ \delta' δ ′ that of the operators.
The proved threshold is consistent with the computation: on M 2 M_2 M 2 and M 3 M_3 M 3 the search finds no violation of item 12 for δ ′ ≤ 0.95 \delta'\le0.95 δ ′ ≤ 0.95 and δ ∈ { 0.3 , 0.6 , 0.9 } \delta\in\{0.3,0.6,0.9\} δ ∈ { 0.3 , 0.6 , 0.9 } , and the first violation above 10 − 4 10^{-4} 1 0 − 4 at δ ′ = \delta'= δ ′ = 0.999 for δ = 0.3 \delta=0.3 δ = 0.3 , 0.99 for δ = 0.6 \delta=0.6 δ = 0.6 , 0.99 for δ = 0.9 \delta=0.9 δ = 0.9 (M 2 M_2 M 2 and M 3 M_3 M 3 alike); all these grid values lie above 2 2 / 3 2\sqrt2/3 2 2 /3 (Figure 6 (a)). Unlike item 24 below, whose whole threshold curve we find for n = 2 n=2 n = 2 , here sharpness is proved only in the limit δ → 1 \delta\to1 δ → 1 ; at δ = 0.3 \delta=0.3 δ = 0.3 , 0.6 0.6 0.6 and 0.9 0.9 0.9 the search finds the onset of violations higher. The criterion of Moslehian and Najafi [147 Theorem 2.1] is that A B + B A ≥ 0 AB+BA\ge0 A B + B A ≥ 0 if and only if f ( A + B ) ≤ f ( A ) + f ( B ) f(A+B)\le f(A)+f(B) f ( A + B ) ≤ f ( A ) + f ( B ) for all non-negative operator monotone functions f f f . For the square root alone the weaker hypothesis A 1 / 2 B 1 / 2 + B 1 / 2 A 1 / 2 ≥ 0 A^{1/2}B^{1/2}+B^{1/2}A^{1/2}\ge0 A 1/2 B 1/2 + B 1/2 A 1/2 ≥ 0 suffices, and the proof of sharpness shows that in the extremal configuration it is also necessary.
For the power-mean inequality φ ( A 2 ) 1 / 2 ≤ φ ( A ) \varphi(A^2)^{1/2}\le\varphi(A) φ ( A 2 ) 1/2 ≤ φ ( A ) with φ ( 1 ) = n \varphi(1)=n φ ( 1 ) = n (item 24) a Kantorovich-type argument gives the envelope with both contrasts prescribed in closed form, and with it a sharp threshold.
Proposition 9.6 (Item 24 with both contrasts). Let n ≥ 2 n\ge2 n ≥ 2 , O M C ( ρ ) = δ < 1 \mathop{\mathrm{OMC}}(\rho)=\delta<1 OMC ( ρ ) = δ < 1 , T r ρ = n \mathop{\mathrm{Tr}}\rho=n Tr ρ = n , and let A ≥ 0 A\ge0 A ≥ 0 have O M C ( A ) = δ ′ \mathop{\mathrm{OMC}}(A)=\delta' OMC ( A ) = δ ′ . Put ℓ = 1 − δ 1 + δ \ell=\frac{1-\delta}{1+\delta} ℓ = 1 + δ 1 − δ , ℓ ′ = 1 − δ ′ 1 + δ ′ \ell'=\frac{1-\delta'}{1+\delta'} ℓ ′ = 1 + δ ′ 1 − δ ′ , x 0 = n ℓ ′ + ( 1 − ℓ ′ ) n ℓ n − 1 + ℓ , x ∗ = 2 n ℓ ′ 1 + ℓ ′ , f ( x ) = ( 1 + ℓ ′ ) x − n ℓ ′ x . x_0=n\ell'+(1-\ell')\frac{n\ell}{n-1+\ell},\qquad x_*=\frac{2n\ell'}{1+\ell'},\qquad
f(x)=\frac{\sqrt{(1+\ell')x-n\ell'}}x . x 0 = n ℓ ′ + ( 1 − ℓ ′ ) n − 1 + ℓ n ℓ , x ∗ = 1 + ℓ ′ 2 n ℓ ′ , f ( x ) = x ( 1 + ℓ ′ ) x − n ℓ ′ . Then φ ρ ( A 2 ) 1 / 2 / φ ρ ( A ) ≤ f ( max ( x 0 , x ∗ ) ) \varphi_\rho(A^2)^{1/2}/\varphi_\rho(A)\le f(\max(x_0,x_*)) φ ρ ( A 2 ) 1/2 / φ ρ ( A ) ≤ f ( max ( x 0 , x ∗ )) , so V n ( δ , δ ′ ) ≤ f ( max ( x 0 , x ∗ ) ) − 1 V_n(\delta,\delta')\le f(\max(x_0,x_*))-1 V n ( δ , δ ′ ) ≤ f ( max ( x 0 , x ∗ )) − 1 , with equality if x 0 ≥ x ∗ x_0\ge x_* x 0 ≥ x ∗ or ( n − 1 ) ℓ ℓ ′ ≤ 1 (n-1)\ell\ell'\le1 ( n − 1 ) ℓ ℓ ′ ≤ 1 , in particular always for n = 2 n=2 n = 2 . Consequently:
item 24 is violated for no density if and only if δ ′ ≤ ( n − 1 ) / n \delta'\le\sqrt{(n-1)/n} δ ′ ≤ ( n − 1 ) / n ; for δ ′ > ( n − 1 ) / n \delta'>\sqrt{(n-1)/n} δ ′ > ( n − 1 ) / n it is violated for every δ \delta δ close enough to 1 1 1 ;
for n = 2 n=2 n = 2 the violation threshold is the curve f ( max ( x 0 , x ∗ ) ) = 1 f(\max(x_0,x_*))=1 f ( max ( x 0 , x ∗ )) = 1 in the ( δ , δ ′ ) (\delta,\delta') ( δ , δ ′ ) -square, and it equals 1 / 2 1/\sqrt2 1/ 2 for all δ ≥ 1 / 2 \delta\ge1/\sqrt2 δ ≥ 1/ 2 .
Proof. Both sides are homogeneous of degree one in A A A , so let ∥ A ∥ = 1 \|A\|=1 ∥ A ∥ = 1 ; then the spectrum of A A A lies in [ ℓ ′ , 1 ] [\ell',1] [ ℓ ′ , 1 ] and contains both end points. From ( 1 − A ) ( A − ℓ ′ ) ≥ 0 (1-A)(A-\ell')\ge0 ( 1 − A ) ( A − ℓ ′ ) ≥ 0 we get A 2 ≤ ( 1 + ℓ ′ ) A − ℓ ′ A^2\le(1+\ell')A-\ell' A 2 ≤ ( 1 + ℓ ′ ) A − ℓ ′ , hence, with x = φ ( A ) x=\varphi(A) x = φ ( A ) ,
φ ( A 2 ) 1 / 2 φ ( A ) ≤ ( 1 + ℓ ′ ) x − n ℓ ′ x = f ( x ) , \frac{\varphi(A^2)^{1/2}}{\varphi(A)}\le\frac{\sqrt{(1+\ell')x-n\ell'}}x=f(x), φ ( A ) φ ( A 2 ) 1/2 ≤ x ( 1 + ℓ ′ ) x − n ℓ ′ = f ( x ) , with equality for two-level A A A . If e e e is a unit eigenvector of A A A for the eigenvalue 1 1 1 , then A ≥ ℓ ′ + ( 1 − ℓ ′ ) e e ∗ A\ge\ell'+(1-\ell')ee^* A ≥ ℓ ′ + ( 1 − ℓ ′ ) e e ∗ , so x ≥ n ℓ ′ + ( 1 − ℓ ′ ) ⟨ ρ e , e ⟩ ≥ n ℓ ′ + ( 1 − ℓ ′ ) λ min ( ρ ) x\ge n\ell'+(1-\ell')\langle \rho e,e\rangle\ge n\ell'+(1-\ell')\lambda_{\min}(\rho) x ≥ n ℓ ′ + ( 1 − ℓ ′ ) ⟨ ρ e , e ⟩ ≥ n ℓ ′ + ( 1 − ℓ ′ ) λ m i n ( ρ ) , and λ min ( ρ ) ≥ n ℓ / ( n − 1 + ℓ ) \lambda_{\min}(\rho)\ge n\ell/(n-1+\ell) λ m i n ( ρ ) ≥ n ℓ / ( n − 1 + ℓ ) for T r ρ = n \mathop{\mathrm{Tr}}\rho=n Tr ρ = n (proof of Proposition 5.37 ); so x ≥ x 0 x\ge x_0 x ≥ x 0 . The function f f f increases on [ n ℓ ′ / ( 1 + ℓ ′ ) , x ∗ ] [n\ell'/(1+\ell'),x_*] [ n ℓ ′ / ( 1 + ℓ ′ ) , x ∗ ] and decreases on [ x ∗ , ∞ ) [x_*,\infty) [ x ∗ , ∞ ) , since d d x f ( x ) 2 = x − 3 ( 2 n ℓ ′ − ( 1 + ℓ ′ ) x ) \frac d{dx}f(x)^2=x^{-3}\bigl(2n\ell'-(1+\ell')x\bigr) d x d f ( x ) 2 = x − 3 ( 2 n ℓ ′ − ( 1 + ℓ ′ ) x ) ; hence f ( x ) ≤ f ( max ( x 0 , x ∗ ) ) f(x)\le f(\max(x_0,x_*)) f ( x ) ≤ f ( max ( x 0 , x ∗ )) . For equality take A = ℓ ′ + ( 1 − ℓ ′ ) e e ∗ A=\ell'+(1-\ell')ee^* A = ℓ ′ + ( 1 − ℓ ′ ) e e ∗ with a unit vector e e e , so that x = n ℓ ′ + ( 1 − ℓ ′ ) ⟨ ρ e , e ⟩ x=n\ell'+(1-\ell')\langle \rho e,e\rangle x = n ℓ ′ + ( 1 − ℓ ′ ) ⟨ ρ e , e ⟩ . If x 0 ≥ x ∗ x_0\ge x_* x 0 ≥ x ∗ , let ρ \rho ρ have the spectrum { a , … , a , b } \{a,\dots,a,b\} { a , … , a , b } and e e e an eigenvector for b b b ; then x = x 0 x=x_0 x = x 0 . Otherwise we need ⟨ ρ e , e ⟩ = n ℓ ′ / ( 1 + ℓ ′ ) \langle \rho e,e\rangle=n\ell'/(1+\ell') ⟨ ρ e , e ⟩ = n ℓ ′ / ( 1 + ℓ ′ ) ; the values ⟨ ρ e , e ⟩ \langle \rho e,e\rangle ⟨ ρ e , e ⟩ for the spectra { a , … , a , b } \{a,\dots,a,b\} { a , … , a , b } and { a , b , … , b } \{a,b,\dots,b\} { a , b , … , b } (normalised to T r ρ = n \mathop{\mathrm{Tr}}\rho=n Tr ρ = n ) fill the interval [ n ℓ / ( n − 1 + ℓ ) , n / ( 1 + ( n − 1 ) ℓ ) ] [n\ell/(n-1+\ell),\,n/(1+(n-1)\ell)] [ n ℓ / ( n − 1 + ℓ ) , n / ( 1 + ( n − 1 ) ℓ )] , whose left end is at most n ℓ ′ / ( 1 + ℓ ′ ) n\ell'/(1+\ell') n ℓ ′ / ( 1 + ℓ ′ ) because x 0 ≤ x ∗ x_0\le x_* x 0 ≤ x ∗ , and whose right end is at least n ℓ ′ / ( 1 + ℓ ′ ) n\ell'/(1+\ell') n ℓ ′ / ( 1 + ℓ ′ ) if ( n − 1 ) ℓ ℓ ′ ≤ 1 (n-1)\ell\ell'\le1 ( n − 1 ) ℓ ℓ ′ ≤ 1 .
(a) Since f ( max ( x 0 , x ∗ ) ) ≤ f ( x ∗ ) = ( 1 + ℓ ′ ) / ( 2 n ℓ ′ ) f(\max(x_0,x_*))\le f(x_*)=(1+\ell')/(2\sqrt{n\ell'}) f ( max ( x 0 , x ∗ )) ≤ f ( x ∗ ) = ( 1 + ℓ ′ ) / ( 2 n ℓ ′ ) , there is no violation if ( 1 + ℓ ′ ) 2 ≤ 4 n ℓ ′ (1+\ell')^2\le4n\ell' ( 1 + ℓ ′ ) 2 ≤ 4 n ℓ ′ , i.e. ℓ ′ ≥ ( n − n − 1 ) 2 \ell'\ge(\sqrt n-\sqrt{n-1})^2 ℓ ′ ≥ ( n − n − 1 ) 2 , i.e. δ ′ ≤ ( n − 1 ) / n \delta'\le\sqrt{(n-1)/n} δ ′ ≤ ( n − 1 ) / n . Conversely, as δ → 1 \delta\to1 δ → 1 we have ℓ → 0 \ell\to0 ℓ → 0 , so x 0 → n ℓ ′ < x ∗ x_0\to n\ell'<x_* x 0 → n ℓ ′ < x ∗ and ( n − 1 ) ℓ ℓ ′ ≤ 1 (n-1)\ell\ell'\le1 ( n − 1 ) ℓ ℓ ′ ≤ 1 , and the envelope tends to f ( x ∗ ) − 1 > 0 f(x_*)-1>0 f ( x ∗ ) − 1 > 0 when ( 1 + ℓ ′ ) 2 > 4 n ℓ ′ (1+\ell')^2>4n\ell' ( 1 + ℓ ′ ) 2 > 4 n ℓ ′ . (b) For n = 2 n=2 n = 2 equality always holds; x 0 ≤ x ∗ x_0\le x_* x 0 ≤ x ∗ means 1 − δ ≤ 2 ℓ ′ / ( 1 + ℓ ′ ) 1-\delta\le2\ell'/(1+\ell') 1 − δ ≤ 2 ℓ ′ / ( 1 + ℓ ′ ) , and at ℓ ′ = ( 2 − 1 ) 2 \ell'=(\sqrt2-1)^2 ℓ ′ = ( 2 − 1 ) 2 this is δ ≥ 1 / 2 \delta\ge1/\sqrt2 δ ≥ 1/ 2 . ◻
The computed envelopes of item 24 on M 2 M_2 M 2 and M 3 M_3 M 3 with both contrasts prescribed agree at every grid point with Proposition 9.6 to 2 × 10 − 11 2\times 10^{-11} 2 × 1 0 − 11 (absolute difference). The threshold rises above ( n − 1 ) / n \sqrt{(n-1)/n} ( n − 1 ) / n for smaller δ \delta δ and tends to 1 1 1 as δ → 0 \delta\to0 δ → 0 (Figure 6 (a)): the less tracial the functional, the more moderate the operators that detect it. For monotonicity (item 10) an argument of this kind is not available: for 0 ≤ A ≤ B = A + t v v ∗ 0\le A\le B=A+tvv^* 0 ≤ A ≤ B = A + t v v ∗ with a unit vector v v v that is not an eigenvector of A A A , B 2 − A 2 = t ( ( A + t 2 ) v v ∗ + v v ∗ ( A + t 2 ) ) B^2-A^2=t\bigl((A+\frac t2)vv^*+vv^*(A+\frac t2)\bigr) B 2 − A 2 = t ( ( A + 2 t ) v v ∗ + v v ∗ ( A + 2 t ) ) is indefinite for every t > 0 t>0 t > 0 , however small the contrasts of A A A and B B B , so the operator inequality fails and only the functional can prevent a violation. Correspondingly, the numerically observed onset of violations of item 10 depends on δ \delta δ : on M 2 M_2 M 2 the first grid value of δ ′ \delta' δ ′ with a violation above 10 − 4 10^{-4} 1 0 − 4 is 0.99 for δ = 0.3 \delta=0.3 δ = 0.3 , 0.95 for δ = 0.6 \delta=0.6 δ = 0.6 , 0.8 for δ = 0.9 \delta=0.9 δ = 0.9 .
Figure 7. Contrast of the Jordan product A B + B A AB+BA A B + B A against O M C ( A ) < 0.8 \mathop{\mathrm{OMC}}(A)<0.8 OMC ( A ) < 0.8 , with O M C ( B ) = 0.6 \mathop{\mathrm{OMC}}(B)=0.6 OMC ( B ) = 0.6 (for O M C ( A ) ≤ 0.8 \mathop{\mathrm{OMC}}(A)\le0.8 OMC ( A ) ≤ 0.8 it is positive, Theorem 9.2 (a)): the bounds of Theorem 9.2 (b), the largest value of Proposition 9.3 (orange; on M 3 M_3 M 3 and M 4 M_4 M 4 the value ( 2 − μ ) / ( 2 + μ ) (2-\mu)/(2+\mu) ( 2 − μ ) / ( 2 + μ ) ), and the largest value found by optimisation (markers). In the figure Δ ( ⋅ ) = O M C ( ⋅ ) \Delta(\cdot)=\mathop{\mathrm{OMC}}(\cdot) Δ ( ⋅ ) = OMC ( ⋅ ) .
Figure 8. The commutator test ∣ φ ρ ( A B − B A ) ∣ / ( λ ˉ ( A ) λ ˉ ( B ) ) |\varphi_\rho(AB-BA)|/(\bar\lambda(A)\bar\lambda(B)) ∣ φ ρ ( A B − B A ) ∣/ ( λ ˉ ( A ) λ ˉ ( B )) for random states φ ρ \varphi_\rho φ ρ and random A , B A,B A , B against O M C ( ρ ) \mathop{\mathrm{OMC}}(\rho) OMC ( ρ ) , with O M C ( A ) = O M C ( B ) = 0.6 \mathop{\mathrm{OMC}}(A)=\mathop{\mathrm{OMC}}(B)=0.6 OMC ( A ) = OMC ( B ) = 0.6 , and the sharp bound 2 c n ( δ ) ⋅ 0.36 2c_n(\delta)\cdot0.36 2 c n ( δ ) ⋅ 0.36 of Proposition 9.4 (b). In the figure φ W = φ ρ \varphi_W=\varphi_\rho φ W = φ ρ , Δ ( W ) = O M C ( ρ ) = δ \Delta(W)=\mathop{\mathrm{OMC}}(\rho)=\delta Δ ( W ) = OMC ( ρ ) = δ and Δ ( ⋅ ) = O M C ( ⋅ ) \Delta(\cdot)=\mathop{\mathrm{OMC}}(\cdot) Δ ( ⋅ ) = OMC ( ⋅ ) .
Numerical evidence
Method
The envelopes were computed by constrained gradient ascent with the library of [149 ] (the programs trace_characterizations.py and trace_contrast.py of its repository). The random sampling (trace_montecarlo.py), the computations for composite operators (contrast_operations.py) and Figure 3 (trace_ envelopes_figures.py) are plain NumPy/SciPy programs in the same repository that do not use the library. The density ρ \rho ρ is diagonal with a spectrum parametrised so that O M C ( ρ ) = δ \mathop{\mathrm{OMC}}(\rho)=\delta OMC ( ρ ) = δ holds exactly, the other eigenvalues being free. Positive variables have a free contrast in [ 0 , 1 ] [0,1] [ 0 , 1 ] , so rank-deficient operators are included; there are self-adjoint and unitary variables (polar factors, and projections U D U ∗ UDU^* U D U ∗ with every rank pattern spread over the starts) and positive scalars. In item 2 the search runs over self-adjoint X , Y X,Y X , Y only, which contain the extremiser of Corollary 5.6 . Half of the starts begin at structured points (two-level spectra, rotations in pairs of eigenvectors of ρ \rho ρ ). The lattice operations are the spectral projections of P + Q P+Q P + Q for the eigenvalue 2 2 2 (P ∧ Q P\wedge Q P ∧ Q ) and for its range (P ∨ Q P\vee Q P ∨ Q ), computed by functional calculus with an indicator function. In items 19 and 23 the suprema are approached as P P P and Q Q Q merge, where P ∧ Q P\wedge Q P ∧ Q and P ∨ Q P\vee Q P ∨ Q jump; there the search maximises ( L − R ) / N − 10 − 12 / ∥ P − Q ∥ F 2 (\mathcal{L}-\mathcal{R})/\mathcal{N}-10^{-12}/\|P-Q\|_{\mathrm F}^2 ( L − R ) / N − 1 0 − 12 /∥ P − Q ∥ F 2 , which keeps P P P and Q Q Q about 10 − 3 10^{-3} 1 0 − 3 apart, while the reported values are those of ( L − R ) / N (\mathcal{L}-\mathcal{R})/\mathcal{N} ( L − R ) / N . A grid of 14 14 14 contrasts times 64 64 64 starts forms one batch per item and dimension (800 800 800 gradient steps, n = 2 , 3 , 4 n=2, 3, 4 n = 2 , 3 , 4 ).
The best final point for every contrast is recomputed with 40 40 40 -digit arithmetic (mpmath), after the unitaries are re-orthonormalised and the contrast is recomputed from the stored eigenvalues. Every reported value is therefore the exact violation at an exactly admissible point, and so a rigorous lower bound for the supremum; we call it the computed envelope. It is not an upper bound: where no theorem is available, the true envelope may be larger. The double-precision and high-precision values agree to 5 × 10 − 6 5\times 10^{-6} 5 × 1 0 − 6 (relative), and at δ = 0 \delta=0 δ = 0 no value exceeds 1 × 10 − 16 1\times 10^{-16} 1 × 1 0 − 16 . For items 12–16, independent NumPy/SciPy searches (BFGS from 40 40 40 random starts on M 2 M_2 M 2 at δ = 0.3 , 0.5 , 0.9 \delta=0.3,0.5,0.9 δ = 0.3 , 0.5 , 0.9 ) give values such that the ratio of the batched search to them lies in [ 0.994 , 1.097 ] [0.994,1.097] [ 0.994 , 1.097 ] ; the ratios above 1 1 1 come from item 13, where BFGS stops earlier. The whole computation takes 475 475 475 s on 4 4 4 CPU processes.
Envelopes with the operators free
Table 4. The items with a proved envelope or lower bound, ordered by the order of vanishing at the trace. V ⋆ V^\star V ⋆ : the sharp value or, where marked ≥ \ge ≥ , a proved lower bound; “search”: the computed envelope on M 2 M_2 M 2 (a range over the items of a row). Items 11 and 24 are given for n = 2 n=2 n = 2 .
Table 5. Computed envelopes (certified lower bounds) on M 2 M_2 M 2 at δ = 0.1 , 0.5 , 0.9 \delta=0.1,0.5,0.9 δ = 0.1 , 0.5 , 0.9 and the proved values V ⋆ V^\star V ⋆ ; a ^a a : sharp for n = 2 n=2 n = 2 only, b ^b b : a lower bound, c ^c c : approached but not attained. “order”: leading order of V ⋆ V^\star V ⋆ as δ → 0 \delta\to0 δ → 0 ; “slope”: log 2 ( V ( 0.2 ) / V ( 0.1 ) ) \log_2(V(0.2)/V(0.1)) log 2 ( V ( 0.2 ) / V ( 0.1 )) of the search; “n n n -dep.”: the largest ∣ V n / V 2 − 1 ∣ |V_n/V_2-1| ∣ V n / V 2 − 1∣ for n = 3 , 4 n=3,4 n = 3 , 4 and δ ≥ 0.1 \delta\ge0.1 δ ≥ 0.1 (for item 7 it reflects search shortfalls, for item 13 a genuine increase with n n n ; items 19 and 20 differ for odd n n n , Theorem 5.34 ). “n.r.”: no violation resolved by the search (full run on 4 CPU processes with one thread each).
Table 4 and Table 5 , Figure 2 and the top rows of Figure 9 –Figure 33 summarise the computation. Where an envelope is known exactly and attained (items 1–5, 8, 11, and 9 for n = 2 n=2 n = 2 ; for item 10, exact by Theorem 5.27 and Remark 5.28 , see § 10.6 ), the search reproduces it to 1 × 10 − 7 1\times 10^{-7} 1 × 1 0 − 7 for δ ≥ 0.05 \delta\ge0.05 δ ≥ 0.05 and n = 2 , 3 , 4 n=2,3,4 n = 2 , 3 , 4 ; for items 17, 18, 20–22 and 24 it does so to 3 × 10 − 4 3\times 10^{-4} 3 × 1 0 − 4 , and for the approached suprema of 19 and 23 to 2 × 10 − 6 2\times 10^{-6} 2 × 1 0 − 6 . The values of 17, 18 and 21–23 on M 3 M_3 M 3 and M 4 M_4 M 4 agree with those on M 2 M_2 M 2 to 4 × 10 − 7 4\times 10^{-7} 4 × 1 0 − 7 , as Theorem 5.32 , Corollary 5.6 and Propositions 5.16 and 5.33 predict, and on M 3 M_3 M 3 the search finds the odd-dimensional value 0.0694 0.0694 0.0694 of item 20 at δ = 0.5 \delta=0.5 δ = 0.5 (M 2 M_2 M 2 : 0.0802 0.0802 0.0802 ). These agreements are a check of the method rather than new information.
The new information concerns the items without a proved envelope. All five items 12–16 fail quadratically, with V ≈ 0.5 δ 2 V\approx0.5\,\delta^2 V ≈ 0.5 δ 2 (items 12 and 13), δ 2 / 36 \delta^2/36 δ 2 /36 (item 14, equal to the lower bound of Proposition 5.35 to leading order), 0.064 δ 2 0.064\,\delta^2 0.064 δ 2 (item 15) and 0.09 δ 2 0.09\,\delta^2 0.09 δ 2 (item 16). Item 15 decreases with n n n , because the identity part of e P / 2 e Q e P / 2 e^{P/2}e^Qe^{P/2} e P /2 e Q e P /2 enters N \mathcal{N} N , whereas item 13 increases with n n n at large δ \delta δ (on M 3 M_3 M 3 at δ = 0.9 \delta=0.9 δ = 0.9 the value 0.3064 0.3064 0.3064 against 0.2197 0.2197 0.2197 on M 2 M_2 M 2 ); items 12, 14 and 16 give the same values for n = 2 , 3 , 4 n=2,3,4 n = 2 , 3 , 4 (to 4 × 10 − 3 4\times 10^{-3} 4 × 1 0 − 3 ). In item 16 the best points have large ∥ A ∥ \|A\| ∥ A ∥ and ∥ B ∥ \|B\| ∥ B ∥ , and the supremum is approached as the scale grows. Among items 6, 7 and 25 (marked b ^b b in Table 5 ) and item 10 (exact, Theorem 5.27 ), whose computed envelopes were compared with the proved bounds, no computed value exceeds the bound (the largest relative difference, search minus bound, is − 1 × 10 − 13 -1\times 10^{-13} − 1 × 1 0 − 13 ) except for item 6 in dimension n ≥ 3 n\ge3 n ≥ 3 : on M 2 M_2 M 2 the search never exceeds G − 1 G-1 G − 1 , but for n ≥ 3 n\ge3 n ≥ 3 it finds larger violations, V ≥ 0.847 V\ge0.847 V ≥ 0.847 at n = 4 n=4 n = 4 , δ = 0.8 \delta=0.8 δ = 0.8 , where G − 1 = 0.667 G-1=0.667 G − 1 = 0.667 . At n = 3 n=3 n = 3 , δ = 0.9 \delta=0.9 δ = 0.9 the present run reaches only 1.2372 < G − 1 = 1.2942 1.2372<G-1=1.2942 1.2372 < G − 1 = 1.2942 , but a point found by an earlier run and stored in the test suite of the repository gives V > 2.1 V>2.1 V > 2.1 there (recomputed with 60 digits). For item 6 we therefore only know G − 1 ≤ V n ≤ κ − 1 G-1\le V_n\le \kappa-1 G − 1 ≤ V n ≤ κ − 1 (Proposition 5.43 ). Items 9 and 10 give the same values for n = 3 , 4 n=3,4 n = 3 , 4 as for n = 2 n=2 n = 2 , to 3 × 10 − 10 3\times 10^{-10} 3 × 1 0 − 10 and 3 × 10 − 8 3\times 10^{-8} 3 × 1 0 − 8 .
Except for items 6 (n ≥ 3 n\ge3 n ≥ 3 ), 11 (A = ρ − 1 A=\rho^{-1} A = ρ − 1 ) and 25 (A A A of rank two with eigenvalue ratio ≈ δ 2 / 8 \approx\delta^2/8 ≈ δ 2 /8 ), the best points are rank-one operators or matrix units in the span of eigenvectors of ρ \rho ρ for its extreme eigenvalues, and they do not commute with ρ \rho ρ . The search has two limitations. The proved lower bounds of items 6 and 7 are approached only as an operator becomes singular; gradient ascent reaches 0.924 0.924 0.924 and 0.796 0.796 0.796 of the lower bound at δ = 0.5 \delta=0.5 δ = 0.5 and finds no violation of item 6 for δ ≤ 0.2 \delta\le0.2 δ ≤ 0.2 . Item 25 is resolved for δ ≥ 0.3 \delta\ge0.3 δ ≥ 0.3 ; the search reaches g 25 g_{25} g 25 to 10 − 3 10^{-3} 1 0 − 3 for δ ≥ 0.7 \delta\ge0.7 δ ≥ 0.7 and does not exceed it. Finally, on M 2 ⊕ M 3 M_2\oplus M_3 M 2 ⊕ M 3 with ρ = 1 ⊕ s 1 \rho=1\oplus s1 ρ = 1 ⊕ s 1 and ρ = d i a g ( 1 , ℓ ) ⊕ s 1 \rho=\mathrm{diag}(1,\ell)\oplus s1 ρ = diag ( 1 , ℓ ) ⊕ s 1 , s ∈ { 1 4 , 4 } s\in\{\frac14,4\} s ∈ { 4 1 , 4 } , and intra-block contrast δ ∈ { 0.3 , 0.5 , 0.7 } \delta\in\{0.3,0.5,0.7\} δ ∈ { 0.3 , 0.5 , 0.7 } , the computation confirms Corollary 8.2 : for central ρ \rho ρ the largest value found for the items that hold for every trace is 1 × 10 − 16 1\times 10^{-16} 1 × 1 0 − 16 ; for items 11 and 24 the search reproduces the exact values for the fixed density to 2 × 10 − 11 2\times 10^{-11} 2 × 1 0 − 11 ; with intra-block contrast no value exceeds the M 2 M_2 M 2 envelope (largest relative difference − 7 × 10 − 14 -7\times 10^{-14} − 7 × 1 0 − 14 ), and where Corollary 8.2 gives equality the search reaches it to 10 − 3 10^{-3} 1 0 − 3 in 63 63 63 of 72 72 72 cases.
Both contrasts
For the 14 items with operator variables that are not projections, a second computation prescribes the contrast δ ′ \delta' δ ′ of all operator variables as well (in items 10 and 25 the variables of prescribed contrast are A A A and B − A B-A B − A ; 64 64 64 starts per grid point, 800 800 800 gradient steps, best point recomputed with 40 40 40 digits; 760 760 760 s on four CPU processes). The bottom rows of Figure 9 –Figure 33 (Appendix § 12 ) show the computed V ( 0.9 , δ ′ ) V(0.9,\delta') V ( 0.9 , δ ′ ) . At δ ′ = 0 \delta'=0 δ ′ = 0 no item is violated, as Proposition 3.5 requires, but beyond that the dependence on δ ′ \delta' δ ′ takes several forms. For item 1 the values follow Theorem 6.1 to 1 × 10 − 9 1\times 10^{-9} 1 × 1 0 − 9 (relative, n = 2 , 3 n=2,3 n = 2 , 3 , all grid points). For items 2 and 3 the theorem gives only the upper bound G ( δ δ ′ ) − 1 G(\delta\delta')-1 G ( δ δ ′ ) − 1 ; on M 2 M_2 M 2 the search reaches it (to 9 × 10 − 10 9\times 10^{-10} 9 × 1 0 − 10 ), and on M 3 M_3 M 3 it reaches the odd-dimensional value of item 1 up to the relative differences − 2 × 10 − 4 -2\times 10^{-4} − 2 × 1 0 − 4 (item 2) and − 2 × 10 − 9 -2\times 10^{-9} − 2 × 1 0 − 9 (item 3). Item 8 follows Proposition 5.22 on M 2 M_2 M 2 ; here the violation grows smoothly with both contrasts. Items 6, 10, 12, 13, 24 and 25, in contrast, show no violation above 10 − 4 10^{-4} 1 0 − 4 for δ ′ < 0.8 \delta'<0.8 δ ′ < 0.8 : at δ = 0.9 \delta=0.9 δ = 0.9 the first grid value of δ ′ \delta' δ ′ with a violation above 10 − 4 10^{-4} 1 0 − 4 is 6 (n = 2 n=2 n = 2 ): 0.9; 6 (n = 3 n=3 n = 3 ): 0.8; 10 (n = 2 n=2 n = 2 ): 0.8; 10 (n = 3 n=3 n = 3 ): 0.8; 12 (n = 2 n=2 n = 2 ): 0.99; 12 (n = 3 n=3 n = 3 ): 0.99; 13 (n = 2 n=2 n = 2 ): 0.9; 13 (n = 3 n=3 n = 3 ): 0.9; 24 (n = 2 n=2 n = 2 ): 0.8; 24 (n = 3 n=3 n = 3 ): 0.9; 25 (n = 2 n=2 n = 2 ): –; 25 (n = 3 n=3 n = 3 ): 0.99 (“–”: none on the grid). For items 12 and 24 § 9.1 proves such thresholds, and for item 24 the computed values agree with the exact envelope of Proposition 9.6 ; for items 6, 10, 13 and 25 the thresholds remain observations. The determinant inequality (item 11) behaves differently again: it is violated most strongly when the contrast of A A A is close to that of ρ \rho ρ , since its extremiser is ρ − 1 \rho^{-1} ρ − 1 , and in M 2 M_2 M 2 items 4 and 7 also have an interior maximum in δ ′ \delta' δ ′ (for item 4 at the contrast of ∣ A ∣ |A| ∣ A ∣ for the extremiser A = a e 1 e 2 ∗ + b e 2 e 1 ∗ A=\sqrt a\,e_1e_2^*+\sqrt b\,e_2e_1^* A = a e 1 e 2 ∗ + b e 2 e 1 ∗ ), which the additional dimension of M 3 M_3 M 3 removes. Finally, item 16 does not see the contrast of the operators at all: it is invariant under A ↦ A + c A\mapsto A+c A ↦ A + c , B ↦ B + c B\mapsto B+c B ↦ B + c , which makes both contrasts arbitrarily small without changing the quotient, so V ( δ , δ ′ ) V(\delta,\delta') V ( δ , δ ′ ) equals the envelope with free operators for every δ ′ > 0 \delta'>0 δ ′ > 0 and jumps to 0 0 0 at δ ′ = 0 \delta'=0 δ ′ = 0 .
The same computation restricted to two-level operators of contrast δ ′ \delta' δ ′ reproduces the violations found with general spectra on M 3 M_{3} M 3 (to 1 % 1\% 1% , wherever they exceed 10 − 4 10^{-4} 1 0 − 4 ) except in the items and at the grid values 2 (δ ′ ∈ { 0.4 , 0.6 } \delta'\in\{0.4, 0.6\} δ ′ ∈ { 0.4 , 0.6 } ); 4 (δ ′ ≥ 0.2 \delta'\ge 0.2 δ ′ ≥ 0.2 ); 6 (δ ′ ≥ 0.8 \delta'\ge 0.8 δ ′ ≥ 0.8 ); 7 (δ ′ ≥ 0.6 \delta'\ge 0.6 δ ′ ≥ 0.6 ); 11 (δ ′ ∈ { 0.4 , 0.8 , 0.9 , 0.99 } \delta'\in\{0.4, 0.8, 0.9, 0.99\} δ ′ ∈ { 0.4 , 0.8 , 0.9 , 0.99 } ); 13 (δ ′ ≥ 0.9 \delta'\ge 0.9 δ ′ ≥ 0.9 ); 25 (δ ′ ≥ 0.99 \delta'\ge 0.99 δ ′ ≥ 0.99 ). In item 4 the extremisers on M 3 M_3 M 3 apparently need a third eigenvalue (with the contrast free, the extremiser has the eigenvalues a , b , 0 \sqrt a,\sqrt b,0 a , b , 0 ; § 10.4 ), and in items 6 and 7 the two-level values fall further behind as δ ′ → 1 \delta'\to1 δ ′ → 1 . In item 25 two-level operators give no violation at δ = 0.9 \delta=0.9 δ = 0.9 , δ ′ = 0.99 \delta'=0.99 δ ′ = 0.99 and 0.999 0.999 0.999 , where general spectra give 1.9 × 10 − 4 1.9\times10^{-4} 1.9 × 1 0 − 4 and 1.5 × 10 − 3 1.5\times10^{-3} 1.5 × 1 0 − 3 ; in item 11 they fall short most at δ = 0.6 \delta=0.6 δ = 0.6 , δ ′ = 0.9 \delta'=0.9 δ ′ = 0.9 (0.017 0.017 0.017 against 0.045 0.045 0.045 ), and in item 2 by about 10 % 10\% 10% at δ = 0.3 \delta=0.3 δ = 0.3 . In item 13 at δ = 0.9 \delta=0.9 δ = 0.9 two-level operators give no violation for δ ′ = 0.9 \delta'=0.9 δ ′ = 0.9 –0.99 0.99 0.99 , where general spectra give 0.019 0.019 0.019 –0.11 0.11 0.11 , but this is a failure of the restricted search rather than evidence: on M 3 M_3 M 3 it returns the same values at every δ ′ ≤ 0.99 \delta'\le0.99 δ ′ ≤ 0.99 , i.e. it does not move away from its starting points. On M 2 M_2 M 2 , where every operator of contrast δ ′ \delta' δ ′ is two-level, the two searches should agree; they differ in item 7 (δ = 0.3 \delta=0.3 δ = 0.3 , δ ′ = 0.9 \delta'=0.9 δ ′ = 0.9 ) and in item 13 (δ = 0.9 \delta=0.9 δ = 0.9 , δ ′ = 0.9 \delta'=0.9 δ ′ = 0.9 and 0.95 0.95 0.95 , where the restricted search finds the larger value), which indicates the reliability of the search near δ ′ = 1 \delta'=1 δ ′ = 1 . In item 10 the two classes differ only where neither violates the inequality.
Projections and two-level operators
For the items to which the vertex principle (Lemma 3.6 ) applies, two-level operators attain the supremum, and in eleven items the variables are projections by definition. To see how far this extends, we repeated the search for the other 14 items with the positive variables (the moduli, for the others) restricted to two-level operators c ( ℓ + ( 1 − ℓ ) P ) c(\ell+(1-\ell)P) c ( ℓ + ( 1 − ℓ ) P ) with free ℓ ∈ [ 0 , 1 ] \ell\in[0,1] ℓ ∈ [ 0 , 1 ] , c > 0 c>0 c > 0 and a projection P P P of any rank (for self-adjoint variables α + β P \alpha+\beta P α + β P with free real α , β \alpha,\beta α , β ), and to scaled projections c P cP c P (Table 6 ).
Table 6. Computed envelopes (certified lower bounds) over all operators, over two-level operators and over scaled projections, for the items whose variables are not projections. In M 2 M_2 M 2 every operator is two-level, so there the middle column only checks the restricted parametrisation. “–”: item 11 needs an invertible A A A .
The result is an observation with clear exceptions, not a general principle. Scaled projections attain the computed envelope for items 1, 2, 3, 8, 10, 12, 16 (to within 0.73 % 0.73\% 0.73% ) and 24 in M 2 M_2 M 2 and M 3 M_3 M 3 , and for item 13 in M 2 M_2 M 2 . The restricted searches are less reliable than the unrestricted one: two-level operators include the scaled projections, so every entry of the two-level column below the projection column is a failure of the restricted search, as are, on M 2 M_2 M 2 , where every operator is two-level, the values 0 0 0 for item 6 and 0.0329 0.0329 0.0329 for item 13 at δ = 0.5 \delta=0.5 δ = 0.5 . In items 6 and 7 the supremum is approached only in a degenerate limit in which an operator becomes a projection and both sides of the inequality vanish; restricting the search to projections (item 6) brings it closer to the proved lower bound G − 1 G-1 G − 1 than the unrestricted search, on M 2 M_2 M 2 and on M 3 M_3 M 3 , and restricting it to two-level operators (item 7) brings it closer to G 2 − 1 G^2-1 G 2 − 1 on M 2 M_2 M 2 (but not on M 3 M_3 M 3 ); for item 7 the violation at projections themselves is zero. Two-level operators that are not projections are needed for item 4 (the extremiser has ∣ A ∣ |A| ∣ A ∣ with the eigenvalues a , b \sqrt a,\sqrt b a , b ), item 11 (A = ρ − 1 A=\rho^{-1} A = ρ − 1 ) and item 25 (rank-one A A A gives no violation, Proposition 5.31 ). In M 3 M_3 M 3 the extremiser of item 4 lives on a two-dimensional subspace and vanishes on its complement, so it has the three eigenvalues a , b , 0 \sqrt a,\sqrt b,0 a , b , 0 and the restricted search misses it. Finally, general operators give strictly larger violations than the restricted searches find in two cases: in M 3 M_3 M 3 at δ = 0.9 \delta=0.9 δ = 0.9 item 13 is violated by 0.3048 0.3048 0.3048 , against 0.2197 0.2197 0.2197 for projections and for M 2 M_2 M 2 , and for item 6 with n ≥ 3 n\ge3 n ≥ 3 the unrestricted search of the previous sections exceeds G − 1 G-1 G − 1 , while the restricted searches stay below it. In short, projections and their affine images are a good first test of a characterisation, but they do not always realise the largest violation.
Random sampling
Random sampling is the method of [85 Sect. 5] . We extended it to all items: for n = 2 , 3 , 4 n=2, 3, 4 n = 2 , 3 , 4 and each item we draw 200000 200000 200000 random functionals and operators (ρ \rho ρ with a uniformly distributed contrast and the other eigenvalues uniform in between, positive operators U d i a g ( λ ) U ∗ U\mathrm{diag}(\lambda)U^* U diag ( λ ) U ∗ with Haar-random U U U , uniformly distributed contrast and random scale, arbitrary operators V ∣ X ∣ V|X| V ∣ X ∣ with Haar-random V V V , self-adjoint ones with random signs, and Haar-random projections), and a second sample in which every positive operator (every modulus) is replaced by a scaled projection of random rank. Appendix § 12 shows all samples together with the computed envelopes, and Table 7 the fraction of violating samples.
Table 7. Random sampling, 200000 200000 200000 samples per item, class and dimension: the percentage of samples that violate the inequality, for random operators and for scaled projections. “–”: not applicable (the variables are projections, or an invertible operator is needed).
No sample exceeds the computed envelope by more than the interpolation between grid values allows: the largest ratio of a sampled violation to the envelope, interpolated linearly in O M C ( ρ ) \mathop{\mathrm{OMC}}(\rho) OMC ( ρ ) at the sample’s own contrast, is 1.003 1.003 1.003 over all items and dimensions. Ratios slightly above 1 1 1 are possible because the linear interpolation underestimates the envelope where it is concave between grid values, and for the items without a closed form the computed envelope is only a lower bound. This is an independent consistency check of the computed values. The samples also show why sampling alone is a poor test of a characterisation. Random operators rarely come close to the envelope, and for monotonicity, subadditivity, the Powers–Størmer inequality and the weighted monotonicity fewer than 0.1 % 0.1\% 0.1% of the samples (or none) violate the inequality at all, so a Monte Carlo study alone would suggest that these inequalities hold for every φ ρ \varphi_\rho φ ρ . Sampling also deteriorates with the dimension: for Gardner’s inequality the fraction of violating random samples drops from 4.4 % 4.4\% 4.4% on M 2 M_2 M 2 to below 0.1 % 0.1\% 0.1% on M 3 M_3 M 3 and M 4 M_4 M 4 , and the best sample stays far below the envelope (Figure 9 ), because the extremal configurations, rank-one operators on a two-dimensional subspace aligned with the extreme eigenvectors of ρ \rho ρ , have a Haar measure that decreases rapidly with n n n . Sampling scaled projections helps: the fraction of violating samples rises, for Gardner’s inequality from 4.4 % 4.4\% 4.4% to 19 % 19\% 19% on M 2 M_2 M 2 and from below 0.1 % 0.1\% 0.1% to 0.6 % 0.6\% 0.6% on M 4 M_4 M 4 , for subadditivity from below 0.1 % 0.1\% 0.1% to 21 % 21\% 21% , 9 % 9\% 9% and 5 % 5\% 5% on M 2 M_2 M 2 , M 3 M_3 M 3 , M 4 M_4 M 4 , and for the power-mean inequality from 7 % 7\% 7% to 50 % 50\% 50% on M 2 M_2 M 2 , and the best sample comes close to the envelope (Figure 9 , Figure 20 and Figure 32 ). The exceptions are those of § 10.4 : items 4 and 7, whose extremisers are two-level operators that are not projections, and items 6 and 11, which need invertible operators.
Comparison with the constants of § 5
Several computed envelopes can now be compared with constants proved in finite von Neumann algebras. For item 10 the computed envelope equals ( G − 1 ) 2 / ( 4 G ) (G-1)^2/(4G) ( G − 1 ) 2 / ( 4 G ) (Theorem 5.27 and Remark 5.28 ) to the digits of Table 5 , and for n = 3 , 4 n=3,4 n = 3 , 4 it agrees with n = 2 n=2 n = 2 to 3 × 10 − 8 3\times 10^{-8} 3 × 1 0 − 8 , as the theorem predicts. For item 12 it agrees with the lower bound t 1 / 2 ( κ ) − 1 \mathsf t_{1/2}(\kappa)-1 t 1/2 ( κ ) − 1 of Theorem 5.40 to three decimal places (§ 5.10 ), and for item 16 it lies below and within 3 ⋅ 10 − 3 3\cdot10^{-3} 3 ⋅ 1 0 − 3 of the lower bound b 2 ( κ ) − 1 b_2(\kappa)-1 b 2 ( κ ) − 1 of Proposition 5.50 . For item 7 the search reaches 0.796 0.796 0.796 of the proved lower bound G 2 − 1 G^2-1 G 2 − 1 at δ = 0.5 \delta=0.5 δ = 0.5 ; the multistart value c 2 , 1 ≥ 2.99 c_{2,1}\ge2.99 c 2 , 1 ≥ 2.99 for κ = 10 \kappa=10 κ = 10 of Remark 5.52 is also below the proved G 2 = 3.025 G^2=3.025 G 2 = 3.025 . For item 6 the search exceeds G − 1 G-1 G − 1 for n ≥ 3 n\ge3 n ≥ 3 but stays below the upper bound κ − 1 \kappa-1 κ − 1 of Proposition 5.43 .
Other numerical observations
The numerical statements in Remarks 5.13 , 5.26 , 5.45 , 5.48 and 5.52 , after Conjecture 5.24 and after Theorem 5.40 , concern quantities not covered by the certified computation: the constants of items 27, 30, 34–36 and 38; the constant c 2 , 1 c_{2,1} c 2 , 1 of item 7 at κ = 10 \kappa=10 κ = 10 ; and the constants c P B c_{\rm PB} c PB (item 16) and b b b at the values κ = 2 , 4 , 10 , 100 \kappa=2,4,10,100 κ = 2 , 4 , 10 , 100 on M 2 M_2 M 2 and on M 3 M_3 M 3 and M 2 ⊕ M 2 M_2\oplus M_2 M 2 ⊕ M 2 , where no certified computation was made. They were obtained by multistart optimisation in double precision on M 2 M_2 M 2 , M 3 M_3 M 3 , M 4 M_4 M 4 and M 2 ⊕ M 2 M_2\oplus M_2 M 2 ⊕ M 2 ; unlike the computed envelopes of this section they are not recomputed at exactly admissible points, so they are indications, not certified lower bounds. No proof uses them.
Conclusions and open problems
A characterisation of the trace by an inequality becomes quantitative once the distance of a functional from the traces is measured by the Michelson contrast of its density on a factor, and by its noncentrality on a general finite von Neumann algebra. For Gardner’s inequality the answer is as simple as it could be: the sharp constant is the secant of the angle whose sine is the product of the contrasts of the density and of the operator, and with the operator free it is cosh L N C ( ρ ) \cosh\mathop{\mathrm{LNC}}(\rho) cosh LNC ( ρ ) . For the other characterisations the finite proved matrix envelopes vanish at the trace to the orders δ \delta δ , δ 2 \delta^2 δ 2 or δ 4 \delta^4 δ 4 (item 37 is + ∞ +\infty + ∞ for every non-central ρ \rho ρ , i.e. non-scalar on M n M_n M n ), and the weighted monotonicity (item 25) has a proved lower bound of order δ 8 \delta^8 δ 8 , with the classical reverse-inequality constants as exact values where we could determine them. Of the 38 trace characterisations, 18 (items 1–5, 8, 10, 17, 18, 21–23, 26, 28, 29, 31, 33 and 37, the last being + ∞ +\infty + ∞ ) have constants that are explicit functions of κ Z \kappa_\mathcal{Z} κ Z in every finite von Neumann algebra, five more are exact on matrix algebras (items 11, 19, 24; item 9 for n = 2 n=2 n = 2 ; item 20 for even n n n ), 13 are bounded, and items 13 and 15 are known only numerically; on direct sums only the contrast inside the blocks counts. The rules by which contrasts compose (rapidities add for products and means, angles add for Jordan products, three contrasts multiply in the commutator test) explain some of the thresholds in the contrast of the operators that the computation reveals, and prove them for subadditivity of the square root and for the power-mean inequality.
Matrix envelopes
The computation supports the following values. (The value ( G − 1 ) 2 / ( 4 G ) (G-1)^2/(4G) ( G − 1 ) 2 / ( 4 G ) of item 10, which the search had also suggested, is proved in Theorem 5.27 and Remark 5.28 .)
Conjecture 11.1. V n = G 2 − 1 V_n=G^2-1 V n = G 2 − 1 for item 7 and V n = g 25 ( δ ) V_n=g_{25}(\delta) V n = g 25 ( δ ) for item 25 in every dimension; V 2 = G − 1 V_2=G-1 V 2 = G − 1 for item 6; the value of Proposition 5.17 holds for item 9 in every dimension; and the lower bound of Proposition 5.22 is the two-contrast envelope of item 8 for even n n n .
Several questions remain open. What is the threshold of item 12 for δ < 1 \delta<1 δ < 1 , where the search places it above the value 2 2 / 3 2\sqrt2/3 2 2 /3 of Proposition 9.5 , and what are the thresholds of items 6, 10, 13 and 25, which the search shows but which no operator inequality of the kind used in § 9.1 can explain for item 10? What is the envelope of item 6 for n ≥ 3 n\ge3 n ≥ 3 , where the search exceeds G − 1 G-1 G − 1 , and does it grow with n n n ? Do items 12–16 have closed forms? The search gives the leading coefficients 1 2 \frac12 2 1 , 1 2 \frac12 2 1 , 1 36 \frac1{36} 36 1 , 0.064 0.064 0.064 and 0.09 0.09 0.09 ; for item 14 the coefficient ( 1 − 2 p ) 2 / 4 (1-\frac2p)^2/4 ( 1 − p 2 ) 2 /4 is a proved lower bound, and for items 12 and 16 the computed envelopes agree with the proved lower bounds t 1 / 2 ( κ ) − 1 \mathsf t_{1/2}(\kappa)-1 t 1/2 ( κ ) − 1 and b 2 ( κ ) − 1 b_2(\kappa)-1 b 2 ( κ ) − 1 (§ 10.6 ), which suggests that these are the values. Finally, the same questions can be asked for other characterisations, such as those of [146 ] with the arithmetic mean.
Von Neumann algebras
We first record two observations that bear on problems (7) and (8) below.
Young’s inequality for positive operators. Prove Conjecture 5.24 , in particular c 2 + = cosh L N C c_2^+=\cosh\mathop{\mathrm{LNC}} c 2 + = cosh LNC . Only the positivity of B B B matters. For B = B ∗ B=B^* B = B ∗ the constant is ( 4 G Z 2 − 3 ) 1 / 2 > G Z (4G_\mathcal{Z}^2-3)^{1/2}>G_\mathcal{Z} ( 4 G Z 2 − 3 ) 1/2 > G Z (Proposition (?) ); so if the conjecture holds, the positivity of B B B cannot be weakened to self-adjointness. By Remark 5.26 , c 2 + = G Z c_2^+=G_\mathcal{Z} c 2 + = G Z is equivalent to (6) , which for rank-one B B B is the Kantorovich inequality. For p > 2 p>2 p > 2 , find c p + c_p^+ c p + ; is it a function of κ Z \kappa_\mathcal{Z} κ Z ?
Monotonicity. Is m p ( ρ ) = m p ( κ Z ( ρ ) ) m_p(\rho)=\mathsf m_p(\kappa_\mathcal{Z}(\rho)) m p ( ρ ) = m p ( κ Z ( ρ )) , i.e. is the 2 × 2 2\times2 2 × 2 problem extremal for every p > 1 p>1 p > 1 ? Find a closed form of m p \mathsf m_p m p for p ≠ 2 p\ne2 p = 2 . Does the trace inequality (8) hold in every finite von Neumann algebra (Remark 5.30 )? If so, the bounds m p ≤ K ( κ Z 1 / p , p ) m_p\le K(\kappa_\mathcal{Z}^{1/p},p) m p ≤ K ( κ Z 1/ p , p ) and m exp ≤ S ( κ Z ) m_{\exp}\le S(\kappa_\mathcal{Z}) m e x p ≤ S ( κ Z ) of Proposition 5.29 m:e hold without the assumption that M \mathcal{M} M is AFD.
Petz–Zemánek for k ≥ 3 k\ge3 k ≥ 3 . By Propositions (?) and (?) , G Z ≤ c k ≤ κ Z 1 − 1 / k G_\mathcal{Z}\le c_k\le\kappa_\mathcal{Z}^{1-1/k} G Z ≤ c k ≤ κ Z 1 − 1/ k . By Propositions 5.11 and 5.12 , c k c_k c k is not a function of κ Z \kappa_\mathcal{Z} κ Z . Determine c k c_k c k ; which spectral data of ρ \rho ρ govern it? Numerically c k > G Z c_k>G_\mathcal{Z} c k > G Z for k ≥ 3 k\ge3 k ≥ 3 , and c k c_k c k is increasing in k k k (Remark 5.13 ); prove this. Proposition (?) gives lim inf k c k ≥ c ∞ \liminf_kc_k\ge c_\infty lim inf k c k ≥ c ∞ ; is lim k c k = c ∞ \lim_kc_k=c_\infty lim k c k = c ∞ ?
Subadditivity. Is c p T = t p ( κ Z ) c_p^T=\mathsf t_p(\kappa_\mathcal{Z}) c p T = t p ( κ Z ) ? Numerically (not used in the proofs) no configuration exceeding t p ( κ Z ) \mathsf t_p(\kappa_\mathcal{Z}) t p ( κ Z ) was found in M 2 ( C ) M_2(\mathbb{C}) M 2 ( C ) , M 3 ( C ) M_3(\mathbb{C}) M 3 ( C ) or M 2 ( C ) ⊕ M 2 ( C ) M_2(\mathbb{C})\oplus M_2(\mathbb{C}) M 2 ( C ) ⊕ M 2 ( C ) for p ∈ { 0.1 , 0.25 , 0.5 , 0.75 } p\in\{0.1,0.25,0.5,0.75\} p ∈ { 0.1 , 0.25 , 0.5 , 0.75 } . For an operator monotone f ≥ 0 f\ge0 f ≥ 0 with f ( 0 ) = 0 f(0)=0 f ( 0 ) = 0 the constant of the f f f -version is at most 1 + κ Z 2 \frac{1+\kappa_\mathcal{Z}}2 2 1 + κ Z (Remark 5.41 ). Since (R) fails, on what does it depend? For p = 1 2 p=\frac12 p = 2 1 (item 12) the certified matrix computation agrees with t 1 / 2 ( κ ) \mathsf t_{1/2}(\kappa) t 1/2 ( κ ) (§ 10.6 ).
The Alhasan–Fawwaz inequality. For A ≥ 0 A\ge0 A ≥ 0 and B = B ∗ B=B^* B = B ∗ [86 ] , the constant a ( ρ ) a(\rho) a ( ρ ) satisfies 1 < a ≤ G Z 1<a\le G_\mathcal{Z} 1 < a ≤ G Z for ρ ∉ Z \rho\notin\mathcal{Z} ρ ∈ / Z . It is not a function of κ Z \kappa_\mathcal{Z} κ Z , it is < 2 <2 < 2 on M 2 ( C ) M_2(\mathbb{C}) M 2 ( C ) , and it is < cosh O L C <\cosh\mathop{\mathrm{OLC}} < cosh OLC on every M n ( C ) M_n(\mathbb{C}) M n ( C ) (Proposition 5.44 ). Numerically it depends on the multiplicity of min σ ( ρ ) \min\sigma(\rho) min σ ( ρ ) (Remark 5.45 ). Determine a a a . Is a a a bounded on M n ( C ) M_n(\mathbb{C}) M n ( C ) independently of ρ \rho ρ ? Is a = G Z a=G_\mathcal{Z} a = G Z in II1 _1 1 factors? The same questions arise for the variant with B ≥ 0 B\ge0 B ≥ 0 .
Exponential and Araki–Lieb–Thirring inequalities. Bikchentaev’s characterisation of traces by φ ( e P Q P ) ≤ φ ( e Q ) \varphi(e^{PQP})\le\varphi(e^Q) φ ( e P QP ) ≤ φ ( e Q ) concerns projections [99 ] . For traces it holds for all Q ≥ 0 Q\ge0 Q ≥ 0 , but not for all Q = Q ∗ Q=Q^* Q = Q ∗ . Its constants lie in ( 1 , κ Z ] (1,\kappa_\mathcal{Z}] ( 1 , κ Z ] and are not functions of κ Z \kappa_\mathcal{Z} κ Z (Proposition 5.47 be:c ). The shift-invariant constant b b b satisfies b 2 ( κ Z ) ≤ b ≤ κ Z b_2(\kappa_\mathcal{Z})\le b\le\kappa_\mathcal{Z} b 2 ( κ Z ) ≤ b ≤ κ Z (Proposition 5.47 ). For the Peierls–Bogoliubov inequality of [98 ] in the form with φ ( B e A ) \varphi(Be^A) φ ( B e A ) the constant is infinite for ρ ∉ Z \rho\notin\mathcal{Z} ρ ∈ / Z . In the symmetric form, b ≤ c P B ≤ κ Z b\le c_{\rm PB}\le\kappa_\mathcal{Z} b ≤ c PB ≤ κ Z (Proposition 5.50 ). Is c P B = b = b 2 ( κ Z ) c_{\rm PB}=b=b_2(\kappa_\mathcal{Z}) c PB = b = b 2 ( κ Z ) , as the numerics in Remarks 5.48 and 5.52 and the computed envelope of item 16 (§ 10.6 ) suggest? For the Araki–Lieb–Thirring inequality of [98 ] (exponents r > 1 r>1 r > 1 and q > 0 q>0 q > 0 , all A , B ≥ 0 A,B\ge0 A , B ≥ 0 ), 1 < c r , q ≤ κ Z 1<c_{r,q}\le\kappa_\mathcal{Z} 1 < c r , q ≤ κ Z and sup q c r , q = κ Z \sup_qc_{r,q}=\kappa_\mathcal{Z} sup q c r , q = κ Z (Proposition 5.51 ); find c r , q c_{r,q} c r , q ; for ( r , q ) = ( 2 , 1 ) (r,q)=(2,1) ( r , q ) = ( 2 , 1 ) , G Z 2 ≤ c 2 , 1 ≤ κ Z G_\mathcal{Z}^2\le c_{2,1}\le\kappa_\mathcal{Z} G Z 2 ≤ c 2 , 1 ≤ κ Z (Proposition (?) ). For the spectral geometric mean of [105 ] , G Z ≤ c 6 ≤ κ Z G_\mathcal{Z}\le c_6\le\kappa_\mathcal{Z} G Z ≤ c 6 ≤ κ Z (Proposition 5.43 ); determine c 6 c_6 c 6 .
Weights and non-finite algebras. Several characterisations hold for normal semifinite weights [125 , 160 ] . For a faithful normal state, ∥ Δ φ ∥ \|\Delta_\varphi\| ∥ Δ φ ∥ extends κ Z \kappa_\mathcal{Z} κ Z (Remark 11.2 ). Do the results of § 5 and § 7 hold with ∥ Δ φ ∥ \|\Delta_\varphi\| ∥ Δ φ ∥ in place of κ Z \kappa_\mathcal{Z} κ Z whenever ∥ Δ φ ∥ < ∞ \|\Delta_\varphi\|<\infty ∥ Δ φ ∥ < ∞ ? This includes φ = τ ( ρ ⋅ ) \varphi=\tau(\rho\,\cdot) φ = τ ( ρ ⋅ ) with ρ \rho ρ or ρ − 1 \rho^{-1} ρ − 1 unbounded. For faithful normal states (and positive functionals), ∥ Δ φ ∥ < ∞ \|\Delta_\varphi\|<\infty ∥ Δ φ ∥ < ∞ forces M \mathcal{M} M to be finite. For weights it does not: φ = T r ( h ⋅ ) \varphi=\mathop{\mathrm{Tr}}(h\,\cdot) φ = Tr ( h ⋅ ) on B ( H ) B(\mathcal{H}) B ( H ) with h h h positive and invertible has ∥ Δ φ ∥ = κ ( h ) \|\Delta_\varphi\|=\kappa(h) ∥ Δ φ ∥ = κ ( h ) (Remark 11.2 ). For faithful normal semifinite weights with bounded Δ φ \Delta_\varphi Δ φ , do the results hold with ∥ Δ φ ∥ \|\Delta_\varphi\| ∥ Δ φ ∥ in place of κ Z \kappa_\mathcal{Z} κ Z ?
C ∗ C^* C ∗ -algebras. Central envelopes need not exist in a C ∗ C^* C ∗ -algebra, but κ Z A ( ρ ) \kappa^A_\mathcal{Z}(\rho) κ Z A ( ρ ) is defined, and by Remark 11.3 it is (modulo the sketch there) max ζ κ ( ρ + J ζ ) \max_\zeta\kappa(\rho+J_\zeta) max ζ κ ( ρ + J ζ ) . Do the results extend with κ Z A \kappa^A_\mathcal{Z} κ Z A ? Can the inequality sup P κ ( ρ + P ) ≤ κ Z A ( ρ ) \sup_P\kappa(\rho+P)\le\kappa^A_\mathcal{Z}(\rho) sup P κ ( ρ + P ) ≤ κ Z A ( ρ ) be strict? For inner derivations, the analogous comparison between ∥ D a ∥ \|D_a\| ∥ D a ∥ and the distance from a a a to the centre is governed by the structure of the primitive ideal space [90 , 156 ] , which suggests that it can be.
Two contrasts beyond matrices. Does Theorem A hold in finite von Neumann algebras, with δ \delta δ replaced by N C ( ρ ) \mathop{\mathrm{NC}}(\rho) NC ( ρ ) and δ ′ \delta' δ ′ by a suitable central version of the contrast of ∣ X ∣ |X| ∣ X ∣ ? The rescaling argument of Lemma 4.12 does not apply, since X ↦ z X X\mapsto zX X ↦ z X changes the contrast of ∣ X ∣ |X| ∣ X ∣ . The same question arises for the matrix-only items 9, 11, 13–15, 19, 24 and 25.
Declarations
Author contributions. All authors contributed to the conception of the work, the proofs, the computations and the writing, and approved the final manuscript.
Funding. No funding was received for this work.
Competing interests. The authors declare no competing interests.
Data availability. No data sets were generated or analysed apart from the numerical computations of § 10 , which are reproduced by the programs described below. No proof depends on a floating-point computation: the computed envelopes are certified lower bounds that support the conjectures and illustrate the theorems, and the further numerical observations of § 10.7 are not used in any proof. The proof of Proposition 5.44 af:d uses an exact computation in rational arithmetic, which can be checked by hand.
Code availability. The library [149 ] , the programs that compute all numbers, tables and figures of § 6 , § 9 and § 10 and of Appendix § 12 , and the tests are available at https://github.com/andre-a-no/von_neumann_type_i (MIT licence), in the version that accompanies the arXiv posting of this paper.
Reproducing the computations. The programs are in the directory scripts of the repository and run in this order, since later ones read the results of earlier ones: trace_characterizations.py --mode full (the envelopes with free operators, Table 4 and Table 5 ; 475 475 475 s), trace_contrast.py --mode full (both contrasts and the restricted searches, Table 6 ; 760 760 760 s), then trace_montecarlo.py (Table 7 and Appendix § 12 ), contrast_operations.py (§ 9 ), trace_envelopes_figures.py (Figure 3 ) and trace_paper_schematics.py (Figure 1 , Figure 2 and Figure 6 ). Figure 4 comes from bounds_search.py --mode full of the pipeline of [149 ] . The programs write the numbers as LaTeX macros and tables, which the source of this paper reads; the figures label the contrast of the density Δ ( W ) \Delta(W) Δ ( W ) , which is O M C ( ρ ) \mathop{\mathrm{OMC}}(\rho) OMC ( ρ ) in the notation of this paper.
Random samples and computed envelopes, item by item
Each figure shows one inequality, with one panel per dimension. In the figures the contrast of the density is labelled Δ ( W ) \Delta(W) Δ ( W ) , which is δ = O M C ( ρ ) \delta=\mathop{\mathrm{OMC}}(\rho) δ = OMC ( ρ ) , and the contrast δ ′ \delta' δ ′ of the operators is labelled δ ′ \delta' δ ′ . Top row: the normalised violation ( L − R ) / N (\mathcal{L}-\mathcal{R})/\mathcal{N} ( L − R ) / N of the random samples (grey) and, where the variables are not projections, of the scaled-projection samples (orange) against O M C ( ρ ) \mathop{\mathrm{OMC}}(\rho) OMC ( ρ ) on M 2 M_2 M 2 , M 3 M_3 M 3 , M 4 M_4 M 4 ; black: the computed envelope V ( δ ) V(\delta) V ( δ ) , a certified lower bound for the supremum, joined by straight lines. Bottom row (items with operator variables): samples with O M C ( ρ ) = 0.9 \mathop{\mathrm{OMC}}(\rho)=0.9 OMC ( ρ ) = 0.9 and all operator variables of contrast δ ′ \delta' δ ′ , against δ ′ \delta' δ ′ , on M 2 M_2 M 2 and M 3 M_3 M 3 , with the computed V ( 0.9 , δ ′ ) V(0.9,\delta') V ( 0.9 , δ ′ ) . Samples far below zero are cut off. Where an envelope is concave between grid points (item 11 below its maximum at δ ′ = δ \delta'=\delta δ ′ = δ and after its steep drop near δ ′ = 1 \delta'=1 δ ′ = 1 ), a few samples lie above the straight segments; on M 2 M_2 M 2 they lie below the closed form 1 − ( w 1 ℓ ′ + w 2 ) / ( 2 ℓ ′ ) 1-(w_1\ell'+w_2)/(2\sqrt{\ell'}) 1 − ( w 1 ℓ ′ + w 2 ) / ( 2 ℓ ′ ) , ℓ ′ = 1 − δ ′ 1 + δ ′ \ell'=\frac{1-\delta'}{1+\delta'} ℓ ′ = 1 + δ ′ 1 − δ ′ , w = 2 1 + ℓ ( 1 , ℓ ) w=\frac{2}{1+\ell}(1,\ell) w = 1 + ℓ 2 ( 1 , ℓ ) .
Figure 9. Item 1, ∣ φ ( X ) ∣ ≤ φ ( ∣ X ∣ ) |\varphi(X)|\le\varphi(|X|) ∣ φ ( X ) ∣ ≤ φ ( ∣ X ∣ ) : random samples and the computed envelope (a certified lower bound); top: against the contrast δ = O M C ( ρ ) \delta=\mathop{\mathrm{OMC}}(\rho) δ = OMC ( ρ ) of the density, bottom: against the contrast δ ′ \delta^\prime δ ′ of the operators at δ = 0.9 \delta=0.9 δ = 0.9 . In the figure Δ ( W ) = O M C ( ρ ) = δ \Delta(W)=\mathop{\mathrm{OMC}}(\rho)=\delta Δ ( W ) = OMC ( ρ ) = δ .
Figure 10. Item 2, φ ( ∣ X + Y ∣ ) ≤ φ ( ∣ X ∣ ) + φ ( ∣ Y ∣ ) \varphi(|X+Y|)\le\varphi(|X|)+\varphi(|Y|) φ ( ∣ X + Y ∣ ) ≤ φ ( ∣ X ∣ ) + φ ( ∣ Y ∣ ) : random samples and the computed envelope (a certified lower bound); top: against the contrast δ = O M C ( ρ ) \delta=\mathop{\mathrm{OMC}}(\rho) δ = OMC ( ρ ) of the density, bottom: against the contrast δ ′ \delta^\prime δ ′ of the operators at δ = 0.9 \delta=0.9 δ = 0.9 . In the figure Δ ( W ) = O M C ( ρ ) = δ \Delta(W)=\mathop{\mathrm{OMC}}(\rho)=\delta Δ ( W ) = OMC ( ρ ) = δ .
Figure 11. Item 3, φ ( A + A ∗ ) ≤ 2 φ ( ∣ A ∗ ∣ ) \varphi(A+A^*)\le2\varphi(|A^*|) φ ( A + A ∗ ) ≤ 2 φ ( ∣ A ∗ ∣ ) : random samples and the computed envelope (a certified lower bound); top: against the contrast δ = O M C ( ρ ) \delta=\mathop{\mathrm{OMC}}(\rho) δ = OMC ( ρ ) of the density, bottom: against the contrast δ ′ \delta^\prime δ ′ of the operators at δ = 0.9 \delta=0.9 δ = 0.9 . In the figure Δ ( W ) = O M C ( ρ ) = δ \Delta(W)=\mathop{\mathrm{OMC}}(\rho)=\delta Δ ( W ) = OMC ( ρ ) = δ .
Figure 12. Item 4, ∣ φ ( A 2 ) ∣ ≤ φ ( A ∗ A ) |\varphi(A^2)|\le\varphi(A^*A) ∣ φ ( A 2 ) ∣ ≤ φ ( A ∗ A ) : random samples and the computed envelope (a certified lower bound); top: against the contrast δ = O M C ( ρ ) \delta=\mathop{\mathrm{OMC}}(\rho) δ = OMC ( ρ ) of the density, bottom: against the contrast δ ′ \delta^\prime δ ′ of the operators at δ = 0.9 \delta=0.9 δ = 0.9 . In the figure Δ ( W ) = O M C ( ρ ) = δ \Delta(W)=\mathop{\mathrm{OMC}}(\rho)=\delta Δ ( W ) = OMC ( ρ ) = δ .
Figure 13. Item 5, φ ( ∣ P − R ∣ ) ≤ φ ( ∣ P − Q ∣ ) + φ ( ∣ Q − R ∣ ) \varphi(|P-R|)\le\varphi(|P-Q|)+\varphi(|Q-R|) φ ( ∣ P − R ∣ ) ≤ φ ( ∣ P − Q ∣ ) + φ ( ∣ Q − R ∣ ) : random samples and the computed envelope (a certified lower bound) against the contrast δ = O M C ( ρ ) \delta=\mathop{\mathrm{OMC}}(\rho) δ = OMC ( ρ ) of the density, labelled Δ ( W ) \Delta(W) Δ ( W ) in the figure.
Figure 14. Item 6, φ ( A ♮ B ) ≤ φ ( A ) φ ( B ) \varphi(A\natural B)\le\sqrt{\varphi(A)\varphi(B)} φ ( A ♮ B ) ≤ φ ( A ) φ ( B ) : random samples and the computed envelope (a certified lower bound); top: against the contrast δ = O M C ( ρ ) \delta=\mathop{\mathrm{OMC}}(\rho) δ = OMC ( ρ ) of the density, bottom: against the contrast δ ′ \delta^\prime δ ′ of the operators at δ = 0.9 \delta=0.9 δ = 0.9 . In the figure Δ ( W ) = O M C ( ρ ) = δ \Delta(W)=\mathop{\mathrm{OMC}}(\rho)=\delta Δ ( W ) = OMC ( ρ ) = δ .
Figure 15. Item 7, φ ( ( B 1 / 2 A B 1 / 2 ) 2 ) ≤ φ ( B A 2 B ) \varphi((B^{1/2}AB^{1/2})^2)\le\varphi(BA^2B) φ (( B 1/2 A B 1/2 ) 2 ) ≤ φ ( B A 2 B ) : random samples and the computed envelope (a certified lower bound); top: against the contrast δ = O M C ( ρ ) \delta=\mathop{\mathrm{OMC}}(\rho) δ = OMC ( ρ ) of the density, bottom: against the contrast δ ′ \delta^\prime δ ′ of the operators at δ = 0.9 \delta=0.9 δ = 0.9 . In the figure Δ ( W ) = O M C ( ρ ) = δ \Delta(W)=\mathop{\mathrm{OMC}}(\rho)=\delta Δ ( W ) = OMC ( ρ ) = δ .
Figure 16. Item 8, φ ( ∣ A B ∗ ∣ ) ≤ 1 2 φ ( ∣ A ∣ 2 ) + 1 2 φ ( ∣ B ∣ 2 ) \varphi(|AB^*|)\le\frac12\varphi(|A|^2)+\frac12\varphi(|B|^2) φ ( ∣ A B ∗ ∣ ) ≤ 2 1 φ ( ∣ A ∣ 2 ) + 2 1 φ ( ∣ B ∣ 2 ) : random samples and the computed envelope (a certified lower bound); top: against the contrast δ = O M C ( ρ ) \delta=\mathop{\mathrm{OMC}}(\rho) δ = OMC ( ρ ) of the density, bottom: against the contrast δ ′ \delta^\prime δ ′ of the operators at δ = 0.9 \delta=0.9 δ = 0.9 . In the figure Δ ( W ) = O M C ( ρ ) = δ \Delta(W)=\mathop{\mathrm{OMC}}(\rho)=\delta Δ ( W ) = OMC ( ρ ) = δ .
Figure 17. Item 9, φ ( P Q + Q P ) ≤ 2 φ ( Q P Q ) \varphi(PQ+QP)\le2\varphi(QPQ) φ ( P Q + QP ) ≤ 2 φ ( QP Q ) : random samples and the computed envelope (a certified lower bound) against the contrast δ = O M C ( ρ ) \delta=\mathop{\mathrm{OMC}}(\rho) δ = OMC ( ρ ) of the density, labelled Δ ( W ) \Delta(W) Δ ( W ) in the figure.
Figure 18. Item 10, 0 ≤ A ≤ B ⇒ φ ( A 2 ) ≤ φ ( B 2 ) 0\le A\le B\Rightarrow\varphi(A^2)\le\varphi(B^2) 0 ≤ A ≤ B ⇒ φ ( A 2 ) ≤ φ ( B 2 ) : random samples and the computed envelope (a certified lower bound); top: against the contrast δ = O M C ( ρ ) \delta=\mathop{\mathrm{OMC}}(\rho) δ = OMC ( ρ ) of the density, bottom: against the contrast δ ′ \delta^\prime δ ′ of the operators at δ = 0.9 \delta=0.9 δ = 0.9 . In the figure Δ ( W ) = O M C ( ρ ) = δ \Delta(W)=\mathop{\mathrm{OMC}}(\rho)=\delta Δ ( W ) = OMC ( ρ ) = δ .
Figure 19. Item 11, ( det A ) 1 / n ≤ 1 n φ ( A ) (\det A)^{1/n}\le\frac1n\varphi(A) ( det A ) 1/ n ≤ n 1 φ ( A ) , φ ( 1 ) = n \varphi(1)=n φ ( 1 ) = n : random samples and the computed envelope (a certified lower bound); top: against the contrast δ = O M C ( ρ ) \delta=\mathop{\mathrm{OMC}}(\rho) δ = OMC ( ρ ) of the density, bottom: against the contrast δ ′ \delta^\prime δ ′ of the operators at δ = 0.9 \delta=0.9 δ = 0.9 . In the figure Δ ( W ) = O M C ( ρ ) = δ \Delta(W)=\mathop{\mathrm{OMC}}(\rho)=\delta Δ ( W ) = OMC ( ρ ) = δ .
Figure 20. Item 12, φ ( ( A + B ) 1 / 2 ) ≤ φ ( A 1 / 2 ) + φ ( B 1 / 2 ) \varphi((A+B)^{1/2})\le\varphi(A^{1/2})+\varphi(B^{1/2}) φ (( A + B ) 1/2 ) ≤ φ ( A 1/2 ) + φ ( B 1/2 ) : random samples and the computed envelope (a certified lower bound); top: against the contrast δ = O M C ( ρ ) \delta=\mathop{\mathrm{OMC}}(\rho) δ = OMC ( ρ ) of the density, bottom: against the contrast δ ′ \delta^\prime δ ′ of the operators at δ = 0.9 \delta=0.9 δ = 0.9 . In the figure Δ ( W ) = O M C ( ρ ) = δ \Delta(W)=\mathop{\mathrm{OMC}}(\rho)=\delta Δ ( W ) = OMC ( ρ ) = δ .
Figure 21. Item 13, φ ( A + B − ∣ A − B ∣ ) ≤ 2 φ ( A 1 / 4 B 1 / 2 A 1 / 4 ) \varphi(A+B-|A-B|)\le2\varphi(A^{1/4}B^{1/2}A^{1/4}) φ ( A + B − ∣ A − B ∣ ) ≤ 2 φ ( A 1/4 B 1/2 A 1/4 ) : random samples and the computed envelope (a certified lower bound); top: against the contrast δ = O M C ( ρ ) \delta=\mathop{\mathrm{OMC}}(\rho) δ = OMC ( ρ ) of the density, bottom: against the contrast δ ′ \delta^\prime δ ′ of the operators at δ = 0.9 \delta=0.9 δ = 0.9 . In the figure Δ ( W ) = O M C ( ρ ) = δ \Delta(W)=\mathop{\mathrm{OMC}}(\rho)=\delta Δ ( W ) = OMC ( ρ ) = δ .
Figure 22. Item 14, ∣ φ ( P Q ) ∣ ≤ φ ( P ) 1 / 3 φ ( Q ) 2 / 3 |\varphi(PQ)|\le\varphi(P)^{1/3}\varphi(Q)^{2/3} ∣ φ ( P Q ) ∣ ≤ φ ( P ) 1/3 φ ( Q ) 2/3 : random samples and the computed envelope (a certified lower bound) against the contrast δ = O M C ( ρ ) \delta=\mathop{\mathrm{OMC}}(\rho) δ = OMC ( ρ ) of the density, labelled Δ ( W ) \Delta(W) Δ ( W ) in the figure.
Figure 23. Item 15, φ ( e P + Q ) ≤ φ ( e P / 2 e Q e P / 2 ) \varphi(e^{P+Q})\le\varphi(e^{P/2}e^Qe^{P/2}) φ ( e P + Q ) ≤ φ ( e P /2 e Q e P /2 ) : random samples and the computed envelope (a certified lower bound) against the contrast δ = O M C ( ρ ) \delta=\mathop{\mathrm{OMC}}(\rho) δ = OMC ( ρ ) of the density, labelled Δ ( W ) \Delta(W) Δ ( W ) in the figure.
Figure 24. Item 16, φ ( e A ) e φ ( e A / 2 B e A / 2 ) / φ ( e A ) ≤ φ ( e A + B ) \varphi(e^A)e^{\varphi(e^{A/2}Be^{A/2})/\varphi(e^A)}\le\varphi(e^{A+B}) φ ( e A ) e φ ( e A /2 B e A /2 ) / φ ( e A ) ≤ φ ( e A + B ) : random samples and the computed envelope (a certified lower bound); top: against the contrast δ = O M C ( ρ ) \delta=\mathop{\mathrm{OMC}}(\rho) δ = OMC ( ρ ) of the density, bottom: against the contrast δ ′ \delta^\prime δ ′ of the operators at δ = 0.9 \delta=0.9 δ = 0.9 . In the figure Δ ( W ) = O M C ( ρ ) = δ \Delta(W)=\mathop{\mathrm{OMC}}(\rho)=\delta Δ ( W ) = OMC ( ρ ) = δ .
Figure 25. Item 17, φ ( P Q P ) ≤ φ ( Q ) \varphi(PQP)\le\varphi(Q) φ ( P QP ) ≤ φ ( Q ) : random samples and the computed envelope (a certified lower bound) against the contrast δ = O M C ( ρ ) \delta=\mathop{\mathrm{OMC}}(\rho) δ = OMC ( ρ ) of the density, labelled Δ ( W ) \Delta(W) Δ ( W ) in the figure.
Figure 26. Item 18, φ ( ∣ P − Q ∣ ) ≤ φ ( P + Q ) \varphi(|P-Q|)\le\varphi(P+Q) φ ( ∣ P − Q ∣ ) ≤ φ ( P + Q ) : random samples and the computed envelope (a certified lower bound) against the contrast δ = O M C ( ρ ) \delta=\mathop{\mathrm{OMC}}(\rho) δ = OMC ( ρ ) of the density, labelled Δ ( W ) \Delta(W) Δ ( W ) in the figure.
Figure 27. Item 19, φ ( P + Q ) ≤ φ ( 1 + P ∧ Q ) \varphi(P+Q)\le\varphi(1+P\wedge Q) φ ( P + Q ) ≤ φ ( 1 + P ∧ Q ) : random samples and the computed envelope (a certified lower bound) against the contrast δ = O M C ( ρ ) \delta=\mathop{\mathrm{OMC}}(\rho) δ = OMC ( ρ ) of the density, labelled Δ ( W ) \Delta(W) Δ ( W ) in the figure.
Figure 28. Item 20, φ ( e Q P Q ) ≤ φ ( e P ) \varphi(e^{QPQ})\le\varphi(e^P) φ ( e QP Q ) ≤ φ ( e P ) : random samples and the computed envelope (a certified lower bound) against the contrast δ = O M C ( ρ ) \delta=\mathop{\mathrm{OMC}}(\rho) δ = OMC ( ρ ) of the density, labelled Δ ( W ) \Delta(W) Δ ( W ) in the figure.
Figure 29. Item 21, φ ( R e P Q ) ≥ 0 \varphi(\mathrm{Re}\,PQ)\ge0 φ ( Re P Q ) ≥ 0 : random samples and the computed envelope (a certified lower bound) against the contrast δ = O M C ( ρ ) \delta=\mathop{\mathrm{OMC}}(\rho) δ = OMC ( ρ ) of the density, labelled Δ ( W ) \Delta(W) Δ ( W ) in the figure.
Figure 30. Item 22, φ ( R e P Q ) ≤ φ ( ∣ P Q ∣ ) \varphi(\mathrm{Re}\,PQ)\le\varphi(|PQ|) φ ( Re P Q ) ≤ φ ( ∣ P Q ∣ ) : random samples and the computed envelope (a certified lower bound) against the contrast δ = O M C ( ρ ) \delta=\mathop{\mathrm{OMC}}(\rho) δ = OMC ( ρ ) of the density, labelled Δ ( W ) \Delta(W) Δ ( W ) in the figure.
Figure 31. Item 23, φ ( P ∨ Q ) ≤ φ ( P ) + φ ( Q ) \varphi(P\vee Q)\le\varphi(P)+\varphi(Q) φ ( P ∨ Q ) ≤ φ ( P ) + φ ( Q ) : random samples and the computed envelope (a certified lower bound) against the contrast δ = O M C ( ρ ) \delta=\mathop{\mathrm{OMC}}(\rho) δ = OMC ( ρ ) of the density, labelled Δ ( W ) \Delta(W) Δ ( W ) in the figure.
Figure 32. Item 24, φ ( A 2 ) 1 / 2 ≤ φ ( A ) \varphi(A^2)^{1/2}\le\varphi(A) φ ( A 2 ) 1/2 ≤ φ ( A ) , φ ( 1 ) = n \varphi(1)=n φ ( 1 ) = n : random samples and the computed envelope (a certified lower bound); top: against the contrast δ = O M C ( ρ ) \delta=\mathop{\mathrm{OMC}}(\rho) δ = OMC ( ρ ) of the density, bottom: against the contrast δ ′ \delta^\prime δ ′ of the operators at δ = 0.9 \delta=0.9 δ = 0.9 . In the figure Δ ( W ) = O M C ( ρ ) = δ \Delta(W)=\mathop{\mathrm{OMC}}(\rho)=\delta Δ ( W ) = OMC ( ρ ) = δ .
Figure 33. Item 25, 0 ≤ A ≤ B ⇒ φ ( A 4 ) ≤ φ ( A B 2 A ) 0\le A\le B\Rightarrow\varphi(A^4)\le\varphi(AB^2A) 0 ≤ A ≤ B ⇒ φ ( A 4 ) ≤ φ ( A B 2 A ) : random samples and the computed envelope (a certified lower bound); top: against the contrast δ = O M C ( ρ ) \delta=\mathop{\mathrm{OMC}}(\rho) δ = OMC ( ρ ) of the density, bottom: against the contrast δ ′ \delta^\prime δ ′ of the operators at δ = 0.9 \delta=0.9 δ = 0.9 . In the figure Δ ( W ) = O M C ( ρ ) = δ \Delta(W)=\mathop{\mathrm{OMC}}(\rho)=\delta Δ ( W ) = OMC ( ρ ) = δ .