Scope and organisation
This is a consolidated research preprint, rather than a new claim of priority for each ingredient. The operator Michelson contrast for positive operators originates in the work of Abed, Nikolaeva and Novikov [1 ] ; that joint work is credited as a prerequisite. The later separate three-author manuscripts on sharp weighted trace inequalities, quantitative centrality and monotonicity defects are not reproduced here.
Chapter 1 consolidates the logarithmic-contrast manuscript of Nikolaeva and Novikov [2 ] , including its quantum-information additions. Chapter 2 presents the finite-matrix continuation, using the later revised text of the Novikov–Nikolaeva manuscript. Chapters 3, 5 and 6 incorporate the associated manuscripts originally authored by Novikov alone. Chapter 4 incorporates the joint infinite-contrast manuscript. These provenance distinctions are retained explicitly; collecting the series does not change the authorship history of the component works.
The notation OMC, OLC is used throughout; the original quantum-information text used QMC, QLC for the same functionals. The chapter structure permits local hypotheses to be stated precisely: an arbitrary invertible operator, a positive element, a finite tracial algebra, and a Euclidean Jordan algebra are different settings. In particular, a statement about ∣ A ∣ |A| ∣ A ∣ alone does not transfer a commutativity assertion to A A A .
Reading guide. Chapters 1–2 establish the matrix framework. Chapter 3 extends it to finite traces, Chapter 4 to unbounded spectra, and Chapter 5 to products. Chapter 6 returns to geometry and generalises the matrix equality phenomenon by Jordan rank.
Preparation of this version. The consolidation, English-language editing, cross-reference reconciliation and selected algebraic checks were assisted by ChatGPT. The present document is a working preprint and has not been independently refereed as a unified text. A separate author-facing preparation note records the source versions and corrections made during consolidation.
Chapter 1
Operator Michelson contrast and spectral geometry
The logarithmic coordinate
For an invertible bounded operator A A A on a complex Hilbert space put
κ ( A ) = ∥ A ∥ ∥ A − 1 ∥ , OMC ( A ) = κ ( A ) − 1 κ ( A ) + 1 , \kappa(A) = \lVert A\rVert\lVert A^{-1}\rVert,\qquad\operatorname{OMC}(A) = \frac{\kappa(A)-1}{\kappa(A)+1}, κ ( A ) = ∥ A ∥ ∥ A − 1 ∥ , OMC ( A ) = κ ( A ) + 1 κ ( A ) − 1 ,
OLC ( A ) = artanh OMC ( A ) = 1 2 log κ ( A ) . \operatorname{OLC}(A) = \operatorname{artanh}\operatorname{OMC}(A) = \frac{1}{2}\log\kappa(A). OLC ( A ) = artanh OMC ( A ) = 2 1 log κ ( A ) .
For positive A A A , with m = inf σ ( A ) > 0 m=\inf\sigma(A)>0 m = inf σ ( A ) > 0 and M = sup σ ( A ) M=\sup\sigma(A) M = sup σ ( A ) , this gives OMC ( A ) = ( M − m ) / ( M + m ) \operatorname{OMC}(A)=(M-m)/(M+m) OMC ( A ) = ( M − m ) / ( M + m ) , the operator Michelson contrast of [1 ] . The identity artanh ( ( M − m ) / ( M + m ) ) = 1 2 log ( M / m ) \operatorname{artanh}((M-m)/(M+m))=\frac{1}{2}\log(M/m) artanh (( M − m ) / ( M + m )) = 2 1 log ( M / m ) is elementary; condition-number inequalities and the scalar change of coordinate are background, not new inequalities. The logarithmic formulation developed in [2 ] makes the equality and geometric questions explicit.
Polar decomposition gives κ ( A ) = κ ( ∣ A ∣ ) \kappa(A)=\kappa(\lvert A\rvert) κ ( A ) = κ (∣ A ∣) , so these functionals extend from the positive cone to all invertibles. They are invariant under multiplication by a nonzero scalar, left or right multiplication by a unitary, adjoint and inversion. They are nonnegative and vanish exactly on scalar multiples of unitaries. On nonzero noninvertible operators we use the extended conventions OMC = 1 \operatorname{OMC}=1 OMC = 1 and OLC = + ∞ \operatorname{OLC}=+\infty OLC = + ∞ ; the zero operator is excluded from this convention.
Remark 1.1.1. Passing to ∣ A ∣ \lvert A\rvert ∣ A ∣ extends a functional, not the Thompson metric, and does not preserve commutation questions. For example, the Pauli matrices σ x , σ y \sigma_x,\sigma_y σ x , σ y have absolute value I I I and satisfy σ x σ y = − σ y σ x \sigma_x\sigma_y=-\sigma_y\sigma_x σ x σ y = − σ y σ x .
Projective Thompson geometry and continuity
For positive invertible A , B A,B A , B , the Thompson metric and its quotient by positive scalar multiplication are
d T ( A , B ) = ∥ log ( A − 1 / 2 B A − 1 / 2 ) ∥ , δ T ( [ A ] , [ B ] ) = inf c > 0 d T ( A , c B ) . d_T(A,B)=\lVert\log(A^{-1/2}BA^{-1/2})\rVert,\qquad\delta_T([A],[B])=\inf_{c>0}d_T(A,cB). d T ( A , B ) = ∥ log ( A − 1/2 B A − 1/2 )∥ , δ T ([ A ] , [ B ]) = c > 0 inf d T ( A , c B ) .
This quotient is one half of Hilbert’s projective metric. The underlying order and metric results are classical; see [3 , 4 ] .
Proposition 1.2.1. For A > 0 A > 0 A > 0 invertible,
δ T ( [ I ] , [ A ] ) = OLC ( A ) . \delta_T([I],[A]) = \operatorname{OLC}(A). δ T ([ I ] , [ A ]) = OLC ( A ) . Consequently ∣ OLC ( A ) − OLC ( B ) ∣ ≤ δ T ( [ A ] , [ B ] ) |\operatorname{OLC}(A)-\operatorname{OLC}(B)| \le\delta_T([A],[B]) ∣ OLC ( A ) − OLC ( B ) ∣ ≤ δ T ([ A ] , [ B ]) .
Proof. Writing m = inf σ ( A ) m=\inf\sigma(A) m = inf σ ( A ) and M = sup σ ( A ) M=\sup\sigma(A) M = sup σ ( A ) , d T ( I , c A ) = max { ∣ log ( c m ) ∣ , ∣ log ( c M ) ∣ } d_T(I,cA)=\max\{|\log(cm)|,|\log(cM)|\} d T ( I , c A ) = max { ∣ log ( c m ) ∣ , ∣ log ( c M ) ∣ } . The minimum occurs at c = ( m M ) − 1 / 2 c=(mM)^{-1/2} c = ( m M ) − 1/2 and equals 1 2 log ( M / m ) \frac{1}{2}\log(M/m) 2 1 log ( M / m ) . The second statement is the reverse triangle inequality in the metric on rays.
Proposition 1.2.2. If invertible A j A_j A j converge in norm to a nonzero noninvertible A A A , then κ ( A j ) → ∞ \kappa(A_j)\to\infty κ ( A j ) → ∞ and OLC ( A j ) → ∞ \operatorname{OLC}(A_j)\to\infty OLC ( A j ) → ∞ .
Proof. We have ∥ A j ∥ → ∥ A ∥ > 0 \lVert A_j\rVert\to\lVert A\rVert>0 ∥ A j ∥ → ∥ A ∥ > 0 . If ∥ A j − 1 ∥ \lVert A_j^{-1}\rVert ∥ A j − 1 ∥ had a bounded subsequence, then along it ∥ A j − 1 ( A − A j ) ∥ < 1 \lVert A_j^{-1}(A-A_j)\rVert<1 ∥ A j − 1 ( A − A j )∥ < 1 eventually. The Neumann series would make A = A j ( I + A j − 1 ( A − A j ) ) A=A_j(I+A_j^{-1}(A-A_j)) A = A j ( I + A j − 1 ( A − A j )) invertible, a contradiction.
On an order interval a I ≤ A , B ≤ b I aI\le A,B\le bI a I ≤ A , B ≤ b I with a > 0 a>0 a > 0 , ordinary norm continuity is also quantitative: extremal eigenvalues are norm-Lipschitz, hence ∣ OLC ( A ) − OLC ( B ) ∣ ≤ ∥ A − B ∥ / a |\operatorname{OLC}(A)-\operatorname{OLC}(B)|\le\lVert A-B\rVert/a ∣ OLC ( A ) − OLC ( B ) ∣ ≤ ∥ A − B ∥ / a . A bound on condition numbers alone, without a lower scale bound, does not give this norm-Lipschitz conclusion.
Products, equality and dimension
Theorem 1.3.1. For invertible A , B A,B A , B ,
OLC ( A B ) ≤ OLC ( A ) + OLC ( B ) , OMC ( A B ) ≤ OMC ( A ) + OMC ( B ) 1 + OMC ( A ) OMC ( B ) . \operatorname{OLC}(AB)\le\operatorname{OLC}(A)+\operatorname{OLC}(B),\qquad
\operatorname{OMC}(AB)\le\frac{\operatorname{OMC}(A)+\operatorname{OMC}(B)}{1+\operatorname{OMC}(A)\operatorname{OMC}(B)}. OLC ( A B ) ≤ OLC ( A ) + OLC ( B ) , OMC ( A B ) ≤ 1 + OMC ( A ) OMC ( B ) OMC ( A ) + OMC ( B ) . Proof. Submultiplicativity of the norm, applied to A B AB A B and B − 1 A − 1 B^{-1}A^{-1} B − 1 A − 1 , gives κ ( A B ) ≤ κ ( A ) κ ( B ) \kappa(AB)\le\kappa(A)\kappa(B) κ ( A B ) ≤ κ ( A ) κ ( B ) . Take half logarithms, then apply the increasing function tanh \tanh tanh and its addition formula.
Lemma 1.3.2. For positive definite X , Y ∈ M n ( C ) X,Y\in M_n(\mathbb{C}) X , Y ∈ M n ( C ) , equality in Theorem 1.3.1 holds if and only if
E max ( X ) ∩ E max ( Y ) ≠ { 0 } , E min ( X ) ∩ E min ( Y ) ≠ { 0 } . E_{\max}(X)\cap E_{\max}(Y)\ne\{0\},\qquad E_{\min}(X)\cap E_{\min}(Y)\ne\{0\}. E m a x ( X ) ∩ E m a x ( Y ) = { 0 } , E m i n ( X ) ∩ E m i n ( Y ) = { 0 } . Proof. Equality of the product of the two norm bounds forces equality in both bounds. Choose a unit vector v v v attaining ∥ X Y ∥ \lVert XY\rVert ∥ X Y ∥ . Equality throughout ∥ X Y v ∥ ≤ ∥ X ∥ ∥ Y v ∥ ≤ ∥ X ∥ ∥ Y ∥ \lVert XYv\rVert\le\lVert X\rVert\lVert Yv\rVert\le\lVert X\rVert\lVert Y\rVert ∥ X Y v ∥ ≤ ∥ X ∥ ∥ Y v ∥ ≤ ∥ X ∥ ∥ Y ∥ implies that v v v is a top eigenvector of Y Y Y and that Y v Yv Y v is a top eigenvector of X X X . Positivity makes Y v Yv Y v a positive scalar multiple of v v v . Apply the same argument to Y − 1 X − 1 Y^{-1}X^{-1} Y − 1 X − 1 for the lower endpoint. Conversely, common top and bottom eigenvectors attain both bounds.
Theorem 1.3.3. If X , Y ∈ M n ( C ) X,Y\in M_n(\mathbb{C}) X , Y ∈ M n ( C ) are positive definite and OLC ( X Y ) = OLC ( X ) + OLC ( Y ) \operatorname{OLC}(XY)=\operatorname{OLC}(X)+\operatorname{OLC}(Y) OLC ( X Y ) = OLC ( X ) + OLC ( Y ) , then X Y = Y X XY=YX X Y = Y X for n ≤ 3 n\le3 n ≤ 3 . Noncommuting equality pairs exist in every dimension n ≥ 4 n\ge4 n ≥ 4 .
Proof. If either matrix is scalar there is nothing to prove. Otherwise common top and bottom eigenvectors supplied by Lemma 1.3.2 are orthogonal. Their spans reduce both self-adjoint matrices. The orthogonal complement has dimension at most one when n ≤ 3 n\le3 n ≤ 3 , so both matrices are diagonal on a common orthonormal basis. For n = 4 n=4 n = 4 take
X = ( 4 0 0 0 0 3 0 0 0 0 1 0 0 0 0 1 ) , Y = ( 4 0 0 0 0 2 1 0 0 1 2 0 0 0 0 1 ) . X=\begin{pmatrix}
4&0&0&0\\
0&3&0&0\\
0&0&1&0\\
0&0&0&1
\end{pmatrix},
\qquad
Y=\begin{pmatrix}
4&0&0&0\\
0&2&1&0\\
0&1&2&0\\
0&0&0&1
\end{pmatrix}. X = 4 0 0 0 0 3 0 0 0 0 1 0 0 0 0 1 , Y = 4 0 0 0 0 2 1 0 0 1 2 0 0 0 0 1 .
Both have maximal eigenvalue 4 4 4 and minimal eigenvalue 1 1 1 , sharing e 1 e_1 e 1 and e 4 e_4 e 4 at the two endpoints. Thus κ ( X Y ) = 16 \kappa(XY)=16 κ ( X Y ) = 16 , while the middle 2 × 2 2 \times2 2 × 2 blocks do not commute. Direct sums with identity blocks preserve the example.
Commutativity is necessary in low dimension, but is never sufficient by itself: the endpoint order must also align. For example, X = diag ( a , a − 1 ) X=\operatorname{diag}(a,a^{-1}) X = diag ( a , a − 1 ) and Y = X − 1 Y=X^{-1} Y = X − 1 have scalar product and positive individual contrasts when a > 1 a>1 a > 1 .
Tensor products and local contrasts
The spatial tensor norm and the inverse formula give
OLC ( A ⊗ B ) = OLC ( A ) + OLC ( B ) , OLC ( A ⊗ N ) = N OLC ( A ) . \operatorname{OLC}(A \otimes B)=\operatorname{OLC}(A)+\operatorname{OLC}(B), \qquad\operatorname{OLC}(A^{\otimes N})=N\operatorname{OLC}(A). OLC ( A ⊗ B ) = OLC ( A ) + OLC ( B ) , OLC ( A ⊗ N ) = N OLC ( A ) .
Proposition 1.3.4. Let X = ⨂ i X i X=\bigotimes_i X_i X = ⨂ i X i , Y = ⨂ i Y i Y=\bigotimes_i Y_i Y = ⨂ i Y i , where X i , Y i X_i,Y_i X i , Y i are positive definite matrices. Equality for X , Y X,Y X , Y holds if and only if it holds for every pair X i , Y i X_i,Y_i X i , Y i . In particular, if all factor dimensions are at most three, equality forces X Y = Y X XY=YX X Y = Y X , regardless of the total dimension.
Proof. The deficit OLC ( X ) + OLC ( Y ) − OLC ( X Y ) \operatorname{OLC}(X)+\operatorname{OLC}(Y)-\operatorname{OLC}(XY) OLC ( X ) + OLC ( Y ) − OLC ( X Y ) is the sum of the nonnegative factor deficits. It vanishes precisely when every summand vanishes. Apply Theorem 1.3.3 in each factor.
For instance, no noncommuting equality pair can have simultaneous pure tensor decompositions on C 3 ⊗ C 2 \mathbb{C}^3 \otimes\mathbb{C}^2 C 3 ⊗ C 2 . This is a restriction on the specified factorisation, not an assertion that every noncommuting equality pair is irreducible under arbitrary direct sums.
If both tensor factors are finite-dimensional and Z Z Z is invertible, positivity and the order bounds for ∣ Z ∣ |Z| ∣ Z ∣ imply
OLC ( Tr 2 ∣ Z ∣ ) ≤ OLC ( Z ) , OLC ( Tr 1 ∣ Z ∣ ) ≤ OLC ( Z ) . \operatorname{OLC}(\operatorname{Tr}_2 |Z|) \le\operatorname{OLC}(Z), \qquad\operatorname{OLC}(\operatorname{Tr}_1 |Z|) \le\operatorname{OLC}(Z). OLC ( Tr 2 ∣ Z ∣ ) ≤ OLC ( Z ) , OLC ( Tr 1 ∣ Z ∣ ) ≤ OLC ( Z ) .
Indeed, m I ≤ ∣ Z ∣ ≤ M I mI \le|Z| \le MI m I ≤ ∣ Z ∣ ≤ M I gives d 2 m I ≤ Tr 2 ∣ Z ∣ ≤ d 2 M I d_2mI \le\operatorname{Tr}_2 |Z| \le d_2MI d 2 m I ≤ Tr 2 ∣ Z ∣ ≤ d 2 M I . There is no reverse upper estimate by the local contrasts: for 0 < ε < 1 0<\varepsilon<1 0 < ε < 1 , Z = I ⊗ I + ε σ z ⊗ σ z Z=I\otimes I+\varepsilon\sigma_z\otimes\sigma_z Z = I ⊗ I + ε σ z ⊗ σ z has scalar partial traces and nonzero contrast. For nearest-neighbour invertible factors on a finite chain,
OLC ( A 12 A 23 ⋯ A N − 1 , N ) ≤ ∑ i = 1 N − 1 OLC ( A i , i + 1 ) . \operatorname{OLC}(A_{12}A_{23}\cdots A_{N-1,N}) \le\sum_{i=1}^{N-1}\operatorname{OLC}(A_{i,i+1}). OLC ( A 12 A 23 ⋯ A N − 1 , N ) ≤ i = 1 ∑ N − 1 OLC ( A i , i + 1 ) .
This is repeated subadditivity. Products occurring at intermediate steps need not be positive; the positive-matrix alignment lemma must not be applied to them without further hypotheses.
Powers, means, sums and contractions
For positive invertible A A A and z ∈ C z\in\mathbb{C} z ∈ C , functional calculus gives
∣ A z ∣ = A ℜ z , OLC ( A z ) = ∣ ℜ z ∣ OLC ( A ) . |A^z|=A^{\Re z}, \qquad\operatorname{OLC}(A^z)=|\Re z|\operatorname{OLC}(A). ∣ A z ∣ = A ℜ z , OLC ( A z ) = ∣ℜ z ∣ OLC ( A ) .
Hence, for 0 ≤ t ≤ 1 0\le t\le1 0 ≤ t ≤ 1 and positive invertible A , B A,B A , B ,
OLC ( A t B 1 − t ) ≤ t OLC ( A ) + ( 1 − t ) OLC ( B ) . \operatorname{OLC}(A^tB^{1-t}) \le t\operatorname{OLC}(A)+(1-t)\operatorname{OLC}(B). OLC ( A t B 1 − t ) ≤ t OLC ( A ) + ( 1 − t ) OLC ( B ) .
For the weighted geometric mean A # t B = A 1 / 2 ( A − 1 / 2 B A − 1 / 2 ) t A 1 / 2 A\#_tB=A^{1/2}(A^{-1/2}BA^{-1/2})^tA^{1/2} A # t B = A 1/2 ( A − 1/2 B A − 1/2 ) t A 1/2 the corresponding estimate is
OLC ( A # t B ) ≤ ( 1 − t ) OLC ( A ) + t OLC ( B ) . \operatorname{OLC}(A\#_tB) \le(1-t)\operatorname{OLC}(A)+t\operatorname{OLC}(B). OLC ( A # t B ) ≤ ( 1 − t ) OLC ( A ) + t OLC ( B ) .
Indeed, monotonicity of the weighted geometric mean yields m A 1 − t m B t I ≤ A # t B ≤ M A 1 − t M B t I m_A^{1-t}m_B^t I \le A\#_t B \le M_A^{1-t}M_B^t I m A 1 − t m B t I ≤ A # t B ≤ M A 1 − t M B t I , with the weights used in the definition of # t \#_t # t . More generally, this order argument applies to any positive mean with the stated endpoint bounds.
If a map on positive invertibles fixes the identity ray and is nonexpansive for δ T \delta_T δ T , Proposition 1.2.1 gives OLC ( F ( A ) ) ≤ OLC ( A ) \operatorname{OLC}(F(A)) \le\operatorname{OLC}(A) OLC ( F ( A )) ≤ OLC ( A ) . A separate sufficient condition is spectral range compression: m F ( A ) ≥ m A m_{F(A)} \ge m_A m F ( A ) ≥ m A and M F ( A ) ≤ M A M_{F(A)} \le M_A M F ( A ) ≤ M A . Unital positive linear maps have this property. Reversing both endpoint inequalities gives contrast expansion.
Proposition 1.4.1. For positive invertible A 1 , … , A N A_1,\ldots,A_N A 1 , … , A N and positive scalars w i w_i w i ,
OLC ( ∑ i w i A i ) ≤ max i OLC ( A i ) . \operatorname{OLC}\left(\sum_i w_i A_i\right) \le\max_i \operatorname{OLC}(A_i). OLC ( i ∑ w i A i ) ≤ i max OLC ( A i ) . Proof. Put K = max i κ ( A i ) K=\max_i \kappa(A_i) K = max i κ ( A i ) . The sum lies between ( ∑ i w i m i ) I (\sum_i w_i m_i)I ( ∑ i w i m i ) I and ( ∑ i w i M i ) I (\sum_i w_i M_i)I ( ∑ i w i M i ) I , and ∑ i w i M i ≤ K ∑ i w i m i \sum_i w_i M_i \le K\sum_i w_i m_i ∑ i w i M i ≤ K ∑ i w i m i .
Thus Cesàro means have contrast bounded by the supremum of the individual contrasts. If such means converge in norm to an invertible limit, continuity passes the bound to the limit. Similarly,
OLC ( A 1 ⋯ A N ) ≤ ∑ i = 1 N OLC ( A i ) ; \operatorname{OLC}(A_1\cdots A_N) \le\sum_{i=1}^{N}\operatorname{OLC}(A_i); OLC ( A 1 ⋯ A N ) ≤ i = 1 ∑ N OLC ( A i ) ;
if the products converge in norm to an invertible P P P and the series on the right converges, then OLC ( P ) ≤ ∑ i ≥ 1 OLC ( A i ) \operatorname{OLC}(P) \le\sum_{i\ge1}\operatorname{OLC}(A_i) OLC ( P ) ≤ ∑ i ≥ 1 OLC ( A i ) .
For A > 0 A>0 A > 0 one also has
OLC ( A ) ≥ 1 2 log ( 1 + ∥ I − A / ∥ A ∥ ∥ ) , \operatorname{OLC}(A) \ge\frac{1}{2}\log\left(1+\left\lVert I-A/\lVert A\rVert\right\rVert\right), OLC ( A ) ≥ 2 1 log ( 1 + ∥ I − A / ∥ A ∥ ∥ ) ,
with equality exactly for scalar A A A . To see this, set r = m / M ∈ ( 0 , 1 ] r=m/M\in(0,1] r = m / M ∈ ( 0 , 1 ] ; the inequality is r − 1 ≥ 2 − r r^{-1}\ge2-r r − 1 ≥ 2 − r , equivalent to ( 1 − r ) 2 ≥ 0 (1-r)^2\ge0 ( 1 − r ) 2 ≥ 0 . For X > 0 X>0 X > 0 and invertible Y Y Y ,
OLC ( X 1 / 2 Y X 1 / 2 ) ≤ OLC ( X ) + OLC ( Y ) . \operatorname{OLC}(X^{1/2}YX^{1/2}) \le\operatorname{OLC}(X)+\operatorname{OLC}(Y). OLC ( X 1/2 Y X 1/2 ) ≤ OLC ( X ) + OLC ( Y ) .
If Y > 0 Y>0 Y > 0 , the equality case is the same endpoint alignment as in Lemma 1.3.2 ; see the Jordan-algebra proof in Chapter 6.
Quantum states: qubits and higher dimensions
For a positive definite density matrix ρ \rho ρ on C d \mathbb{C}^d C d , OLC ( ρ ) = 1 2 log ( λ max / λ min ) \operatorname{OLC}(\rho)=\frac{1}{2}\log(\lambda_{\max}/\lambda_{\min}) OLC ( ρ ) = 2 1 log ( λ m a x / λ m i n ) and OLC ( ρ ) ≤ − 1 2 log λ min \operatorname{OLC}(\rho)\le-\frac{1}{2}\log\lambda_{\min} OLC ( ρ ) ≤ − 2 1 log λ m i n . For d ≥ 2 d\ge2 d ≥ 2 equality in this last bound is unattainable at a full-rank state, but the bound is asymptotically attained along spectra approaching a pure state. The maximally mixed state has zero contrast; any nonzero rank-deficient state has infinite logarithmic contrast.
Proposition 1.5.1. For a qubit ρ = 1 2 ( I + r ⋅ σ ) \rho=\frac{1}{2}(I+\mathbf{r}\cdot\boldsymbol{\sigma}) ρ = 2 1 ( I + r ⋅ σ ) with r = ∣ r ∣ < 1 r=\lvert\mathbf{r}\rvert<1 r = ∣ r ∣ < 1 ,
OMC ( ρ ) = r , OLC ( ρ ) = artanh r , \operatorname{OMC}(\rho)=r,\qquad\operatorname{OLC}(\rho)=\operatorname{artanh}r, OMC ( ρ ) = r , OLC ( ρ ) = artanh r , and, with natural logarithms,
S ( ρ ) = H ( r ) : = log 2 − 1 2 ( ( 1 + r ) log ( 1 + r ) + ( 1 − r ) log ( 1 − r ) ) . S(\rho)=H(r):=\log2-\frac{1}{2}\bigl((1+r)\log(1+r)+(1-r)\log(1-r)\bigr). S ( ρ ) = H ( r ) := log 2 − 2 1 ( ( 1 + r ) log ( 1 + r ) + ( 1 − r ) log ( 1 − r ) ) . This is a strictly decreasing bijection from r ∈ [ 0 , 1 ) r\in[0,1) r ∈ [ 0 , 1 ) to S ∈ ( 0 , log 2 ] S\in(0,\log2] S ∈ ( 0 , log 2 ] . Furthermore
1 2 ∥ ρ − I / 2 ∥ 1 = r 2 , ∥ ρ − I / 2 ∥ 2 = r 2 , Tr ρ 2 = 1 + r 2 2 , \frac{1}{2}\lVert\rho-I/2\rVert_1=\frac{r}{2},\qquad
\lVert\rho-I/2\rVert_2=\frac{r}{\sqrt{2}},\qquad
\operatorname{Tr}\rho^2=\frac{1+r^2}{2}, 2 1 ∥ ρ − I /2 ∥ 1 = 2 r , ∥ ρ − I /2 ∥ 2 = 2 r , Tr ρ 2 = 2 1 + r 2 , where these two Schatten norms are unnormalised.
Proof. The Pauli identity gives ( r ⋅ σ ) 2 = r 2 I (\mathbf{r}\cdot\boldsymbol{\sigma})^2=r^2I ( r ⋅ σ ) 2 = r 2 I , hence eigenvalues ( 1 ± r ) / 2 (1\pm r)/2 ( 1 ± r ) /2 for ρ \rho ρ . All formulas follow by substitution. In particular H ′ ( r ) = 1 2 log ( ( 1 − r ) / ( 1 + r ) ) < 0 H'(r)=\frac{1}{2}\log((1-r)/(1+r))<0 H ′ ( r ) = 2 1 log (( 1 − r ) / ( 1 + r )) < 0 for r > 0 r>0 r > 0 .
In terms of η = OLC ( ρ ) \eta=\operatorname{OLC}(\rho) η = OLC ( ρ ) , the equivalent formula S ( η ) = log ( 2 cosh η ) − η tanh η S(\eta)=\log(2\cosh\eta)-\eta\tanh\eta S ( η ) = log ( 2 cosh η ) − η tanh η gives
S ( η ) = log 2 − 1 2 η 2 + O ( η 4 ) ( η → 0 ) , S ( η ) = ( 1 + 2 η ) e − 2 η ( 1 + O ( e − 2 η ) ) ( η → ∞ ) . S(\eta)=\log2-\frac{1}{2}\eta^2+O(\eta^4)\quad(\eta\to0),\qquad S(\eta)=(1+2\eta)e^{-2\eta}(1+O(e^{-2\eta}))\quad(\eta\to\infty). S ( η ) = log 2 − 2 1 η 2 + O ( η 4 ) ( η → 0 ) , S ( η ) = ( 1 + 2 η ) e − 2 η ( 1 + O ( e − 2 η )) ( η → ∞ ) .
These follow from Taylor expansion at zero and the exact representation S ( η ) = log ( 1 + e − 2 η ) + 2 η / ( 1 + e 2 η ) S(\eta)=\log(1+e^{-2\eta})+2\eta/(1+e^{2\eta}) S ( η ) = log ( 1 + e − 2 η ) + 2 η / ( 1 + e 2 η ) . There is no exact symmetry under η ↦ 1 / η \eta\mapsto1/\eta η ↦ 1/ η .
Entropy at fixed qutrit condition number
Fix κ > 1 \kappa>1 κ > 1 . Every ordered qutrit spectrum with this condition number is
( λ 1 , λ 2 , λ 3 ) = ( t , 1 − ( 1 + 1 / κ ) t , t / κ ) , κ 2 κ + 1 ≤ t ≤ κ κ + 2 . (\lambda_1,\lambda_2,\lambda_3)=\left(t,1-(1+1/\kappa)t,t/\kappa\right),\qquad\frac{\kappa}{2\kappa+1}\leq t\leq\frac{\kappa}{\kappa+2}. ( λ 1 , λ 2 , λ 3 ) = ( t , 1 − ( 1 + 1/ κ ) t , t / κ ) , 2 κ + 1 κ ≤ t ≤ κ + 2 κ .
Theorem 1.5.2. The entropy range is a nondegenerate interval [ S min ( κ ) , S max ( κ ) ] [S_{\min}(\kappa),S_{\max}(\kappa)] [ S m i n ( κ ) , S m a x ( κ )] . Its minimum and unique maximising spectrum are
S min = log ( κ + 2 ) − κ κ + 2 log κ , ( λ 1 ∗ , λ 2 ∗ , λ 3 ∗ ) = ( 1 , q , κ − 1 ) 1 + q + κ − 1 , q = κ − 1 / ( κ + 1 ) . S_{\min}=\log(\kappa+2)-\frac{\kappa}{\kappa+2}\log\kappa,\qquad(\lambda_1^*,\lambda_2^*,\lambda_3^*)=\frac{(1,q,\kappa^{-1})}{1+q+\kappa^{-1}},\qquad q=\kappa^{-1/(\kappa+1)}. S m i n = log ( κ + 2 ) − κ + 2 κ log κ , ( λ 1 ∗ , λ 2 ∗ , λ 3 ∗ ) = 1 + q + κ − 1 ( 1 , q , κ − 1 ) , q = κ − 1/ ( κ + 1 ) . For every finite κ > 1 \kappa>1 κ > 1 , log 2 < S max ( κ ) < log 3 \log2<S_{\max}(\kappa)<\log3 log 2 < S m a x ( κ ) < log 3 ; as κ → ∞ \kappa\to\infty κ → ∞ , S min → 0 S_{\min}\to0 S m i n → 0 and S max → log 2 S_{\max}\to\log2 S m a x → log 2 .
Proof. Entropy is strictly concave along the nonconstant affine family. Its derivative vanishes exactly when ( κ + 1 ) log λ 2 = κ log λ 1 + log λ 3 (\kappa+1)\log\lambda_2=\kappa\log\lambda_1+\log\lambda_3 ( κ + 1 ) log λ 2 = κ log λ 1 + log λ 3 , giving q q q above; κ − 1 < q < 1 \kappa^{-1}<q<1 κ − 1 < q < 1 places the maximum in the interior. The two endpoint entropies are S P = log ( 2 κ + 1 ) − 2 κ 2 κ + 1 log κ S_P=\log(2\kappa+1)-\frac{2\kappa}{2\kappa+1}\log\kappa S P = log ( 2 κ + 1 ) − 2 κ + 1 2 κ log κ and S Q = S min S_Q=S_{\min} S Q = S m i n . Their difference vanishes at 1 and has derivative
( S P − S Q ) ′ = 6 ( κ 2 − 1 ) log κ ( 2 κ + 1 ) 2 ( κ + 2 ) 2 > 0. (S_P-S_Q)'=\frac{6(\kappa^2-1)\log\kappa}{(2\kappa+1)^2(\kappa+2)^2}>0. ( S P − S Q ) ′ = ( 2 κ + 1 ) 2 ( κ + 2 ) 2 6 ( κ 2 − 1 ) log κ > 0.
Thus S Q < S P S_Q<S_P S Q < S P . Finally S P ′ ( κ ) = − 2 log κ / ( 2 κ + 1 ) 2 < 0 S_P'(\kappa)=-2\log\kappa/(2\kappa+1)^2<0 S P ′ ( κ ) = − 2 log κ / ( 2 κ + 1 ) 2 < 0 and S P → log 2 S_P\to\log2 S P → log 2 , so S max > S P > log 2 S_{\max}>S_P>\log2 S m a x > S P > log 2 ; the strict upper bound log 3 \log3 log 3 follows because no admissible spectrum is uniform. The limits follow from the displayed spectra.
For every d > 3 d>3 d > 3 , entropy likewise cannot be determined by contrast. One direct construction fixes unnormalised extreme eigenvalues κ \kappa κ and 1 1 1 , puts the other d − 2 d-2 d − 2 eigenvalues equal to s ∈ [ 1 , κ ] s\in[1,\kappa] s ∈ [ 1 , κ ] , and normalises by Z = κ + 1 + ( d − 2 ) s Z=\kappa+1+(d-2)s Z = κ + 1 + ( d − 2 ) s . Its entropy has derivative
d − 2 Z 2 ( κ log κ − ( κ + 1 ) log s ) , \frac{d-2}{Z^2}\left(\kappa\log\kappa-(\kappa+1)\log s\right), Z 2 d − 2 ( κ log κ − ( κ + 1 ) log s ) ,
which is not identically zero, while its condition number is always κ \kappa κ .
Purity at fixed qutrit condition number
Proposition 1.5.3. On the same family,
P min ( κ ) = κ 2 + 1 2 ( κ 2 + κ + 1 ) , P max ( κ ) = κ 2 + 2 ( κ + 2 ) 2 . P_{\min}(\kappa)=\frac{\kappa^2+1}{2(\kappa^2+\kappa+1)},\qquad P_{\max}(\kappa)=\frac{\kappa^2+2}{(\kappa+2)^2}. P m i n ( κ ) = 2 ( κ 2 + κ + 1 ) κ 2 + 1 , P m a x ( κ ) = ( κ + 2 ) 2 κ 2 + 2 . The minimum is attained at t ∗ = κ ( κ + 1 ) / [ 2 ( κ 2 + κ + 1 ) ] t_*=\kappa(\kappa+1)/[2(\kappa^2+\kappa+1)] t ∗ = κ ( κ + 1 ) / [ 2 ( κ 2 + κ + 1 )] , and the maximum at the endpoint with λ 2 = λ 3 \lambda_2=\lambda_3 λ 2 = λ 3 .
Proof. Put a = 1 + 1 / κ a = 1 + 1/\kappa a = 1 + 1/ κ and b = 1 + a 2 + 1 / κ 2 b = 1 + a^2 + 1/\kappa^2 b = 1 + a 2 + 1/ κ 2 . Then P ( t ) = b t 2 − 2 a t + 1 P(t) = bt^2 - 2at + 1 P ( t ) = b t 2 − 2 a t + 1 is strictly convex, with vertex a / b a/b a / b strictly between the endpoints for κ > 1 \kappa> 1 κ > 1 . Substituting gives P min = 1 − a 2 / b P_{\min} = 1 - a^2/b P m i n = 1 − a 2 / b .
The difference between the two endpoint purities, with the larger-eigenvalue multiplicity one first, is
κ 2 + 2 ( κ + 2 ) 2 − 2 κ 2 + 1 ( 2 κ + 1 ) 2 = 2 ( κ − 1 ) 3 ( κ + 1 ) ( κ + 2 ) 2 ( 2 κ + 1 ) 2 > 0. \frac{\kappa^2 + 2}{(\kappa+ 2)^2} - \frac{2\kappa^2 + 1}{(2\kappa+ 1)^2} = \frac{2(\kappa- 1)^3(\kappa+ 1)}{(\kappa+ 2)^2(2\kappa+ 1)^2} > 0. ( κ + 2 ) 2 κ 2 + 2 − ( 2 κ + 1 ) 2 2 κ 2 + 1 = ( κ + 2 ) 2 ( 2 κ + 1 ) 2 2 ( κ − 1 ) 3 ( κ + 1 ) > 0.
□ \square □
For κ > 1 \kappa> 1 κ > 1 , 1 / 3 < P min < 1 / 2 1/3 < P_{\min} < 1/2 1/3 < P m i n < 1/2 and 1 / 3 < P max < 1 1/3 < P_{\max} < 1 1/3 < P m a x < 1 . Setting κ = e 2 η \kappa= e^{2\eta} κ = e 2 η gives
1 − P max = 4 e − 2 η ( 1 + O ( e − 2 η ) ) , 1 2 − P min = 1 2 e − 2 η ( 1 + O ( e − 2 η ) ) . 1 - P_{\max} = 4e^{-2\eta}(1 + O(e^{-2\eta})), \qquad\frac{1}{2} - P_{\min} = \frac{1}{2}e^{-2\eta}(1 + O(e^{-2\eta})). 1 − P m a x = 4 e − 2 η ( 1 + O ( e − 2 η )) , 2 1 − P m i n = 2 1 e − 2 η ( 1 + O ( e − 2 η )) .
The second limit approaches a rank-two mixed state, not a pure state.
Separability and ensemble mixing
The matrices in Theorem 1.3.3 , normalised to trace one, give a noncommuting equality pair without requiring entanglement. Indeed, X / 9 X/9 X /9 is diagonal in the computational two-qubit basis and the partial transpose of Y / 9 Y/9 Y /9 is
1 9 ( 4 0 0 1 0 2 0 0 0 0 2 0 1 0 0 1 ) , \frac{1}{9}
\begin{pmatrix}
4 & 0 & 0 & 1 \\
0 & 2 & 0 & 0 \\
0 & 0 & 2 & 0 \\
1 & 0 & 0 & 1
\end{pmatrix}, 9 1 4 0 0 1 0 2 0 0 0 0 2 0 1 0 0 1 ,
with eigenvalues ( 5 ± 13 ) / 18 (5 \pm\sqrt{13})/18 ( 5 ± 13 ) /18 , 2 / 9 2/9 2/9 , 2 / 9 2/9 2/9 , all positive. The two-qubit PPT criterion therefore implies separability [5 ] . The tensor restriction of Proposition 1.3.4 concerns simultaneous pure tensor decompositions and is compatible with this separable example.
The classical link between Michelson contrast and Jensen–Shannon divergence was studied in [6 ] . The qubit formulas below use the same binary-entropy function after passing to eigenvalues.
For a binary ensemble of qubits with Bloch-vector lengths r 1 r_1 r 1 , r 2 r_2 r 2 and angle θ \theta θ , put
r ˉ = 1 2 r 1 2 + r 2 2 + 2 r 1 r 2 cos θ . \bar{r} = \frac{1}{2}\sqrt{r_1^2 + r_2^2 + 2r_1r_2\cos\theta}. r ˉ = 2 1 r 1 2 + r 2 2 + 2 r 1 r 2 cos θ .
Then the quantum Jensen–Shannon divergence is exactly
QJSD ( ρ 1 , ρ 2 ) = H ( r ˉ ) − 1 2 H ( r 1 ) − 1 2 H ( r 2 ) . \operatorname{QJSD}(\rho_1, \rho_2) = H(\bar{r}) - \frac{1}{2}H(r_1) - \frac{1}{2}H(r_2). QJSD ( ρ 1 , ρ 2 ) = H ( r ˉ ) − 2 1 H ( r 1 ) − 2 1 H ( r 2 ) .
This follows because the Bloch vector of the average state is the average of the Bloch vectors. Parallel vectors reduce the formula to a classical Bernoulli divergence; antiparallel unit vectors give log 2 \log2 log 2 . For the antipodal pair ρ ± = ( I ± r σ z ) / 2 \rho_{\pm} = (I \pm r\sigma_z)/2 ρ ± = ( I ± r σ z ) /2 , the formula becomes QJSD ( ρ + , ρ − ) = log 2 − H ( r ) \operatorname{QJSD}(\rho_+, \rho_-) = \log2 - H(r) QJSD ( ρ + , ρ − ) = log 2 − H ( r ) , a strictly increasing function of their common Michelson contrast r r r on [ 0 , 1 ] [0, 1] [ 0 , 1 ] . This is the binary entropy–contrast correspondence. For a general pair, the angle remains an additional variable even for qubits. In any dimension Proposition 1.4.1 gives OLC ( ∑ i p i ρ i ) ≤ max i OLC ( ρ i ) \operatorname{OLC}(\sum_i p_i\rho_i) \le\max_i \operatorname{OLC}(\rho_i) OLC ( ∑ i p i ρ i ) ≤ max i OLC ( ρ i ) . The elementary inequality OLC ( ρ ) ≥ 1 2 ( S ( ρ ) − log d ) \operatorname{OLC}(\rho) \ge\frac{1}{2}(S(\rho) - \log d) OLC ( ρ ) ≥ 2 1 ( S ( ρ ) − log d ) is valid but vacuous as a lower bound, since S ≤ log d S \le\log d S ≤ log d ; combining it with the mixing inequality gives no improvement over the standard binary information bound QJSD ≤ log 2 \operatorname{QJSD} \le\log2 QJSD ≤ log 2 . Useful entropy-deficit bounds in terms of C p C_p C p appear in Chapter 2.
Log-spectral measures and the scope of analogies
For a positive n × n n \times n n × n matrix define ν A = n − 1 ∑ j δ log λ j ( A ) \nu_A = n^{-1}\sum_j \delta_{\log\lambda_j(A)} ν A = n − 1 ∑ j δ l o g λ j ( A ) . Then
OLC ( A ) = 1 2 diam supp ν A , ν ^ A ( t ) = n − 1 Tr A i t , ν A ⊗ B = ν A ∗ ν B . \operatorname{OLC}(A) = \frac{1}{2}\operatorname{diam}\operatorname{supp}\nu_A,\qquad\widehat{\nu}_A(t) = n^{-1}\operatorname{Tr} A^{it},\qquad\nu_{A\otimes B} = \nu_A * \nu_B. OLC ( A ) = 2 1 diam supp ν A , ν A ( t ) = n − 1 Tr A i t , ν A ⊗ B = ν A ∗ ν B .
The Fourier transform of this finite atomic measure is a finite exponential sum; it need not decay at infinity. For two eigenvalues, after subtracting their logarithmic midpoint, it is cos ( t OLC ( A ) ) \cos(t\operatorname{OLC}(A)) cos ( t OLC ( A )) . Before that centring it has an additional scalar phase. These are exact spectral identities. They contain no information about relative eigenvectors and therefore cannot by themselves detect whether two operators commute. Chapter 2 develops the corresponding Laplace profile and its homometric ambiguity.
The scalar relation OMC = tanh OLC \operatorname{OMC} = \tanh\operatorname{OLC} OMC = tanh OLC resembles the velocity–rapidity relation, but does not identify arbitrary operator geometry with spacetime geometry. A precise Lorentz-cone statement, with its product law and rank hypotheses, is given in Chapter 6. Analogies with neural-network activations, colour-channel counts, representation dimensions, or spin-chain gap criteria are not used as proofs or as results of this preprint.
Contrasts beyond the operator norm
Symmetric gauges and normalised contrast
Throughout this chapter n ≥ 2 n \ge2 n ≥ 2 , unless a one-dimensional case is explicitly mentioned. Let Φ \Phi Φ be a symmetric gauge function on R n \mathbb{R}^{n} R n , and write ∥ A ∥ Φ = Φ ( s ( A ) ) \lVert A\rVert_{\Phi} = \Phi(s(A)) ∥ A ∥ Φ = Φ ( s ( A )) for the associated unitarily invariant norm. Here s 1 ( A ) ≥ ⋯ ≥ s n ( A ) > 0 s_1(A) \ge\cdots\ge s_n(A) > 0 s 1 ( A ) ≥ ⋯ ≥ s n ( A ) > 0 are the singular values of an invertible matrix, 1 = ( 1 , … , 1 ) \mathbf{1} = (1,\ldots,1) 1 = ( 1 , … , 1 ) , and e j e_j e j is a coordinate vector. Put
C Φ ( A ) = 1 2 log Φ ( s ( A ) ) Φ ( s ( A ) − 1 ) Φ ( 1 ) 2 , d Φ = log Φ ( 1 ) Φ ( e 1 ) . C_{\Phi}(A) = \frac{1}{2}\log\frac{\Phi(s(A))\Phi(s(A)^{-1})}{\Phi(\mathbf{1})^2}, \qquad d_{\Phi} = \log\frac{\Phi(\mathbf{1})}{\Phi(e_1)}. C Φ ( A ) = 2 1 log Φ ( 1 ) 2 Φ ( s ( A )) Φ ( s ( A ) − 1 ) , d Φ = log Φ ( e 1 ) Φ ( 1 ) .
The reciprocal vector is understood up to permutation. We use standard symmetric-gauge and majorisation facts [7 ] :
∣ x ∣ ≤ ∣ y ∣ ⟹ Φ ( x ) ≤ Φ ( y ) , |x| \le|y| \Longrightarrow\Phi(x) \le\Phi(y), ∣ x ∣ ≤ ∣ y ∣ ⟹ Φ ( x ) ≤ Φ ( y ) ,
Φ ( e 1 ) ∥ x ∥ ∞ ≤ Φ ( x ) ≤ Φ ( 1 ) ∥ x ∥ ∞ , \Phi(e_1)\lVert x\rVert_{\infty} \le\Phi(x) \le\Phi(\mathbf{1})\lVert x\rVert_{\infty}, Φ ( e 1 ) ∥ x ∥ ∞ ≤ Φ ( x ) ≤ Φ ( 1 ) ∥ x ∥ ∞ ,
Φ ( 1 ) n ∥ x ∥ 1 ≤ Φ ( x ) ≤ Φ ( e 1 ) ∥ x ∥ 1 , \frac{\Phi(\mathbf{1})}{n}\lVert x\rVert_1 \le\Phi(x) \le\Phi(e_1)\lVert x\rVert_1, n Φ ( 1 ) ∥ x ∥ 1 ≤ Φ ( x ) ≤ Φ ( e 1 ) ∥ x ∥ 1 ,
Φ ( x t y 1 − t ) ≤ Φ ( x ) t Φ ( y ) 1 − t ( x , y ≥ 0 , 0 ≤ t ≤ 1 ) . \Phi(x^t y^{1-t}) \le\Phi(x)^t\Phi(y)^{1-t} \qquad(x,y \ge0,\ 0 \le t \le1). Φ ( x t y 1 − t ) ≤ Φ ( x ) t Φ ( y ) 1 − t ( x , y ≥ 0 , 0 ≤ t ≤ 1 ) .
The last inequality follows by homogeneity from the coordinatewise weighted arithmetic–geometric mean inequality. The lower ℓ 1 \ell_1 ℓ 1 bound follows by averaging all coordinate permutations. A symmetric convex function is monotone under majorisation because a majorised vector is a convex combination of permutations.
Proposition 2.1.1. The contrast is unchanged by scaling Φ \Phi Φ , by A ↦ λ U A V A \mapsto\lambda UAV A ↦ λ U A V with U , V U,V U , V unitary and λ ≠ 0 \lambda\ne0 λ = 0 , and by adjoint, inversion or absolute value. Moreover
0 ≤ C 1 ( A ) ≤ C Φ ( A ) ≤ OLC ( A ) , C Φ ( A ) ≥ OLC ( A ) − d Φ . 0 \le C_1(A) \le C_{\Phi}(A) \le\operatorname{OLC}(A), \qquad C_{\Phi}(A) \ge\operatorname{OLC}(A) - d_{\Phi}. 0 ≤ C 1 ( A ) ≤ C Φ ( A ) ≤ OLC ( A ) , C Φ ( A ) ≥ OLC ( A ) − d Φ . It vanishes exactly on scalar multiples of unitaries.
Proof. The invariances follow from the singular values. The gauge bounds give
Φ ( s ) Φ ( s − 1 ) Φ ( 1 ) 2 ≥ AM ( s ) HM ( s ) ≥ 1 , \frac{\Phi(s)\Phi(s^{-1})}{\Phi(\mathbf{1})^2} \ge\frac{\operatorname{AM}(s)}{\operatorname{HM}(s)} \ge1, Φ ( 1 ) 2 Φ ( s ) Φ ( s − 1 ) ≥ HM ( s ) AM ( s ) ≥ 1 , with equality in the latter inequality only for constant s s s . The upper and lower ℓ ∞ \ell_{\infty} ℓ ∞ estimates give the two comparisons with OLC. □ \square □
The sharp defect and rigidity
Theorem 2.2.1. For n ≥ 2 n \ge2 n ≥ 2 , invertible A A A , B B B , positive definite P P P , and s ≥ 1 s \ge1 s ≥ 1 ,
C Φ ( A B ) ≤ C Φ ( A ) + C Φ ( B ) + d Φ , C_{\Phi}(AB) \le C_{\Phi}(A) + C_{\Phi}(B) + d_{\Phi}, C Φ ( A B ) ≤ C Φ ( A ) + C Φ ( B ) + d Φ , s C Φ ( P ) ≤ C Φ ( P s ) ≤ s C Φ ( P ) + ( s − 1 ) d Φ . sC_{\Phi}(P) \le C_{\Phi}(P^{s}) \le sC_{\Phi}(P) + (s-1)d_{\Phi}. s C Φ ( P ) ≤ C Φ ( P s ) ≤ s C Φ ( P ) + ( s − 1 ) d Φ . Both constants are optimal: for P t = diag ( t , 1 , … , 1 , t − 1 ) P_t = \operatorname{diag}(t,1,\ldots,1,t^{-1}) P t = diag ( t , 1 , … , 1 , t − 1 ) ,
lim t → ∞ ( C Φ ( P t s ) − s C Φ ( P t ) ) = ( s − 1 ) d Φ . \lim_{t\to\infty}\left(C_{\Phi}(P_t^{s}) - sC_{\Phi}(P_t)\right) = (s-1)d_{\Phi}. t → ∞ lim ( C Φ ( P t s ) − s C Φ ( P t ) ) = ( s − 1 ) d Φ .
Proof. Normalise Φ ( 1 ) = 1 \Phi(1)=1 Φ ( 1 ) = 1 and set a = Φ ( e 1 ) a=\Phi(e_1) a = Φ ( e 1 ) . The ideal norm inequality gives ∥ A B ∥ Φ ≤ ∥ A ∥ ∥ B ∥ Φ ≤ a − 1 ∥ A ∥ Φ ∥ B ∥ Φ \lVert AB\rVert_{\Phi} \le\lVert A\rVert\lVert B\rVert_{\Phi} \le a^{-1}\lVert A\rVert_{\Phi}\lVert B\rVert_{\Phi} ∥ A B ∥ Φ ≤ ∥ A ∥ ∥ B ∥ Φ ≤ a − 1 ∥ A ∥ Φ ∥ B ∥ Φ . Apply it also to the inverses and take half logarithms. For an eigenvalue vector λ > 0 \lambda>0 λ > 0 , the gauge Hölder inequality with exponent 1 / s 1/s 1/ s gives Φ ( λ s ) ≥ Φ ( λ ) s \Phi(\lambda^{s}) \ge\Phi(\lambda)^{s} Φ ( λ s ) ≥ Φ ( λ ) s . Coordinatewise monotonicity gives Φ ( λ s ) ≤ ∥ λ ∥ ∞ s − 1 Φ ( λ ) ≤ a 1 − s Φ ( λ ) s \Phi(\lambda^{s}) \le\lVert\lambda\rVert_{\infty}^{s-1}\Phi(\lambda) \le a^{1-s}\Phi(\lambda)^{s} Φ ( λ s ) ≤ ∥ λ ∥ ∞ s − 1 Φ ( λ ) ≤ a 1 − s Φ ( λ ) s . Repeat for λ − 1 \lambda^{-1} λ − 1 . Finally t − s s ( P t s ) → e 1 t^{-s}s(P_t^{s}) \to e_1 t − s s ( P t s ) → e 1 and the corresponding inverse vector tends to e n e_n e n , so C Φ ( P t s ) = s log t + log a + o ( 1 ) C_{\Phi}(P_t^{s}) = s\log t+\log a+o(1) C Φ ( P t s ) = s log t + log a + o ( 1 ) . Subtraction proves optimality.
The norm-product estimate is the classical submultiplicativity of a gauge normalised by Φ ( e 1 ) = 1 \Phi(e_1)=1 Φ ( e 1 ) = 1 . The power inequalities are related to Marshall–Olkin’s condition-number inequalities [8 ] ; the normalisation and optimal additive constants are the point of this formulation. For 0 < s ≤ 1 0<s\le1 0 < s ≤ 1 , the equivalent bounds are s C Φ ( P ) − ( 1 − s ) d Φ ≤ C Φ ( P s ) ≤ s C Φ ( P ) sC_{\Phi}(P)-(1-s)d_{\Phi}\le C_{\Phi}(P^{s})\le sC_{\Phi}(P) s C Φ ( P ) − ( 1 − s ) d Φ ≤ C Φ ( P s ) ≤ s C Φ ( P ) .
Corollary 2.2.2. The following are equivalent: product subadditivity of C Φ C_{\Phi} C Φ for all invertibles; the inequality C Φ ( P 2 ) ≤ 2 C Φ ( P ) C_{\Phi}(P^{2})\le2C_{\Phi}(P) C Φ ( P 2 ) ≤ 2 C Φ ( P ) for all P > 0 P>0 P > 0 ; exact positive-power homogeneity; d Φ = 0 d_{\Phi}=0 d Φ = 0 ; and Φ \Phi Φ being a positive multiple of the operator-norm gauge.
Proof. If d Φ > 0 d_{\Phi}>0 d Φ > 0 , the family P t P_t P t violates square subadditivity. If d Φ = 0 d_{\Phi}=0 d Φ = 0 , then Φ ( e 1 ) = Φ ( 1 ) \Phi(e_1)=\Phi(1) Φ ( e 1 ) = Φ ( 1 ) , and the two ℓ ∞ \ell_{\infty} ℓ ∞ bounds force Φ ( x ) = Φ ( e 1 ) ∥ x ∥ ∞ \Phi(x)=\Phi(e_1)\lVert x\rVert_{\infty} Φ ( x ) = Φ ( e 1 ) ∥ x ∥ ∞ . All remaining implications follow from the operator-norm power law and Theorem 2.2.1 .
Applying the power bound to ∣ A ∣ 2 = A ∗ A \lvert A\rvert^{2}=A^{*}A ∣ A ∣ 2 = A ∗ A gives 2 C Φ ( A ) ≤ C Φ ( A ∗ A ) ≤ 2 C Φ ( A ) + d Φ 2C_{\Phi}(A)\le C_{\Phi}(A^{*}A)\le2C_{\Phi}(A)+d_{\Phi} 2 C Φ ( A ) ≤ C Φ ( A ∗ A ) ≤ 2 C Φ ( A ) + d Φ . The full range of defects is 0 ≤ d Φ ≤ log n 0\le d_{\Phi}\le\log n 0 ≤ d Φ ≤ log n , with the endpoints attained exactly by multiples of the operator and trace norms, respectively. For the latter assertion, Φ ( 1 ) = n Φ ( e 1 ) \Phi(1)=n\Phi(e_1) Φ ( 1 ) = n Φ ( e 1 ) makes the two ℓ 1 \ell_{1} ℓ 1 bounds coincide. At fixed individual contrasts, the product excess is also bounded by min { d Φ , OLC ( A ) + OLC ( B ) } \min\{d_{\Phi},\operatorname{OLC}(A)+\operatorname{OLC}(B)\} min { d Φ , OLC ( A ) + OLC ( B )} . If Φ ( x ) > Φ ( e 1 ) ∥ x ∥ ∞ \Phi(x)>\Phi(e_1)\lVert x\rVert_{\infty} Φ ( x ) > Φ ( e 1 ) ∥ x ∥ ∞ whenever x x x has at least two nonzero coordinates, the defect is not attained by invertibles. It can be attained for other gauges: for n = 4 n=4 n = 4 , Φ ( x ) = max { ∥ x ∥ 1 / 2 , ∥ x ∥ ∞ } \Phi(x)=\max\{\lVert x\rVert_{1}/2,\lVert x\rVert_{\infty}\} Φ ( x ) = max {∥ x ∥ 1 /2 , ∥ x ∥ ∞ } and P = diag ( 3 , 1 , 1 , 1 / 3 ) P=\operatorname{diag}(3,1,1,1/3) P = diag ( 3 , 1 , 1 , 1/3 ) give C Φ ( P 2 ) − 2 C Φ ( P ) = log 2 = d Φ C_{\Phi}(P^{2})-2C_{\Phi}(P)=\log2=d_{\Phi} C Φ ( P 2 ) − 2 C Φ ( P ) = log 2 = d Φ .
Schatten contrasts and logarithmic spectral profiles
Let τ = n − 1 Tr \tau=n^{-1}\operatorname{Tr} τ = n − 1 Tr , x i = log s i ( A ) x_i=\log s_i(A) x i = log s i ( A ) , and
ν A = 1 n ∑ i δ x i , L A ( t ) = log ∫ e t x d ν A ( x ) , C p ( A ) = L A ( p ) + L A ( − p ) 2 p ( p > 0 ) . \nu_A=\frac{1}{n}\sum_i\delta_{x_i},\qquad L_A(t)=\log\int e^{tx}\,\mathrm{d}\nu_A(x),\qquad C_p(A)=\frac{L_A(p)+L_A(-p)}{2p}\qquad(p>0). ν A = n 1 i ∑ δ x i , L A ( t ) = log ∫ e t x d ν A ( x ) , C p ( A ) = 2 p L A ( p ) + L A ( − p ) ( p > 0 ) .
For p ≥ 1 p\ge1 p ≥ 1 this is the contrast for the Schatten norm; the same moment formula defines it for 0 < p < 1 0<p<1 0 < p < 1 . Set C ∞ = OLC C_{\infty}=\operatorname{OLC} C ∞ = OLC . The tilted probability measure is d ν A ( t ) = e t x − L A ( t ) d ν A \mathrm{d}\nu_A^{(t)}=e^{tx-L_A(t)}\mathrm{d}\nu_A d ν A ( t ) = e t x − L A ( t ) d ν A . Standard cumulant identities [9 ] yield
L A ′ ( t ) = ∫ x d ν A ( t ) , L A ′ ′ ( t ) = Var ( ν A ( t ) ) = 1 2 ∬ ( x − y ) 2 d ν A ( t ) ( x ) d ν A ( t ) ( y ) . L_A'(t)=\int x\,\mathrm{d}\nu_A^{(t)},\qquad L_A''(t)=\operatorname{Var}(\nu_A^{(t)})=\frac{1}{2}\iint(x-y)^2\,\mathrm{d}\nu_A^{(t)}(x)\mathrm{d}\nu_A^{(t)}(y). L A ′ ( t ) = ∫ x d ν A ( t ) , L A ′′ ( t ) = Var ( ν A ( t ) ) = 2 1 ∬ ( x − y ) 2 d ν A ( t ) ( x ) d ν A ( t ) ( y ) .
Thus p C p pC_p p C p is the even part of L A L_A L A , unaffected by centring the log-spectrum, and O L C = 1 2 ( L A ′ ( + ∞ ) − L A ′ ( − ∞ ) ) \mathrm{OLC}=\frac{1}{2}(L'_A(+\infty)-L'_A(-\infty)) OLC = 2 1 ( L A ′ ( + ∞ ) − L A ′ ( − ∞ )) .
Theorem 2.3.1. For p > 0 p>0 p > 0 put T p = e 2 p C p T_p=e^{2pC_p} T p = e 2 p C p . Then
T p ( A ) = ∬ cosh ( p ( x − y ) ) d ν A ( x ) d ν A ( y ) , T_p(A)=\iint\cosh(p(x-y))\,\mathrm{d}\nu_A(x)\mathrm{d}\nu_A(y), T p ( A ) = ∬ cosh ( p ( x − y )) d ν A ( x ) d ν A ( y ) , d d p C p ( A ) = D ( ν A ( p ) ∥ ν A ) + D ( ν A ( − p ) ∥ ν A ) 2 p 2 ≥ 0 , \frac{\mathrm{d}}{\mathrm{d}p}C_p(A)=\frac{D(\nu_A^{(p)}\mathbin{\|}\nu_A)+D(\nu_A^{(-p)}\mathbin{\|}\nu_A)}{2p^2}\geq0, d p d C p ( A ) = 2 p 2 D ( ν A ( p ) ∥ ν A ) + D ( ν A ( − p ) ∥ ν A ) ≥ 0 , lim p → 0 C p ( A ) p = 1 2 Var ( ν A ) , lim p → ∞ C p ( A ) = O L C ( A ) , \lim_{p\to0}\frac{C_p(A)}{p}=\frac{1}{2}\operatorname{Var}(\nu_A),\qquad\lim_{p\to\infty}C_p(A)=\mathrm{OLC}(A), p → 0 lim p C p ( A ) = 2 1 Var ( ν A ) , p → ∞ lim C p ( A ) = OLC ( A ) , C p ( ∣ A ∣ s ) = ∣ s ∣ C ∣ s ∣ p ( A ) ( s ≠ 0 ) , C p ( A ⊗ B ) = C p ( A ) + C p ( B ) . C_p(|A|^s)=|s|C_{|s|p}(A)\quad(s\ne0),\qquad C_p(A\otimes B)=C_p(A)+C_p(B). C p ( ∣ A ∣ s ) = ∣ s ∣ C ∣ s ∣ p ( A ) ( s = 0 ) , C p ( A ⊗ B ) = C p ( A ) + C p ( B ) . The function C p C_p C p is real-analytic, and its derivative is strictly positive unless A A A is a scalar multiple of a unitary.
Proof. Symmetrise the product of the two moment integrals to obtain the cosh formula. Since D ( ν A ( t ) ∥ ν A ) = t L A ′ ( t ) − L A ( t ) D(\nu_A^{(t)}\mathbin{\|}\nu_A)=tL'_A(t)-L_A(t) D ( ν A ( t ) ∥ ν A ) = t L A ′ ( t ) − L A ( t ) , differentiation gives the second formula. Relative entropy vanishes here only for a point mass. Taylor expansion gives L A ( p ) + L A ( − p ) = p 2 Var ( ν A ) + O ( p 4 ) L_A(p)+L_A(-p)=p^2\operatorname{Var}(\nu_A)+O(p^4) L A ( p ) + L A ( − p ) = p 2 Var ( ν A ) + O ( p 4 ) ; at infinity the largest exponential in each sum gives the limit. The power map sends the measure by x ↦ s x x\mapsto sx x ↦ s x , while tensor products convolve the two log-spectral measures and add their cumulant profiles.
In particular C p ( P s ) > s C p ( P ) C_p(P^s)>sC_p(P) C p ( P s ) > s C p ( P ) for nonscalar P > 0 P>0 P > 0 and s > 1 s>1 s > 1 . The product defect is log n / p \log n/p log n / p , and this remains valid for 0 < p < 1 0<p<1 0 < p < 1 : the singular-value estimate s j ( A B ) ≤ ∥ A ∥ s j ( B ) s_j(AB)\leq\lVert A\rVert s_j(B) s j ( A B ) ≤ ∥ A ∥ s j ( B ) and the bound ∥ A ∥ ≤ n 1 / p ∥ A ∥ p , τ \lVert A\rVert\leq n^{1/p}\lVert A\rVert_{p,\tau} ∥ A ∥ ≤ n 1/ p ∥ A ∥ p , τ give the same proof without using norm convexity.
Endpoint asymptotics and sharp bounds
Let x 1 > ⋯ > x r x_1>\cdots>x_r x 1 > ⋯ > x r be the distinct log-singular values, of multiplicities m 1 , … , m r m_1,\ldots,m_r m 1 , … , m r , with r ≥ 2 r\geq2 r ≥ 2 . Set g + = x 1 − x 2 g_+=x_1-x_2 g + = x 1 − x 2 and g − = x r − 1 − x r g_-=x_{r-1}-x_r g − = x r − 1 − x r .
Theorem 2.4.1. For every p > 0 p>0 p > 0 ,
C p ( A ) = O L C ( A ) − 1 2 p log n 2 m 1 m r + R A ( p ) 2 p , C_p(A)=\mathrm{OLC}(A)-\frac{1}{2p}\log\frac{n^2}{m_1m_r}+\frac{R_A(p)}{2p}, C p ( A ) = OLC ( A ) − 2 p 1 log m 1 m r n 2 + 2 p R A ( p ) , where
0 < R A ( p ) ≤ n − m 1 m 1 e − p g + + n − m r m r e − p g − , 0<R_A(p)\leq\frac{n-m_1}{m_1}e^{-pg_+}+\frac{n-m_r}{m_r}e^{-pg_-}, 0 < R A ( p ) ≤ m 1 n − m 1 e − p g + + m r n − m r e − p g − , R A ( p ) = m 2 m 1 e − p g + ( 1 + o ( 1 ) ) + m r − 1 m r e − p g − ( 1 + o ( 1 ) ) . R_A(p)=\frac{m_2}{m_1}e^{-pg_+}(1+o(1))+\frac{m_{r-1}}{m_r}e^{-pg_-}(1+o(1)). R A ( p ) = m 1 m 2 e − p g + ( 1 + o ( 1 )) + m r m r − 1 e − p g − ( 1 + o ( 1 )) . Hence p ( O L C − C p ) → log n − 1 2 log ( m 1 m r ) p(\mathrm{OLC}-C_p)\to\log n-\frac{1}{2}\log(m_1m_r) p ( OLC − C p ) → log n − 2 1 log ( m 1 m r ) .
Proof. Factor the top exponential out of each moment sum. The remaining logarithms are
R A ( p ) = log ( 1 + ∑ j ≥ 2 m j m 1 e − p ( x 1 − x j ) ) + log ( 1 + ∑ j < r m j m r e − p ( x j − x r ) ) . R_A(p)=\log\left(1+\sum_{j\geq2}\frac{m_j}{m_1}e^{-p(x_1-x_j)}\right)+\log\left(1+\sum_{j<r}\frac{m_j}{m_r}e^{-p(x_j-x_r)}\right). R A ( p ) = log ( 1 + j ≥ 2 ∑ m 1 m j e − p ( x 1 − x j ) ) + log ( 1 + j < r ∑ m r m j e − p ( x j − x r ) ) .
Use log ( 1 + u ) ≤ u \log(1+u)\leq u log ( 1 + u ) ≤ u for the bound and log ( 1 + u ) ∼ u \log(1+u)\sim u log ( 1 + u ) ∼ u for the asymptotics.
The same elementary Laplace asymptotics give
L A ′ ′ ( t ) ∼ g + 2 m 2 m 1 e − g + t , x 1 − L A ′ ( t ) ∼ g + m 2 m 1 e − g + t ( t → ∞ ) , L''_A(t)\sim g_+^2\frac{m_2}{m_1}e^{-g_+t},\qquad x_1-L'_A(t)\sim g_+\frac{m_2}{m_1}e^{-g_+t}\qquad(t\to\infty), L A ′′ ( t ) ∼ g + 2 m 1 m 2 e − g + t , x 1 − L A ′ ( t ) ∼ g + m 1 m 2 e − g + t ( t → ∞ ) ,
and the analogous lower-edge formula. Thus g ± = − lim t → ± ∞ ∣ t ∣ − 1 log L A ′ ′ ( t ) g_\pm=-\lim_{t\to\pm\infty}|t|^{-1}\log L''_A(t) g ± = − lim t → ± ∞ ∣ t ∣ − 1 log L A ′′ ( t ) . Indeed, in the tilted variance the pair of levels 1 , 2 1,2 1 , 2 supplies the leading term; every other pair has smaller exponential order.
:::
Gibbs dictionary. For A β = e − β H A_{\beta}=e^{-\beta H} A β = e − β H and Z ( s ) = Tr e − s H Z(s)=\operatorname{Tr}e^{-sH} Z ( s ) = Tr e − sH ,
L A β ( t ) = log Z ( β t ) − log n , L A β ′ ′ ( t ) = β 2 Var β t ( H ) , OLC ( A β ) = β 2 ( E max − E min ) . L_{A_{\beta}}(t)=\log Z(\beta t)-\log n,\qquad L_{A_{\beta}}^{\prime\prime}(t)=\beta^{2}\operatorname{Var}_{\beta t}(H),\qquad\operatorname{OLC}(A_{\beta})=\frac{\beta}{2}(E_{\max}-E_{\min}). L A β ( t ) = log Z ( β t ) − log n , L A β ′′ ( t ) = β 2 Var β t ( H ) , OLC ( A β ) = 2 β ( E m a x − E m i n ) .
Also C p ( A β ) = [ log Z ( p β ) + log Z ( − p β ) − 2 log n ] / ( 2 p ) C_p(A_{\beta})=[\log Z(p\beta)+\log Z(-p\beta)-2\log n]/(2p) C p ( A β ) = [ log Z ( pβ ) + log Z ( − pβ ) − 2 log n ] / ( 2 p ) . These are the classical partition-function and low-temperature gap formulas [10 ] . Tensor powers multiply L A L_A L A and the contrasts by N N N , but leave the edge gaps unchanged; the adjacent-to-extreme multiplicity ratios acquire a factor N N N .
Theorem 2.4.2. Let z = OLC ( A ) z=\operatorname{OLC}(A) z = OLC ( A ) , p > 0 p>0 p > 0 , and θ n = 4 ⌊ n / 2 ⌋ ⌈ n / 2 ⌉ / n 2 \theta_n=4\lfloor n/2\rfloor\lceil n/2\rceil/n^2 θ n = 4 ⌊ n /2 ⌋ ⌈ n /2 ⌉ / n 2 . Then
1 p log 2 cosh ( p z ) + n − 2 n ≤ C p ( A ) ≤ 1 2 p log ( 1 + θ n sinh 2 ( p z ) ) ≤ 1 p log cosh ( p z ) . \frac{1}{p}\log\frac{2\cosh(pz)+n-2}{n}\le C_p(A)\le\frac{1}{2p}\log\left(1+\theta_n\sinh^2(pz)\right)\le\frac{1}{p}\log\cosh(pz). p 1 log n 2 cosh ( p z ) + n − 2 ≤ C p ( A ) ≤ 2 p 1 log ( 1 + θ n sinh 2 ( p z ) ) ≤ p 1 log cosh ( p z ) . For z > 0 z>0 z > 0 , equality in the first inequality holds exactly when all interior singular values equal s 1 s n \sqrt{s_1s_n} s 1 s n . Equality in the middle bound holds exactly when all singular values are endpoints with multiplicities ⌊ n / 2 ⌋ , ⌈ n / 2 ⌉ \lfloor n/2\rfloor,\lceil n/2\rceil ⌊ n /2 ⌋ , ⌈ n /2 ⌉ , in either order.
Proof. By scaling, take the log-endpoints to be z , − z z,-z z , − z . The product F ( y ) = ( ∑ e p y i ) ( ∑ e − p y i ) F(y)=\left(\sum e^{py_i}\right)\left(\sum e^{-py_i}\right) F ( y ) = ( ∑ e p y i ) ( ∑ e − p y i ) is strictly convex in each coordinate on [ − z , z ] n [-z,z]^n [ − z , z ] n , so its maximum is at a vertex. With ℓ \ell ℓ top endpoints, F = n 2 + 4 ℓ ( n − ℓ ) sinh 2 ( p z ) F=n^2+4\ell(n-\ell)\sinh^2(pz) F = n 2 + 4 ℓ ( n − ℓ ) sinh 2 ( p z ) , maximised precisely by the balanced multiplicities. For the minimum, put X = ∑ i = 2 n − 1 e p y i X=\sum_{i=2}^{n-1}e^{py_i} X = ∑ i = 2 n − 1 e p y i and W = ∑ i = 2 n − 1 e − p y i W=\sum_{i=2}^{n-1}e^{-py_i} W = ∑ i = 2 n − 1 e − p y i . Then X W ≥ ( n − 2 ) 2 XW\ge(n-2)^2 X W ≥ ( n − 2 ) 2 and X + W ≥ 2 ( n − 2 ) X+W\ge2(n-2) X + W ≥ 2 ( n − 2 ) , whence F = ( 2 cosh ( p z ) + X ) ( 2 cosh ( p z ) + W ) ≥ ( 2 cosh ( p z ) + n − 2 ) 2 F=(2\cosh(pz)+X)(2\cosh(pz)+W)\ge(2\cosh(pz)+n-2)^2 F = ( 2 cosh ( p z ) + X ) ( 2 cosh ( p z ) + W ) ≥ ( 2 cosh ( p z ) + n − 2 ) 2 . Both inequalities are equalities precisely when all interior y i = 0 y_i=0 y i = 0 . The n = 2 n=2 n = 2 case follows directly.
The upper bound at p = 1 p=1 p = 1 is the sharp finite-dimensional Schweitzer–Kantorovich inequality, including the odd-dimensional correction due to Lupaş [11 , 12 , 13 ] . General p p p follows by applying it to ∣ A ∣ p |A|^p ∣ A ∣ p . Here θ n = 1 \theta_n=1 θ n = 1 for even n n n and 1 − n − 2 1-n^{-2} 1 − n − 2 for odd n n n . The lower bound is the companion fixed-endpoint minimisation. Consequences include
T p ≤ cosh ( p OLC ) ≤ n T p − ( n − 2 ) 2 , C p ≤ OLC ≤ C p + log n p , \sqrt{T_p}\le\cosh(p\operatorname{OLC})\le\frac{n\sqrt{T_p}-(n-2)}{2},\qquad C_p\le\operatorname{OLC}\le C_p+\frac{\log n}{p}, T p ≤ cosh ( p OLC ) ≤ 2 n T p − ( n − 2 ) , C p ≤ OLC ≤ C p + p log n ,
and C p ≤ min { OLC , p OLC 2 / 2 } C_p\le\min\{\operatorname{OLC},p\operatorname{OLC}^2/2\} C p ≤ min { OLC , p OLC 2 /2 } . In dimension two, C p = p − 1 log cosh ( p OLC ) C_p=p^{-1}\log\cosh(p\operatorname{OLC}) C p = p − 1 log cosh ( p OLC ) exactly.
For a positive definite density matrix ρ \rho ρ and p > 1 p>1 p > 1 , monotonicity of Rényi entropy [14 ] and Jensen’s inequality give
log d − S ( ρ ) ≤ 2 p p − 1 C p ( ρ ) . \log d-S(\rho)\le\frac{2p}{p-1}C_p(\rho). log d − S ( ρ ) ≤ p − 1 2 p C p ( ρ ) .
Indeed τ ( ρ − p ) ≥ d p \tau(\rho^{-p})\ge d^p τ ( ρ − p ) ≥ d p , so Tr ρ p ≤ d 1 − p T p ( ρ ) \operatorname{Tr}\rho^p\le d^{1-p}T_p(\rho) Tr ρ p ≤ d 1 − p T p ( ρ ) and S ≥ S p ≥ log d − 2 p C p / ( p − 1 ) S\ge S_p\ge\log d-2pC_p/(p-1) S ≥ S p ≥ log d − 2 p C p / ( p − 1 ) . In particular
log d − S ( ρ ) ≤ 4 C 2 ( ρ ) ≤ 2 log cosh ( 2 OLC ( ρ ) ) ≤ 4 OLC ( ρ ) 2 . \log d-S(\rho)\le4C_2(\rho)\le2\log\cosh(2\operatorname{OLC}(\rho))\le4\operatorname{OLC}(\rho)^2. log d − S ( ρ ) ≤ 4 C 2 ( ρ ) ≤ 2 log cosh ( 2 OLC ( ρ )) ≤ 4 OLC ( ρ ) 2 .
Letting p → ∞ p\to\infty p → ∞ gives the weaker linear bound log d − S ≤ 2 OLC \log d-S\le2\operatorname{OLC} log d − S ≤ 2 OLC .
Ky Fan contrasts
For the sum-of-k k k -largest-coordinates gauge Φ ( k ) \Phi_{(k)} Φ ( k ) , write C ( k ) = C Φ ( k ) C_{(k)}=C_{\Phi_{(k)}} C ( k ) = C Φ ( k ) . Thus
e 2 C ( k ) ( A ) = ( 1 k ∑ i = 1 k s i ) ( 1 k ∑ i = 1 k s n + 1 − i − 1 ) . e^{2C_{(k)}(A)}=\left(\frac{1}{k}\sum_{i=1}^{k}s_i\right)\left(\frac{1}{k}\sum_{i=1}^{k}s_{n+1-i}^{-1}\right). e 2 C ( k ) ( A ) = ( k 1 i = 1 ∑ k s i ) ( k 1 i = 1 ∑ k s n + 1 − i − 1 ) .
This is the ratio of the arithmetic mean of the top k k k values to the harmonic mean of the bottom k k k values.
Theorem 2.5.1. The contrasts decrease as k k k increases: O L C = C ( 1 ) ≥ ⋯ ≥ C ( n ) = C 1 ≥ 0 \mathrm{OLC}=C_{(1)}\ge\cdots\ge C_{(n)}=C_1\ge0 OLC = C ( 1 ) ≥ ⋯ ≥ C ( n ) = C 1 ≥ 0 . Their product defect is log k \log k log k and their power defect is ( s − 1 ) log k (s-1)\log k ( s − 1 ) log k for s ≥ 1 s\ge1 s ≥ 1 , both sharp. Writing z = O L C ( A ) z=\mathrm{OLC}(A) z = OLC ( A ) , a = min { k , ⌈ n / 2 ⌉ } a=\min\{k,\lceil n/2\rceil\} a = min { k , ⌈ n /2 ⌉} and b = min { k , ⌊ n / 2 ⌋ } b=\min\{k,\lfloor n/2\rfloor\} b = min { k , ⌊ n /2 ⌋} ,
e 2 C ( k ) ≤ ( a e z + ( k − a ) e − z ) ( b e z + ( k − b ) e − z ) k 2 . e^{2C_{(k)}}\le\frac{\left(ae^z+(k-a)e^{-z}\right)\left(be^z+(k-b)e^{-z}\right)}{k^2}. e 2 C ( k ) ≤ k 2 ( a e z + ( k − a ) e − z ) ( b e z + ( k − b ) e − z ) . For z > 0 z>0 z > 0 and 2 k ≤ n 2k\le n 2 k ≤ n , equality holds exactly when the top k k k singular values coincide and the bottom k k k coincide; the right side is e 2 z e^{2z} e 2 z . For 2 k > n 2k>n 2 k > n , equality requires precisely the two endpoints with balanced multiplicities. The sharp lower bounds are
C ( k ) ≥ log e z + k − 1 k ( k < n ) , C ( n ) ≥ log 2 cosh z + n − 2 n . C_{(k)}\ge\log\frac{e^z+k-1}{k}\quad(k<n),\qquad C_{(n)}\ge\log\frac{2\cosh z+n-2}{n}. C ( k ) ≥ log k e z + k − 1 ( k < n ) , C ( n ) ≥ log n 2 cosh z + n − 2 . For n ≥ 3 n\ge3 n ≥ 3 , 2 ≤ k ≤ n 2\le k\le n 2 ≤ k ≤ n and z > 0 z>0 z > 0 , lower equality holds exactly when s 2 = ⋯ = s n − 1 = s 1 s n s_2=\cdots=s_{n-1}=\sqrt{s_1s_n} s 2 = ⋯ = s n − 1 = s 1 s n .
Proof. The average of the top k k k coordinates is decreasing in k k k ; apply this to s s s and s − 1 s^{-1} s − 1 . The defects follow from Theorem 2.2.1 . Scale s 1 = e z s_1=e^z s 1 = e z , s n = e − z s_n=e^{-z} s n = e − z . Let T T T and B B B be the sums of the largest k k k entries of s s s and s − 1 s^{-1} s − 1 . Holding all but one coordinate x x x fixed, their product is the maximum of
α β , α ( β ′ + x − 1 ) , ( α ′ + x ) β , ( α ′ + x ) ( β ′ + x − 1 ) , \alpha\beta,\qquad\alpha(\beta'+x^{-1}),\qquad(\alpha'+x)\beta,\qquad(\alpha'+x)(\beta'+x^{-1}), α β , α ( β ′ + x − 1 ) , ( α ′ + x ) β , ( α ′ + x ) ( β ′ + x − 1 ) ,
where α , α ′ \alpha,\alpha' α , α ′ are the top-k k k and top-( k − 1 ) (k-1) ( k − 1 ) sums of the other coordinates, and β , β ′ \beta,\beta' β , β ′ the reciprocal counterparts. At k = n k=n k = n only the last expression occurs. Each is convex in x > 0 x>0 x > 0 , so a maximum occurs at cube vertices. A vertex with ℓ \ell ℓ top endpoints has
T = min ( k , ℓ ) e z + ( k − min ( k , ℓ ) ) e − z , B = min ( k , n − ℓ ) e z + ( k − min ( k , n − ℓ ) ) e − z . T=\min(k,\ell)e^z+(k-\min(k,\ell))e^{-z},\qquad B=\min(k,n-\ell)e^z+(k-\min(k,n-\ell))e^{-z}. T = min ( k , ℓ ) e z + ( k − min ( k , ℓ )) e − z , B = min ( k , n − ℓ ) e z + ( k − min ( k , n − ℓ )) e − z .
For 2 k ≤ n 2k\le n 2 k ≤ n , the maximal product is k 2 e 2 z k^2e^{2z} k 2 e 2 z and its equality conditions follow directly from T , B ≤ k e z T,B\le ke^z T , B ≤ k e z . For 2 k > n 2k>n 2 k > n , the product is strictly concave in ℓ \ell ℓ on [ n − k , k ] [n-k,k] [ n − k , k ] and decreases beyond that interval, giving the balanced vertices. There are no other maximisers in this case. An interior coordinate at a maximum would make the coordinatewise convex function constant throughout its interval. Since the nonconstant branches above are strictly monotone or strictly convex, the constant branch α β \alpha\beta α β would dominate throughout. Evaluating at both endpoints would require at least k k k other entries equal to e z e^z e z and k k k equal to e − z e^{-z} e − z , impossible when 2 k > n 2k>n 2 k > n .
For the lower bound with 2 ≤ k < n 2\le k<n 2 ≤ k < n , let u 1 ≥ ⋯ ≥ u n − 2 u_1\ge\cdots\ge u_{n-2} u 1 ≥ ⋯ ≥ u n − 2 be the interior values. Write T = e z + X T=e^z+X T = e z + X , B = e z + W B=e^z+W B = e z + W , where X X X sums the top k − 1 k-1 k − 1 interior values and W W W sums reciprocals of the bottom k − 1 k-1 k − 1 . If Y Y Y sums those bottom values, then Y ≤ X Y\le X Y ≤ X and Y W ≥ ( k − 1 ) 2 YW\ge(k-1)^2 Y W ≥ ( k − 1 ) 2 . Therefore
T B ≥ ( e z + X ) ( e z + ( k − 1 ) 2 X ) ≥ ( e z + k − 1 ) 2 . TB\ge(e^z+X)\left(e^z+\frac{(k-1)^2}{X}\right)\ge(e^z+k-1)^2. T B ≥ ( e z + X ) ( e z + X ( k − 1 ) 2 ) ≥ ( e z + k − 1 ) 2 .
Equality forces the bottom values equal, X = Y X=Y X = Y , and X = k − 1 X=k-1 X = k − 1 ; the ordered top and bottom blocks then force every interior value to be one. The case k = n k=n k = n is Theorem 2.4.2 at p = 1 p=1 p = 1 . □
For k < n k<n k < n this also yields C ( k ) ≤ O L C ≤ log ( k e C ( k ) − k + 1 ) C_{(k)}\le\mathrm{OLC}\le\log(ke^{C_{(k)}}-k+1) C ( k ) ≤ OLC ≤ log ( k e C ( k ) − k + 1 ) . The Ky Fan family and the Schatten family retain different information. For A = diag ( 11 , 11 , 7 , 1 ) A=\operatorname{diag}(11,11,7,1) A = diag ( 11 , 11 , 7 , 1 ) and B = diag ( 12 , 9 , 2 , 1 ) B=\operatorname{diag}(12,9,2,1) B = diag ( 12 , 9 , 2 , 1 ) , all four Ky Fan contrasts of A A A are smaller than those of B B B , while C 3 ( A ) ≈ 0.8734 > 0.8595 ≈ C 3 ( B ) C_3(A)\approx0.8734>0.8595\approx C_3(B) C 3 ( A ) ≈ 0.8734 > 0.8595 ≈ C 3 ( B ) .
Homometric spectra and identifiability
For a finitely supported probability measure ν \nu ν , let ν ~ ( E ) = ν ( − E ) \tilde{\nu}(E)=\nu(-E) ν ~ ( E ) = ν ( − E ) . Measures are homometric if ν ∗ ν ~ \nu*\tilde{\nu} ν ∗ ν ~ agrees. Translation and reflection give the trivial examples. Matrices are isocontrast if their C p C_p C p agree for every p > 0 p>0 p > 0 .
Theorem 2.6.1. For A ∈ G L n ( C ) A \in\mathrm{GL}_n(\mathbb{C}) A ∈ GL n ( C ) and B ∈ G L m ( C ) B \in\mathrm{GL}_m(\mathbb{C}) B ∈ GL m ( C ) , the following are equivalent: isocontrast; equality of C p C_p C p on a subset of ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) having a finite accumulation point; and homometry of ν A , ν B \nu_A,\nu_B ν A , ν B . In that case their operator logarithmic contrasts also agree.
Proof. T p ( A ) = ∫ e p u d ( ν A ∗ ν ~ A ) ( u ) T_p(A)=\int e^{pu}\,\mathrm{d}(\nu_A * \widetilde{\nu}_A)(u) T p ( A ) = ∫ e p u d ( ν A ∗ ν A ) ( u ) is an entire finite exponential sum. Agreement on an accumulating set implies agreement everywhere, and distinct real exponentials are linearly independent. The largest point in the support of the autocorrelation is 2 OLC.
For matrices of the same size, this is equality of the multisets of pairwise nonnegative logarithmic differences. This is the classical turnpike problem [15 , 16 , 17 , 18 , 19 ] ; the spectral interpretation should be distinguished from its combinatorial ingredients.
Theorem 2.6.2. If n ≤ 5 n \le5 n ≤ 5 and A , B ∈ G L n ( C ) A,B \in\mathrm{GL}_n(\mathbb{C}) A , B ∈ GL n ( C ) are isocontrast, their singular-value multisets agree up to scaling or scaling and inversion. For every n ≥ 6 n \ge6 n ≥ 6 there are positive definite counterexamples. They can have simple spectrum for 6 ≤ n ≤ 11 6 \le n \le11 6 ≤ n ≤ 11 and n = 13 n=13 n = 13 .
Proof. For the first assertion see the complete small-cardinality argument in § 4.7 below. For the second, exponentiate a nontrivially homometric pair X , X ′ X,X' X , X ′ of integer multisets: A = diag ( q x ) x ∈ X A=\operatorname{diag}(q^x)_{x\in X} A = diag ( q x ) x ∈ X , B = diag ( q x ) x ∈ X ′ B=\operatorname{diag}(q^x)_{x\in X'} B = diag ( q x ) x ∈ X ′ , q > 1 q>1 q > 1 . The following simple sets have equal difference multisets and are neither translates nor reflected translates:
Table 2.6.
The first pair is Bloom’s example [17 ] , with difference multiset { 1 , 2 , … , 13 , 16 , 17 } \{1,2,\ldots,13,16,17\} { 1 , 2 , … , 13 , 16 , 17 } . For the remaining sizes use Laurent polynomials. If f = P Q f=PQ f = P Q and f ′ = P ( z ) Q ( z − 1 ) f'=P(z)Q(z^{-1}) f ′ = P ( z ) Q ( z − 1 ) have nonnegative integer coefficients, then f f ∗ = f ′ f ′ ∗ ff^*=f'f'^* f f ∗ = f ′ f ′ ∗ , where f ∗ ( z ) = f ( z − 1 ) f^*(z)=f(z^{-1}) f ∗ ( z ) = f ( z − 1 ) . If neither P P P nor Q Q Q is reciprocal up to a signed monomial, cancellation in Z [ z ± 1 ] \mathbb{Z}[z^{\pm1}] Z [ z ± 1 ] rules out trivial homometry. For n = r s n=rs n = r s , r , s ≥ 3 r,s\ge3 r , s ≥ 3 , use P = 1 + ( r − 1 ) z P=1+(r-1)z P = 1 + ( r − 1 ) z , Q = 1 + ( s − 1 ) z Q=1+(s-1)z Q = 1 + ( s − 1 ) z . For n = 3 a + 2 b n=3a+2b n = 3 a + 2 b , 1 ≤ a < b ≤ 2 a 1\le a<b\le2a 1 ≤ a < b ≤ 2 a , use P = 1 − z + z 3 P=1-z+z^3 P = 1 − z + z 3 , Q = a + a z + b z 2 + b z 3 + a z 4 Q=a+az+bz^2+bz^3+az^4 Q = a + a z + b z 2 + b z 3 + a z 4 ; the coefficient vectors of f , z 4 f ′ f,z^4f' f , z 4 f ′ are
( a , 0 , b − a , a , 2 a − b , b − a , b , a ) , ( a , b − a , 0 , 2 a − b , b , b − a , a , a ) . (a,0,b-a,a,2a-b,b-a,b,a), \qquad(a,b-a,0,2a-b,b,b-a,a,a). ( a , 0 , b − a , a , 2 a − b , b − a , b , a ) , ( a , b − a , 0 , 2 a − b , b , b − a , a , a ) .
For n = 23 n=23 n = 23 , use P = 1 − z + z 3 P=1-z+z^3 P = 1 − z + z 3 , Q = 3 + 4 z + 7 z 2 + 5 z 3 + 4 z 4 Q=3+4z+7z^2+5z^3+4z^4 Q = 3 + 4 z + 7 z 2 + 5 z 3 + 4 z 4 , producing vectors ( 3 , 1 , 3 , 1 , 3 , 3 , 5 , 4 ) (3,1,3,1,3,3,5,4) ( 3 , 1 , 3 , 1 , 3 , 3 , 5 , 4 ) and ( 4 , 1 , 2 , 1 , 4 , 4 , 4 , 3 ) (4,1,2,1,4,4,4,3) ( 4 , 1 , 2 , 1 , 4 , 4 , 4 , 3 ) . These cover all sizes. For n ≥ 35 n\ge35 n ≥ 35 , the half-open interval [ n / 7 , n / 5 ) [n/7,n/5) [ n /7 , n /5 ) has length at least two, so it contains an integer a ≡ n ( m o d 2 ) a\equiv n\pmod2 a ≡ n ( mod 2 ) ; then b = ( n − 3 a ) / 2 b=(n-3a)/2 b = ( n − 3 a ) /2 satisfies a < b ≤ 2 a a<b\le2a a < b ≤ 2 a . For 6 ≤ n ≤ 34 6\le n\le34 6 ≤ n ≤ 34 , direct enumeration leaves only 6 , 8 , 10 , 11 , 13 , 23 6,8,10,11,13,23 6 , 8 , 10 , 11 , 13 , 23 outside the product and two-parameter families, and these were supplied explicitly.
Proposition 2.6.3. Suppose A A A has simple singular values whose successive logarithmic gaps are linearly independent over Q \mathbb{Q} Q . If B ∈ G L m ( C ) B\in\mathrm{GL}_m(\mathbb{C}) B ∈ GL m ( C ) is isocontrast with A ∈ G L n ( C ) A\in\mathrm{GL}_n(\mathbb{C}) A ∈ GL n ( C ) , then n n n divides m m m , and the singular values of B B B are those of c A cA c A or ( c A ) − 1 (cA)^{-1} ( c A ) − 1 , each repeated m / n m/n m / n times. The gap hypothesis holds outside a countable union of hyperplanes.
Proof. The gap group Γ \Gamma Γ is free abelian of rank n − 1 n-1 n − 1 and its real group algebra is the Laurent polynomial UFD R [ t 1 ± 1 , … , t n − 1 ± 1 ] \mathbb{R}[t_1^{\pm1},\ldots,t_{n-1}^{\pm1}] R [ t 1 ± 1 , … , t n − 1 ± 1 ] . After translation the measure of A A A corresponds to
f = 1 n ( 1 + t n − 1 + t n − 1 t n − 2 + ⋯ + t n − 1 ⋯ t 1 ) . (1) f=\frac{1}{n}(1+t_{n-1}+t_{n-1}t_{n-2}+\cdots+t_{n-1}\cdots t_1). \tag*{(1)} f = n 1 ( 1 + t n − 1 + t n − 1 t n − 2 + ⋯ + t n − 1 ⋯ t 1 ) . (1)
It is irreducible: as a polynomial in t 1 t_1 t 1 it has degree one with unit leading coefficient; in any factorisation the factor independent of t 1 t_1 t 1 must divide that unit. The same holds after Laurent localisation, and for f ∗ f^* f ∗ . Every difference in supp ν B \operatorname{supp}\nu_B supp ν B belongs to Γ \Gamma Γ , so after translation ν B \nu_B ν B is represented by a nonnegative Laurent polynomial h h h of mass one. Homometry says h h ∗ = f f ∗ hh^*=ff^* h h ∗ = f f ∗ . Unique factorisation forces h h h to be a unit times f f f or f ∗ f^* f ∗ : taking both irreducible factors, or neither, would make the other member of h h ∗ hh^* h h ∗ a unit, an impossibility. Mass one fixes the scalar of the unit at one. Thus the measures agree up to translation/reflection, and each of the n n n atoms has multiplicity m / n m/n m / n . Rational gap dependence is exactly a countable union of integer-coefficient hyperplanes.
Factor reversal is the classical algebraic mechanism for integer homometry [20 ] . For example, P = Q = 1 + 2 z P=Q=1+2z P = Q = 1 + 2 z gives multiplicities ( 1 , 4 , 4 ) (1,4,4) ( 1 , 4 , 4 ) and ( 2 , 5 , 2 ) (2,5,2) ( 2 , 5 , 2 ) on three successive levels, a nine-dimensional isocontrast pair. Even the inverse ambiguity can change von Neumann entropy: normalised spectra proportional to ( 1 , 1 , 4 ) (1,1,4) ( 1 , 1 , 4 ) and ( 4 , 4 , 1 ) (4,4,1) ( 4 , 4 , 1 ) are isocontrast but have different entropy. For Bloom’s pair with q = 2 1 / 4 q=2^{1/4} q = 2 1/4 , the normalised entropies are approximately 1.3272 and 1.4076, while C ( 2 ) C_{(2)} C ( 2 ) is approximately 1.2604 and 1.2876. Thus Ky Fan contrasts can distinguish a pair invisible to all C p C_p C p . The full profile L A ( t ) L_A(t) L A ( t ) , unlike its even part, determines ν A \nu_A ν A uniquely: the top level is lim t → ∞ L A ( t ) / t \lim_{t\to\infty}L_A(t)/t lim t → ∞ L A ( t ) / t , its mass is lim e L A ( t ) − x 1 t \lim e^{L_A(t)-x_1t} lim e L A ( t ) − x 1 t , and subtraction of that exponential permits iteration. This is a classical, potentially ill-conditioned exponential reconstruction problem.
Small-cardinality reconstruction
Proof of Theorem 2.6.2 for $n\leq 5. Write the nonnegative successive gap vector in increasing-point order, and let S S S be its sum, the largest pairwise distance. Orient it so that its last gap is at most its first. For n = 1 n=1 n = 1 the claim is trivial, and for n = 2 n=2 n = 2 the gap is S S S . For n = 3 n=3 n = 3 , with gaps ( a , b ) (a,b) ( a , b ) and b ≤ a b\leq a b ≤ a , removal of one copy of S S S leaves { a , b } \{a,b\} { a , b } , so both are determined. For n = 4 n=4 n = 4 , with gaps ( a , b , c ) (a,b,c) ( a , b , c ) , c ≤ a c\leq a c ≤ a , the largest distance after removing one S S S is a + b a+b a + b ; this determines c c c . The sum of all six distances is $3S+b$, determining b b b and then a a a . For n = 5 n=5 n = 5 , write ( a , b , c , d ) (a,b,c,d) ( a , b , c , d ) , d ≤ a d\leq a d ≤ a . The largest distance after one S S S is removed is a + b + c a+b+c a + b + c , determining d d d . The total of all ten distances is 4 S + 2 ( b + c ) 4S+2(b+c) 4 S + 2 ( b + c ) ; hence K = b + c K=b+c K = b + c and a = S − d − K a=S-d-K a = S − d − K are determined. Removing the known six entries S S S , a + K a+K a + K , K + d K+d K + d , K K K , a a a , d d d leaves
R ( b ) = { b , K − b , a + b , K + d − b } . R(b)=\{b,K-b,a+b,K+d-b\}. R ( b ) = { b , K − b , a + b , K + d − b } .
If another normalised vector ( a , b ′ , K − b ′ , d ) (a,b',K-b',d) ( a , b ′ , K − b ′ , d ) has the same distances, equality of the sums of squares of R ( b ) R(b) R ( b ) , R ( b ′ ) R(b') R ( b ′ ) gives either b ′ = b b'=b b ′ = b or b + b ′ = K + ( d − a ) / 2 b+b'=K+(d-a)/2 b + b ′ = K + ( d − a ) /2 . In the latter case b ′ ∈ R ( b ) b'\in R(b) b ′ ∈ R ( b ) . If b ′ = K − b b'=K-b b ′ = K − b , this equation forces a = d a=d a = d and the vectors are reversals. If b ′ = a + b ≠ b b'=a+b\ne b b ′ = a + b = b , then a > 0 a>0 a > 0 and K = 2 b + ( 3 a − d ) / 2 K=2b+(3a-d)/2 K = 2 b + ( 3 a − d ) /2 . Remove the common entry a + b a+b a + b and subtract b b b from the remaining entries to get
{ ( a − d ) / 2 , 2 a , ( a + d ) / 2 } = { 0 , ( 3 a − d ) / 2 , ( 3 a + d ) / 2 } . \{(a-d)/2,2a,(a+d)/2\}=\{0,(3a-d)/2,(3a+d)/2\}. {( a − d ) /2 , 2 a , ( a + d ) /2 } = { 0 , ( 3 a − d ) /2 , ( 3 a + d ) /2 } .
Since 0 ≤ d ≤ a 0\leq d\leq a 0 ≤ d ≤ a and a > 0 a>0 a > 0 , the zero can match only ( a − d ) / 2 (a-d)/2 ( a − d ) /2 , so again a = d a=d a = d and the vectors reverse. Finally, b ′ = K + d − b b'=K+d-b b ′ = K + d − b forces d = − a d=-a d = − a , hence a = d = 0 a=d=0 a = d = 0 , reducing to the first case. All reasoning is multisets reasoning and allows repeated points and zero gaps.
Geodesic contrasts
An alternative to a product of norms is
G Φ ( A ) = min c ∈ R Φ ( log s ( A ) − c 1 ) Φ ( 1 ) . G_{\Phi}(A)=\min_{c\in\mathbb{R}}\frac{\Phi(\log s(A)-c1)}{\Phi(1)}. G Φ ( A ) = c ∈ R min Φ ( 1 ) Φ ( log s ( A ) − c 1 ) .
The minimum exists by coercivity. Geometrically it is the distance from ∣ A ∣ |A| ∣ A ∣ to the scalar ray in the invariant Finsler metric associated with Φ \Phi Φ , divided by Φ ( 1 ) \Phi(1) Φ ( 1 ) [21 , 22 , 23 ] .
Theorem 2.8.1. The functional G Φ G_{\Phi} G Φ has the scaling, unitary, adjoint and inverse invariances of C Φ C_{\Phi} C Φ , and vanishes exactly on scalar multiples of unitaries. For real s s s ,
G Φ ( ∣ A ∣ s ) = ∣ s ∣ G Φ ( A ) , G Φ ( A B ) ≤ G Φ ( A ) + G Φ ( B ) . G_{\Phi}(|A|^{s})=|s|G_{\Phi}(A), \qquad G_{\Phi}(AB)\leq G_{\Phi}(A)+G_{\Phi}(B). G Φ ( ∣ A ∣ s ) = ∣ s ∣ G Φ ( A ) , G Φ ( A B ) ≤ G Φ ( A ) + G Φ ( B ) . Moreover
Φ ( e 1 + e 2 ) Φ ( 1 ) OLC ( A ) ≤ G Φ ( A ) ≤ OLC ( A ) , \frac{\Phi(e_{1}+e_{2})}{\Phi(1)}\operatorname{OLC}(A)\leq G_{\Phi}(A)\leq\operatorname{OLC}(A), Φ ( 1 ) Φ ( e 1 + e 2 ) OLC ( A ) ≤ G Φ ( A ) ≤ OLC ( A ) , and the lower constant is sharp. Thus G Φ = OLC G_{\Phi}=\operatorname{OLC} G Φ = OLC when n = 2 n=2 n = 2 , or for the operator norm.
Proof. The invariances, zero set and homogeneity follow immediately in log-singular coordinates. Horn’s log-majorisation inequality [24 ] states log s ( A B ) ≺ log s ( A ) + log s ( B ) \log s(AB)\prec\log s(A)+\log s(B) log s ( A B ) ≺ log s ( A ) + log s ( B ) , with the vectors ordered decreasingly; equality of the full sums follows from determinants. Subtracting ( c 1 + c 2 ) 1 (c_{1}+c_{2})\mathbf{1} ( c 1 + c 2 ) 1 , applying gauge monotonicity under majorisation, and then the triangle inequality gives
Φ ( log s ( A B ) − ( c 1 + c 2 ) 1 ) ≤ Φ ( log s ( A ) − c 1 1 ) + Φ ( log s ( B ) − c 2 1 ) . \Phi\bigl(\log s(AB)-(c_{1}+c_{2})\mathbf{1}\bigr)\leq\Phi\bigl(\log s(A)-c_{1}\mathbf{1}\bigr)+\Phi\bigl(\log s(B)-c_{2}\mathbf{1}\bigr). Φ ( log s ( A B ) − ( c 1 + c 2 ) 1 ) ≤ Φ ( log s ( A ) − c 1 1 ) + Φ ( log s ( B ) − c 2 1 ) .
Minimise over c 1 , c 2 c_{1},c_{2} c 1 , c 2 . For the comparison, put x = log s ( A ) x=\log s(A) x = log s ( A ) and z = ( x 1 − x n ) / 2 z=(x_{1}-x_{n})/2 z = ( x 1 − x n ) /2 . Centring at the midpoint gives the upper bound. For any c c c , the endpoint deviations α = ∣ x 1 − c ∣ \alpha=|x_{1}-c| α = ∣ x 1 − c ∣ , β = ∣ x n − c ∣ \beta=|x_{n}-c| β = ∣ x n − c ∣ satisfy α + β ≥ 2 z \alpha+\beta\geq2z α + β ≥ 2 z . Monotonicity and symmetry imply
Φ ( x − c 1 ) ≥ Φ ( α e 1 + β e n ) ≥ α + β 2 Φ ( e 1 + e n ) ≥ z Φ ( e 1 + e 2 ) . \Phi(x-c\mathbf{1})\geq\Phi(\alpha e_{1}+\beta e_{n})\geq\frac{\alpha+\beta}{2}\Phi(e_{1}+e_{n})\geq z\Phi(e_{1}+e_{2}). Φ ( x − c 1 ) ≥ Φ ( α e 1 + β e n ) ≥ 2 α + β Φ ( e 1 + e n ) ≥ z Φ ( e 1 + e 2 ) .
The vector ( z , 0 , … , 0 , − z ) (z,0,\ldots,0,-z) ( z , 0 , … , 0 , − z ) , centred at zero, attains the bound.
For Schatten gauges with p ≥ 1 p\geq1 p ≥ 1 , write G p = min c ( ∫ ∣ x − c ∣ p d ν A ) 1 / p G_{p}=\min_{c}\left(\int|x-c|^{p}\,d\nu_{A}\right)^{1/p} G p = min c ( ∫ ∣ x − c ∣ p d ν A ) 1/ p . Then G p G_{p} G p increases to OLC: minimise the monotone L p L^{p} L p norms over the compact interval of spectral endpoints and use uniform convergence there (or Dini’s theorem). The minimising centre is a median for p = 1 p=1 p = 1 , and the mean for p = 2 p=2 p = 2 . Consequently
G 2 = Var ( ν A ) , G 2 ( A ⊗ B ) 2 = G 2 ( A ) 2 + G 2 ( B ) 2 . G_{2}=\sqrt{\operatorname{Var}(\nu_{A})}, \qquad G_{2}(A\otimes B)^{2}=G_{2}(A)^{2}+G_{2}(B)^{2}. G 2 = Var ( ν A ) , G 2 ( A ⊗ B ) 2 = G 2 ( A ) 2 + G 2 ( B ) 2 .
The cosh representation also gives, for every p > 0 p>0 p > 0 ,
C p ( A ) ≥ 1 2 p log ( 1 + p 2 G 2 ( A ) 2 ) , lim p → 0 C p ( A ) p = 1 2 G 2 ( A ) 2 . C_{p}(A)\geq\frac{1}{2p}\log\bigl(1+p^{2}G_{2}(A)^{2}\bigr), \qquad\lim_{p\to0}\frac{C_{p}(A)}{p}=\frac{1}{2}G_{2}(A)^{2}. C p ( A ) ≥ 2 p 1 log ( 1 + p 2 G 2 ( A ) 2 ) , p → 0 lim p C p ( A ) = 2 1 G 2 ( A ) 2 .
Indeed cosh u ≥ 1 + u 2 / 2 \cosh u\geq1+u^{2}/2 cosh u ≥ 1 + u 2 /2 and the mean squared pair difference is twice the variance.
Connections and remaining questions
The two constructions have different tensor laws: C p C_{p} C p is additive, while G 2 2 G_{2}^{2} G 2 2 is additive. Product subadditivity and power homogeneity hold for every G Φ G_{\Phi} G Φ , but for C Φ C_{\Phi} C Φ only the operator norm has both properties without a defect. The unresolved questions include the maximal defect at fixed individual contrasts, the exceptional homometric configurations beyond the generic reconstruction theorem, and whether the complete Ky Fan family determines singular values up to scaling and inversion. The finite-trace continuation of these questions is developed in Chapter 3.
Schatten contrasts in finite von Neumann algebras
Introduction
For an invertible bounded operator A A A the operator Michelson contrast and the operator logarithmic contrast are
OMC ( A ) = κ ( A ) − 1 κ ( A ) + 1 , OLC ( A ) = artanh OMC ( A ) = 1 2 ln κ ( A ) , \operatorname{OMC}(A)=\frac{\kappa(A)-1}{\kappa(A)+1},\qquad\operatorname{OLC}(A)=\operatorname{artanh}\operatorname{OMC}(A)=\frac{1}{2}\ln\kappa(A), OMC ( A ) = κ ( A ) + 1 κ ( A ) − 1 , OLC ( A ) = artanh OMC ( A ) = 2 1 ln κ ( A ) ,
κ ( A ) = ∥ A ∥ ∥ A − 1 ∥ . \kappa(A)=\lVert A\rVert\lVert A^{-1}\rVert. κ ( A ) = ∥ A ∥ ∥ A − 1 ∥ .
see [1 ] and Nikolaeva and Novikov [2 ] . Replacing the operator norm on M n ( C ) M_n(\mathbb{C}) M n ( C ) by the normalised Schatten norms ∥ X ∥ p = τ ( ∣ X ∣ p ) 1 / p \lVert X\rVert_p=\tau(|X|^p)^{1/p} ∥ X ∥ p = τ ( ∣ X ∣ p ) 1/ p , τ = 1 n Tr \tau=\frac{1}{n}\operatorname{Tr} τ = n 1 Tr , one obtains the contrast
C p ( A ) = 1 2 ln ( ∥ A ∥ p ∥ A − 1 ∥ p ) . C_p(A)=\frac{1}{2}\ln(\lVert A\rVert_p\lVert A^{-1}\rVert_p). C p ( A ) = 2 1 ln (∥ A ∥ p ∥ A − 1 ∥ p ) .
which is nonnegative, vanishes exactly on scalar multiples of unitaries, increases in p p p to OLC ( A ) \operatorname{OLC}(A) OLC ( A ) (Proposition 3.2.3 ), and is subadditive under products up to the sharp defect ln n p \frac{\ln n}{p} p l n n (Theorem 3.4.1 below), while OLC ( A ) − C p ( A ) = O ( 1 p ) \operatorname{OLC}(A)-C_p(A)=O\left(\frac{1}{p}\right) OLC ( A ) − C p ( A ) = O ( p 1 ) as p → ∞ p\to\infty p → ∞ (Theorem 3.5.1). The defect ln n p \frac{\ln n}{p} p l n n is unbounded in the dimension, and the definition uses the trace in an essential way. The present paper asks what remains of the Schatten contrast in an infinite-dimensional algebra, and what it detects there. Three questions organise the chapter.
(Q1) Which functionals give a contrast? If the trace is replaced by a positive functional φ \varphi φ , when is the quantity 1 2 p ln ( φ ( ∣ A ∣ p ) φ ( ∣ A − 1 ∣ p ) ) \frac{1}{2p}\ln\bigl(\varphi(|A|^p)\varphi(|A^{-1}|^p)\bigr) 2 p 1 ln ( φ ( ∣ A ∣ p ) φ ( ∣ A − 1 ∣ p ) ) nonnegative?
(Q2) What do the constants of the matrix case become? What replaces the defect ln n p \frac{\ln n}{p} p l n n and the rate OLC − C p ∼ c p \operatorname{OLC}-C_p\sim\frac{c}{p} OLC − C p ∼ p c in a von Neumann algebra?
(Q3) What happens beyond finite traces: for infinite tensor products of matrix algebras, which may be of type III and then carry no trace?
Background
Characterisations of the trace. Gardner [25 ] proved that the inequality ∣ φ ( x ) ∣ ≤ φ ( ∣ x ∣ ) |\varphi(x)|\leq\varphi(|x|) ∣ φ ( x ) ∣ ≤ φ ( ∣ x ∣ ) characterises traces among positive functionals on a C ∗ C^* C ∗ -algebra. Since then many classical trace inequalities, among them those of Young, Hölder, Cauchy–Schwarz, Golden–Thompson, Peierls–Bogoliubov and Araki–Lieb–Thirring, have been shown to hold for a positive functional only if it is tracial; see [26 , 27 , 28 ] . In a parallel line, central elements of operator algebras, rather than tracial functionals, are characterised by inequalities [29 ] . The Michelson contrast of operators arose in exactly this context, as a quantitative measure of centrality [1 ] ; its logarithmic form OLC and the associated geometry were studied in Nikolaeva and Novikov [2 ] . Passing to Schatten norms puts a trace into the definition of the contrast itself, and question (Q1) asks whether the trace property is forced. Theorem 3.3.1 shows that it is.
Noncommutative integration. The quantities ∥ X ∥ p = τ ( ∣ X ∣ p ) 1 / p \lVert X\rVert_p=\tau\left(\lvert X\rvert^p\right)^{1/p} ∥ X ∥ p = τ ( ∣ X ∣ p ) 1/ p are the norms of the noncommutative L p L^p L p -spaces L p ( M , τ ) L^p(\mathcal{M},\tau) L p ( M , τ ) of a semifinite von Neumann algebra [30 , 31 , 32 ] , and e 2 C p ( A ) = ∥ A ∥ p ∥ A − 1 ∥ p e^{2C_p(A)}=\lVert A\rVert_p\lVert A^{-1}\rVert_p e 2 C p ( A ) = ∥ A ∥ p ∥ A − 1 ∥ p is the condition number of A A A in L p ( M , τ ) L^p(\mathcal{M},\tau) L p ( M , τ ) . Banach spaces of L 1 L^1 L 1 -type associated with a single positive operator affiliated with a von Neumann algebra were studied in [33 , 34 ] , and invertibility of a positive element is characterised by the equivalence of the associated norms with the original ones in [35 ] ; the contrast is a quantitative form of such an equivalence. For algebras without a trace the corresponding spaces are those of Haagerup [36 ] ; we do not treat them here and return to this point in the concluding remarks.
Dimension and types. Murray and von Neumann [37 ] classified finite factors by the range of the trace on projections, { 0 , 1 n , … , 1 } \{0,\frac{1}{n},\ldots,1\} { 0 , n 1 , … , 1 } in type I n \mathrm{I}_n I n and [ 0 , 1 ] [0,1] [ 0 , 1 ] in type I I 1 \mathrm{II}_1 II 1 . Theorems 3.4.1 and 3.5.1 show that the two basic constants of the matrix case, the subadditivity defect and the first-order correction to OLC, are functions of this range. For factors of type III, the work of Powers [38 ] , Araki and Woods [39 ] and Connes [40 ] showed that the type of an infinite tensor product of matrix algebras is governed by the ratios of the eigenvalues of the component states, that is, for 2 × 2 2\times2 2 × 2 components, by their condition numbers. § 6.10 makes this dictionary explicit.
Main results
(A) Positivity of the contrast characterises traces (Theorem 3.3.1 ). For a nonzero positive functional φ \varphi φ on a unital C ∗ C^* C ∗ -algebra and p ∈ ( 0 , ∞ ) p\in(0,\infty) p ∈ ( 0 , ∞ ) , the inequality φ ( ∣ A ∣ p ) φ ( ∣ A − 1 ∣ p ) ≥ φ ( 1 ) 2 \varphi(\lvert A\rvert^p)\varphi(\lvert A^{-1}\rvert^p)\ge\varphi(1)^2 φ (∣ A ∣ p ) φ (∣ A − 1 ∣ p ) ≥ φ ( 1 ) 2 for all invertible A A A holds if and only if φ \varphi φ is tracial; the same is true for the symmetry A ↦ A ∗ A\mapsto A^* A ↦ A ∗ of the contrast.
(B) Monotonicity (Proposition 3.2.5 ). If Ψ \Psi Ψ is a positive unital trace-preserving map between tracial von Neumann algebras and A A A is positive and invertible, then C p ( Ψ ( A ) ) ≤ C p ( A ) C_p(\Psi(A))\le C_p(A) C p ( Ψ ( A )) ≤ C p ( A ) for 1 ≤ p < ∞ 1\le p<\infty 1 ≤ p < ∞ , and this fails for p < 1 p<1 p < 1 . In particular the contrast decreases under conditional expectations and under partial traces, and for p ≥ 1 p\ge1 p ≥ 1 every contrast in the hyperfinite factor of type I I 1 \mathrm{II}_1 II 1 is a monotone limit of matrix contrasts (Example 3.4.3).
(C) The trace must be finite; the defect is a dimension (Proposition 3.2.1, Theorem 3.4.1 , Corollary 3.4.2). If τ ( 1 ) = ∞ \tau(1)=\infty τ ( 1 ) = ∞ , then ∥ A ∥ p ∥ A − 1 ∥ p = ∞ \lVert A\rVert_p\lVert A^{-1}\rVert_p=\infty ∥ A ∥ p ∥ A − 1 ∥ p = ∞ for every invertible τ \tau τ -measurable A A A . For a faithful normal tracial state,
C p ( A B ) ≤ C p ( A ) + C p ( B ) + 1 p ln 1 ε M , ε M = inf { τ ( e ) : e ≠ 0 a projection } , C_p(AB)\le C_p(A)+C_p(B)+\frac{1}{p}\ln\frac{1}{\varepsilon_{\mathcal{M}}},\qquad\varepsilon_{\mathcal{M}}=\inf\{\tau(e):e\ne0\text{ a projection}\}, C p ( A B ) ≤ C p ( A ) + C p ( B ) + p 1 ln ε M 1 , ε M = inf { τ ( e ) : e = 0 a projection } ,
and for factors the constant is optimal: ln n p \frac{\ln n}{p} p l n n in type I n \mathrm{I}_n I n and + ∞ +\infty + ∞ in type I I 1 \mathrm{II}_1 II 1 . Thus the defect is 1 p \frac{1}{p} p 1 times the logarithm of the reciprocal of the smallest Murray–von Neumann dimension of a nonzero projection, and the types I n \mathrm{I}_n I n , I I 1 \mathrm{II}_1 II 1 , I ∞ \mathrm{I}_\infty I ∞ , I I ∞ \mathrm{II}_\infty II ∞ of semifinite factors are distinguished by the behaviour of C p C_p C p .
(D) The rate of convergence to OLC (Theorem 3.5.1, Proposition 3.5.2 ). For invertible A A A with ∣ A ∣ \lvert A\rvert ∣ A ∣ not scalar,
lim p → ∞ 2 p ( OLC ( A ) − C p ( A ) ) = − ln τ ( e + ) − ln τ ( e − ) ∈ [ 2 ln 2 , ∞ ] , \lim_{p\to\infty}2p\bigl(\operatorname{OLC}(A)-C_p(A)\bigr)=-\ln\tau(e_+)-\ln\tau(e_-)\in[2\ln2,\infty], p → ∞ lim 2 p ( OLC ( A ) − C p ( A ) ) = − ln τ ( e + ) − ln τ ( e − ) ∈ [ 2 ln 2 , ∞ ] ,
where e + e_+ e + are the spectral projections of ∣ A ∣ \lvert A\rvert ∣ A ∣ at the top and the bottom of its spectrum. Hence OLC − C p = O ( 1 / p ) \operatorname{OLC}-C_p=O(1/p) OLC − C p = O ( 1/ p ) for all invertible elements if and only if M \mathcal{M} M is finite-dimensional; in M n ( C ) M_n(\mathbb{C}) M n ( C ) … the limits form a finite set, in a factor of type II1 _1 1 they fill [ 2 ln 2 , ∞ ] [2\ln2,\infty] [ 2 ln 2 , ∞ ] , and when e ± = 0 e_{\pm}=0 e ± = 0 and the power-law edge hypotheses of Proposition 3.5.2 hold, the correction is of order ln p p \frac{\ln p}{p} p l n p with a coefficient given by the edge exponents of the spectral measure: 3 2 \frac{3}{2} 2 3 for the exponential of a semicircular element, 1 2 \frac{1}{2} 2 1 for Haar unitaries (Examples 3.5.3, 3.5.4). The second derivative of the profile, L A ′ ′ L_A^{\prime\prime} L A ′′ , decays exponentially exactly when the top of the spectrum is isolated, at the rate given by the gap; this last fact is classical and is recalled in Proposition 3.5.5 .
(E) Araki–Woods factors (§ 6.10 ). For an infinite tensor product of 2 × 2 2\times2 2 × 2 density matrices ρ n \rho_n ρ n we reformulate the classical criteria of Araki–Woods and Connes: types II1 _1 1 and I correspond to summable saturation of the two lower bounds C 1 ≥ 0 C_1\ge0 C 1 ≥ 0 and C 1 ≥ O L C − ln 2 C_1\ge\mathrm{OLC}-\ln2 C 1 ≥ OLC − ln 2 , type III to the divergence of ∑ n min { a n , b n } \sum_n\min\{a_n,b_n\} ∑ n min { a n , b n } , where a n , b n a_n,b_n a n , b n are the two saturation gaps, type II∞ _\infty ∞ to the remaining case, and Connes’ invariant is T = { t : ∑ n ( 1 − O M C ( ρ n ) 2 ) sin 2 ( t O L C ( ρ n ) ) < ∞ } T=\{t:\sum_n(1-\mathrm{OMC}(\rho_n)^2)\sin^2(t\,\mathrm{OLC}(\rho_n))<\infty\} T = { t : ∑ n ( 1 − OMC ( ρ n ) 2 ) sin 2 ( t OLC ( ρ n )) < ∞ } (Proposition 3.6.2 ; this is a reformulation, not a new theorem). For components of arbitrary size with uniformly bounded contrast we prove that the factor is of type II1 _1 1 if and only if the Schatten contrasts of the partial products ρ 1 ⊗ ⋯ ⊗ ρ N \rho_1\otimes\cdots\otimes\rho_N ρ 1 ⊗ ⋯ ⊗ ρ N stay bounded (Theorem 3.6.4 ). Without the bound, boundedness of C p C_p C p for some p ≥ 1 p\ge1 p ≥ 1 still implies type II1 _1 1 (Proposition 3.6.5), but explicit examples show that this fails for p < 1 p<1 p < 1 and that a factor of type II1 _1 1 may have unbounded C p C_p C p for all p ≥ 2 p\ge2 p ≥ 2 , or for all p > 0 p>0 p > 0 (Example 3.6.6).
Relation to classical results
The trace characterisation, the factor defect and the finite-dimensionality criterion are proved below. Monotonicity follows from established Jensen trace inequalities, and the spectral-edge estimates use classical Laplace asymptotics. The two-dimensional Araki–Woods criteria are explicitly presented as reformulations of the classical classification; the bounded-component contrast criterion is obtained by uniform comparison with its square-root criterion. No classification of factors beyond those cited results is claimed.
Scope
Unbounded affiliated operators, for which O L C = ∞ \mathrm{OLC}=\infty OLC = ∞ , and finite sections of unbounded operators on a Hilbert space are not treated here; the definitions of § 6.4 are, however, stated for them, to fix the notation.
Schatten contrasts in a tracial von Neumann algebra
Throughout, M \mathcal{M} M is a von Neumann algebra with a faithful normal semifinite trace τ \tau τ , and for a τ \tau τ -measurable operator X X X we put ∥ X ∥ p = τ ( ∣ X ∣ p ) 1 / p ∈ [ 0 , ∞ ] \lVert X\rVert_p=\tau(|X|^p)^{1/p}\in[0,\infty] ∥ X ∥ p = τ ( ∣ X ∣ p ) 1/ p ∈ [ 0 , ∞ ] , p ∈ ( 0 , ∞ ) p\in(0,\infty) p ∈ ( 0 , ∞ ) [30 , 31 ] . We first show that the trace has to be finite, and then set up the contrast for a tracial state.
The Schatten contrast requires a finite trace
Proposition 3.2.1. Let M \mathcal{M} M be a von Neumann algebra with a faithful normal semifinite trace τ \tau τ such that τ ( 1 ) = ∞ \tau(1)=\infty τ ( 1 ) = ∞ , and let A A A be a τ \tau τ -measurable operator with τ \tau τ -measurable inverse. Then ∥ A ∥ p ∥ A − 1 ∥ p = ∞ \lVert A\rVert_p\lVert A^{-1}\rVert_p=\infty ∥ A ∥ p ∥ A − 1 ∥ p = ∞ for every p ∈ ( 0 , ∞ ) p\in(0,\infty) p ∈ ( 0 , ∞ ) , where ∥ X ∥ p = τ ( ∣ X ∣ p ) 1 / p ∈ [ 0 , ∞ ] \lVert X\rVert_p=\tau(|X|^p)^{1/p}\in[0,\infty] ∥ X ∥ p = τ ( ∣ X ∣ p ) 1/ p ∈ [ 0 , ∞ ] . In particular this holds in B ( H ) B(\mathcal{H}) B ( H ) , dim H = ∞ \dim\mathcal{H}=\infty dim H = ∞ , with the usual trace and the Schatten classes.
Proof. In the polar decomposition A = U ∣ A ∣ A=U|A| A = U ∣ A ∣ the partial isometry U U U is unitary because A A A is invertible, and ∣ A ∣ − 1 = U ∣ A − 1 ∣ U ∗ |A|^{-1}=U|A^{-1}|U^* ∣ A ∣ − 1 = U ∣ A − 1 ∣ U ∗ . Since the generalised singular values are unitarily invariant, μ ( U X U ∗ ) = μ ( X ) \mu(UXU^*)=\mu(X) μ ( U X U ∗ ) = μ ( X ) [31 ] , we have τ ( ∣ A − 1 ∣ p ) = τ ( ∣ A ∣ − p ) \tau(|A^{-1}|^p)=\tau(|A|^{-p}) τ ( ∣ A − 1 ∣ p ) = τ ( ∣ A ∣ − p ) . As t p + t − p ≥ 2 t^p+t^{-p}\ge2 t p + t − p ≥ 2 for t > 0 t>0 t > 0 , the functional calculus gives ∣ A ∣ p + ∣ A ∣ − p ≥ 2 ⋅ 1 |A|^p+|A|^{-p}\ge2\cdot1 ∣ A ∣ p + ∣ A ∣ − p ≥ 2 ⋅ 1 , hence τ ( ∣ A ∣ p ) + τ ( ∣ A ∣ − p ) = τ ( ∣ A ∣ p + ∣ A ∣ − p ) ≥ 2 τ ( 1 ) = ∞ \tau(|A|^p)+\tau(|A|^{-p})=\tau(|A|^p+|A|^{-p})\ge2\tau(1)=\infty τ ( ∣ A ∣ p ) + τ ( ∣ A ∣ − p ) = τ ( ∣ A ∣ p + ∣ A ∣ − p ) ≥ 2 τ ( 1 ) = ∞ , so at least one of the two factors is infinite, while the other is nonzero because τ \tau τ is faithful. □
So the Schatten contrasts require a finite trace. This is consistent with the matrix case, in which the defect ln n p \frac{\ln n}{p} p l n n is unbounded in n n n (Theorem 3.4.1 ).
Contrast in a finite von Neumann algebra
Let M \mathcal{M} M be a von Neumann algebra with a faithful normal tracial state τ \tau τ , and M ~ \widetilde{\mathcal{M}} M the ∗ * ∗ -algebra of closed densely defined operators affiliated with M \mathcal{M} M (all of them are τ \tau τ -measurable since τ \tau τ is finite) [30 , 31 ] . If A ∈ M ~ A \in\widetilde{\mathcal{M}} A ∈ M is injective, then its right support is 1 1 1 ; the left support is equivalent to it, hence equal to 1 1 1 since M \mathcal{M} M is finite, so A A A has dense range and A − 1 ∈ M ~ A^{-1} \in\widetilde{\mathcal{M}} A − 1 ∈ M ; moreover the polar decomposition A = U ∣ A ∣ A = U|A| A = U ∣ A ∣ has U U U unitary, and ∣ A − 1 ∣ = U ∣ A ∣ − 1 U ∗ |A^{-1}| = U|A|^{-1}U^* ∣ A − 1 ∣ = U ∣ A ∣ − 1 U ∗ . The logarithmic spectral measure of A A A is the Borel probability measure
ν A ( E ) = τ ( 1 E ( ln ∣ A ∣ ) ) , E ⊂ R , \nu_A(E) = \tau(1_E(\ln|A|)), \qquad E \subset\mathbb{R}, ν A ( E ) = τ ( 1 E ( ln ∣ A ∣ )) , E ⊂ R ,
so that ∥ A ∥ p p = τ ( ∣ A ∣ p ) = ∫ e p x d ν A \|A\|_p^p = \tau(|A|^p) = \int e^{px}\,d\nu_A ∥ A ∥ p p = τ ( ∣ A ∣ p ) = ∫ e p x d ν A and ∥ A − 1 ∥ p p = τ ( ∣ A ∣ − p ) = ∫ e − p x d ν A \|A^{-1}\|_p^p = \tau(|A|^{-p}) = \int e^{-px}\,d\nu_A ∥ A − 1 ∥ p p = τ ( ∣ A ∣ − p ) = ∫ e − p x d ν A , with values in ( 0 , ∞ ] (0,\infty] ( 0 , ∞ ] . Every Borel probability measure on R \mathbb{R} R arises in this way, already for M = L ∞ [ 0 , 1 ] \mathcal{M}=L^\infty[0,1] M = L ∞ [ 0 , 1 ] with the Lebesgue integral as trace and A A A the multiplication operator by exp ( Q ν ) \exp(Q_\nu) exp ( Q ν ) , Q ν Q_\nu Q ν being the quantile function of ν \nu ν .
Definition 3.2.2. For a Borel probability measure ν \nu ν on R \mathbb{R} R and p ∈ ( 0 , ∞ ) p \in(0,\infty) p ∈ ( 0 , ∞ ) put
C p ( ν ) = 1 2 p ( ln ∫ e p x d ν ( x ) + ln ∫ e − p x d ν ( x ) ) ∈ [ 0 , ∞ ] , C_p(\nu) = \frac{1}{2p}\left(\ln\int e^{px}\,d\nu(x) + \ln\int e^{-px}\,d\nu(x)\right) \in[0,\infty], C p ( ν ) = 2 p 1 ( ln ∫ e p x d ν ( x ) + ln ∫ e − p x d ν ( x ) ) ∈ [ 0 , ∞ ] , C ∞ ( ν ) = 1 2 diam supp ν ∈ [ 0 , ∞ ] , C_\infty(\nu) = \frac{1}{2}\operatorname{diam}\operatorname{supp}\nu\in[0,\infty], C ∞ ( ν ) = 2 1 diam supp ν ∈ [ 0 , ∞ ] , and C p ( A ) = C p ( ν A ) = 1 2 p ln ( ∥ A ∥ p ∥ A − 1 ∥ p ) C_p(A)=C_p(\nu_A)=\frac{1}{2p}\ln(\|A\|_p\|A^{-1}\|_p) C p ( A ) = C p ( ν A ) = 2 p 1 ln ( ∥ A ∥ p ∥ A − 1 ∥ p ) , OLC ( A ) = C ∞ ( ν A ) \operatorname{OLC}(A)=C_\infty(\nu_A) OLC ( A ) = C ∞ ( ν A ) for injective A ∈ M ~ A \in\widetilde{\mathcal{M}} A ∈ M . The critical exponent is p c ( ν ) = sup { p > 0 : C p ( ν ) < ∞ } ∈ [ 0 , ∞ ] p_c(\nu)=\sup\{p>0:C_p(\nu)<\infty\}\in[0,\infty] p c ( ν ) = sup { p > 0 : C p ( ν ) < ∞ } ∈ [ 0 , ∞ ] (with sup ∅ = 0 \sup\varnothing=0 sup ∅ = 0 ).
For M = M n ( C ) \mathcal{M}=M_n(\mathbb{C}) M = M n ( C ) , τ = 1 n Tr \tau=\frac{1}{n}\operatorname{Tr} τ = n 1 Tr , this is the Schatten contrast of the introduction. We write L ν ( t ) = ln ∫ e t x d ν ( x ) ∈ ( − ∞ , ∞ ] L_\nu(t)=\ln\int e^{tx}\,d\nu(x)\in(-\infty,\infty] L ν ( t ) = ln ∫ e t x d ν ( x ) ∈ ( − ∞ , ∞ ] for the cumulant generating function and L A = L ν A L_A=L_{\nu_A} L A = L ν A , so that 2 p C p ( A ) = L A ( p ) + L A ( − p ) 2pC_p(A)=L_A(p)+L_A(-p) 2 p C p ( A ) = L A ( p ) + L A ( − p ) , and we recall that OMC = tanh OLC \operatorname{OMC}=\tanh\operatorname{OLC} OMC = tanh OLC . When ν A \nu_A ν A has a first moment, 2 p C p ( A ) = K A ( p ) + K A ( − p ) 2pC_p(A)=K_A(p)+K_A(-p) 2 p C p ( A ) = K A ( p ) + K A ( − p ) , where K A ( t ) = L A ( t ) − t ∫ x d ν A K_A(t)=L_A(t)-t\int x\,d\nu_A K A ( t ) = L A ( t ) − t ∫ x d ν A is the centred cumulant generating function. We call L A L_A L A the logarithmic spectral profile of A A A ; the quantity p C p ( A ) = 1 2 ( K A ( p ) + K A ( − p ) ) pC_p(A)=\frac{1}{2}(K_A(p)+K_A(-p)) p C p ( A ) = 2 1 ( K A ( p ) + K A ( − p )) is the even part of the centred profile, which is why the edge exponents and the masses of the spectral measure at its edges appear in § 6.8 . We recall the classical facts that L A L_A L A is convex and that, in the interior of the set where L A L_A L A is finite, L A ′ ( t ) = ∫ x d ν A ( t ) L_A'(t)=\int x\,d\nu_A^{(t)} L A ′ ( t ) = ∫ x d ν A ( t ) and L A ′ ′ ( t ) = Var ( ν A ( t ) ) L_A''(t)=\operatorname{Var}(\nu_A^{(t)}) L A ′′ ( t ) = Var ( ν A ( t ) ) for the tilted measure d ν A ( t ) = e t x − L A ( t ) d ν A d\nu_A^{(t)}=e^{tx-L_A(t)}d\nu_A d ν A ( t ) = e t x − L A ( t ) d ν A .
Proposition 3.2.3. Let ν \nu ν be a Borel probability measure on R \mathbb{R} R and 0 < p < ∞ 0<p<\infty 0 < p < ∞ .
(a) C p ( ν ) ≥ 0 C_p(\nu)\geq0 C p ( ν ) ≥ 0 , with equality if and only if ν \nu ν is a point mass.
(b) C p ( ν ) < ∞ C_p(\nu)<\infty C p ( ν ) < ∞ if and only if ∫ e p ∣ x ∣ d ν < ∞ \int e^{p|x|}\,d\nu<\infty ∫ e p ∣ x ∣ d ν < ∞ .
(c) p ↦ C p ( ν ) p\mapsto C_p(\nu) p ↦ C p ( ν ) is nondecreasing and lim p → ∞ C p ( ν ) = C ∞ ( ν ) \lim_{p\to\infty}C_p(\nu)=C_\infty(\nu) lim p → ∞ C p ( ν ) = C ∞ ( ν ) .
(d) C p ( ν ) ≤ 1 p ln cosh ( p C ∞ ( ν ) ) C_p(\nu)\leq\frac{1}{p}\ln\cosh(pC_\infty(\nu)) C p ( ν ) ≤ p 1 ln cosh ( p C ∞ ( ν )) whenever C ∞ ( ν ) < ∞ C_\infty(\nu)<\infty C ∞ ( ν ) < ∞ .
(e) If ν s \nu_s ν s is the image of ν \nu ν under x ↦ s x x\mapsto sx x ↦ s x , s ≠ 0 s\neq0 s = 0 , then C p ( ν s ) = ∣ s ∣ C ∣ s ∣ p ( ν ) C_p(\nu_s)=|s|C_{|s|p}(\nu) C p ( ν s ) = ∣ s ∣ C ∣ s ∣ p ( ν ) ; and C p ( ν ∗ ν ′ ) = C p ( ν ) + C p ( ν ′ ) C_p(\nu*\nu')=C_p(\nu)+C_p(\nu') C p ( ν ∗ ν ′ ) = C p ( ν ) + C p ( ν ′ ) .
Proof. (a) By the Cauchy–Schwarz inequality,
1 = ( ∫ e p x / 2 e − p x / 2 d ν ) 2 ≤ ∫ e p x d ν ∫ e − p x d ν , 1=\left(\int e^{px/2}e^{-px/2}\,d\nu\right)^2\leq\int e^{px}\,d\nu\int e^{-px}\,d\nu, 1 = ( ∫ e p x /2 e − p x /2 d ν ) 2 ≤ ∫ e p x d ν ∫ e − p x d ν ,
with equality iff e p x e^{px} e p x is ν \nu ν -a.e. constant. (b) Both integrals are positive, so their product is finite iff both are, and e p ∣ x ∣ ≤ e p x + e − p x ≤ 2 e p ∣ x ∣ e^{p|x|} \le e^{px} + e^{-px} \le2e^{p|x|} e p ∣ x ∣ ≤ e p x + e − p x ≤ 2 e p ∣ x ∣ . (c) On a probability space the L p L^p L p -norms of the functions e x e^x e x and e − x e^{-x} e − x are nondecreasing in p p p and converge to their essential suprema. (d) The case C ∞ ( ν ) = 0 C_{\infty}(\nu)=0 C ∞ ( ν ) = 0 follows from (a). Let z = C ∞ ( ν ) > 0 z=C_{\infty}(\nu)>0 z = C ∞ ( ν ) > 0 . Since C p C_p C p is invariant under translations of ν \nu ν , we may assume supp ν ⊂ [ − z , z ] \operatorname{supp}\nu\subset[-z,z] supp ν ⊂ [ − z , z ] . By convexity, e ± p y ≤ cosh ( p z ) ± y z sinh ( p z ) e^{\pm py} \le\cosh(pz) \pm\frac{y}{z}\sinh(pz) e ± p y ≤ cosh ( p z ) ± z y sinh ( p z ) for y ∈ [ − z , z ] y \in[-z,z] y ∈ [ − z , z ] ; integrating, with μ = ∫ y d ν \mu=\int y\,d\nu μ = ∫ y d ν ,
∫ e p y d ν ∫ e − p y d ν ≤ cosh 2 ( p z ) − μ 2 z 2 sinh 2 ( p z ) ≤ cosh 2 ( p z ) . \int e^{py}\,d\nu\int e^{-py}\,d\nu\le\cosh^2(pz)-\frac{\mu^2}{z^2}\sinh^2(pz)\le\cosh^2(pz). ∫ e p y d ν ∫ e − p y d ν ≤ cosh 2 ( p z ) − z 2 μ 2 sinh 2 ( p z ) ≤ cosh 2 ( p z ) .
(e) ∫ e ± p x d ν s = ∫ e ± p s x d ν \int e^{\pm px}\,d\nu_s=\int e^{\pm psx}\,d\nu ∫ e ± p x d ν s = ∫ e ± p s x d ν , and ∫ e t x d ( ν ∗ ν ′ ) = ∫ e t x d ν ∫ e t x d ν ′ \int e^{tx}\,d(\nu*\nu')=\int e^{tx}\,d\nu\int e^{tx}\,d\nu' ∫ e t x d ( ν ∗ ν ′ ) = ∫ e t x d ν ∫ e t x d ν ′ .
Remark 3.2.4. For injective elements of M ^ \widehat{\mathcal{M}} M there is no subadditivity estimate at a fixed exponent: if p < p c ( ν A ) < 2 p p<p_c(\nu_A)<2p p < p c ( ν A ) < 2 p for a positive A A A , then C p ( A ) < ∞ C_p(A)<\infty C p ( A ) < ∞ but C p ( A 2 ) = 2 C 2 p ( A ) = ∞ C_p(A^2)=2C_{2p}(A)=\infty C p ( A 2 ) = 2 C 2 p ( A ) = ∞ . For bounded invertible elements the failure of subadditivity is quantified in Theorem 3.4.1 .
Monotonicity under trace-preserving unital positive maps
The contrast of positive elements decreases under coarse-graining. This is a consequence of a form of Jensen’s trace inequality (cf. [41 , 42 ] ); since the function t − p t^{-p} t − p is not defined at 0 and Ψ \Psi Ψ need not be a self-map, we include a short proof.
Proposition 3.2.5. Let ( M , τ ) (\mathcal{M},\tau) ( M , τ ) and ( N , τ ′ ) (\mathcal{N},\tau') ( N , τ ′ ) be von Neumann algebras with faithful normal tracial states, and let Ψ : M → N \Psi:\mathcal{M}\to\mathcal{N} Ψ : M → N be a positive, unital linear map with τ ′ ∘ Ψ = τ \tau'\circ\Psi=\tau τ ′ ∘ Ψ = τ . Let A ∈ M A\in\mathcal{M} A ∈ M be positive and invertible. Then Ψ ( A ) \Psi(A) Ψ ( A ) is positive and invertible,
OLC ( Ψ ( A ) ) ≤ OLC ( A ) , C p ( Ψ ( A ) ) ≤ C p ( A ) for 1 ≤ p < ∞ . \operatorname{OLC}(\Psi(A))\le\operatorname{OLC}(A),\qquad C_p(\Psi(A))\le C_p(A)\quad\text{for }1\le p<\infty. OLC ( Ψ ( A )) ≤ OLC ( A ) , C p ( Ψ ( A )) ≤ C p ( A ) for 1 ≤ p < ∞. In particular this holds for τ \tau τ -preserving conditional expectations onto von Neumann subalgebras.
Proof. Let m ≤ A ≤ M m\le A\le M m ≤ A ≤ M with 0 < m ≤ M 0<m\le M 0 < m ≤ M the extreme points of the spectrum of A A A . Since Ψ \Psi Ψ is positive and unital, m ≤ Ψ ( A ) ≤ M m\le\Psi(A)\le M m ≤ Ψ ( A ) ≤ M , which gives the first two statements. We claim that for every continuous convex function f f f on [ m , M ] [m,M] [ m , M ] ,
τ ′ ( f ( Ψ ( A ) ) ) ≤ τ ( f ( A ) ) . (2) \tau'(f(\Psi(A)))\le\tau(f(A)). \tag*{(2)} τ ′ ( f ( Ψ ( A ))) ≤ τ ( f ( A )) . (2)
Applied to f ( t ) = t p f(t)=t^p f ( t ) = t p , p ≥ 1 p\ge1 p ≥ 1 , and to f ( t ) = t − p f(t)=t^{-p} f ( t ) = t − p this gives τ ′ ( Ψ ( A ) ± p ) ≤ τ ( A ± p ) \tau'(\Psi(A)^{\pm p})\le\tau(A^{\pm p}) τ ′ ( Ψ ( A ) ± p ) ≤ τ ( A ± p ) , hence the claim for C p C_p C p .
To prove (2) put B = Ψ ( A ) B=\Psi(A) B = Ψ ( A ) , let B ⊂ N \mathcal{B}\subset\mathcal{N} B ⊂ N be the abelian von Neumann algebra generated by B B B , and let E E E be the τ ′ \tau' τ ′ -preserving conditional expectation of N \mathcal{N} N onto B \mathcal{B} B . The map Θ ( g ) = E ( Ψ ( g ( A ) ) ) \Theta(g)=E(\Psi(g(A))) Θ ( g ) = E ( Ψ ( g ( A ))) , g ∈ C [ m , M ] g\in C[m,M] g ∈ C [ m , M ] , is positive and unital with values in B \mathcal{B} B , and Θ ( id ) = E ( B ) = B \Theta(\operatorname{id})=E(B)=B Θ ( id ) = E ( B ) = B . If ℓ ( t ) = α t + β \ell(t)=\alpha t+\beta ℓ ( t ) = α t + β is an affine function with ℓ ≤ f \ell\le f ℓ ≤ f on [ m , M ] [m,M] [ m , M ] , then Θ ( f ) ≥ Θ ( ℓ ) = α B + β = ℓ ( B ) \Theta(f)\ge\Theta(\ell)=\alpha B+\beta=\ell(B) Θ ( f ) ≥ Θ ( ℓ ) = α B + β = ℓ ( B ) . As f f f is the supremum of countably many such ℓ \ell ℓ , and B \mathcal{B} B is an algebra of functions, we get Θ ( f ) ≥ f ( B ) \Theta(f)\ge f(B) Θ ( f ) ≥ f ( B ) . Therefore τ ′ ( f ( B ) ) ≤ τ ′ ( Θ ( f ) ) = τ ′ ( Ψ ( f ( A ) ) ) = τ ( f ( A ) ) \tau'(f(B))\le\tau'(\Theta(f))=\tau'(\Psi(f(A)))=\tau(f(A)) τ ′ ( f ( B )) ≤ τ ′ ( Θ ( f )) = τ ′ ( Ψ ( f ( A ))) = τ ( f ( A )) .
Remark 3.2.6. (i) The restriction p ≥ 1 p\ge1 p ≥ 1 is necessary. In M 3 ( C ) M_3(\mathbb{C}) M 3 ( C ) let Ψ ( X ) = 1 2 ( X + U X U ∗ ) \Psi(X)=\frac{1}{2}(X+UXU^*) Ψ ( X ) = 2 1 ( X + U X U ∗ ) , where U U U is the permutation matrix exchanging the first two basis vectors, and A = diag ( 16 , 1 , 1 16 ) A=\operatorname{diag}(16,1,\frac{1}{16}) A = diag ( 16 , 1 , 16 1 ) . Then Ψ ( A ) = diag ( 17 2 , 17 2 , 1 16 ) \Psi(A)=\operatorname{diag}(\frac{17}{2},\frac{17}{2},\frac{1}{16}) Ψ ( A ) = diag ( 2 17 , 2 17 , 16 1 ) and, for p = 1 2 p=\frac{1}{2} p = 2 1 , e C 1 / 2 ( A ) = τ ( A 1 / 2 ) τ ( A − 1 / 2 ) = 49 16 = 3.0625 e^{C_{1/2}(A)}=\tau(A^{1/2})\tau(A^{-1/2})=\frac{49}{16}=3.0625 e C 1/2 ( A ) = τ ( A 1/2 ) τ ( A − 1/2 ) = 16 49 = 3.0625 , whereas e C 1 / 2 ( Ψ ( A ) ) = 3.1661 … e^{C_{1/2}(\Psi(A))}=3.1661\ldots e C 1/2 ( Ψ ( A )) = 3.1661 … More generally, for A = diag ( s , 1 , s − 1 ) A=\operatorname{diag}(s,1,s^{-1}) A = diag ( s , 1 , s − 1 ) one has Ψ ( A ) = diag ( s + 1 2 , s + 1 2 , s − 1 ) \Psi(A)=\operatorname{diag}(\frac{s+1}{2},\frac{s+1}{2},s^{-1}) Ψ ( A ) = diag ( 2 s + 1 , 2 s + 1 , s − 1 ) and
τ ( Ψ ( A ) p ) τ ( Ψ ( A ) − p ) τ ( A p ) τ ( A − p ) ⟶ 2 1 − p ( s → ∞ ) , \frac{\tau(\Psi(A)^p)\,\tau(\Psi(A)^{-p})}{\tau(A^p)\,\tau(A^{-p})}\longrightarrow2^{1-p}\qquad(s\to\infty), τ ( A p ) τ ( A − p ) τ ( Ψ ( A ) p ) τ ( Ψ ( A ) − p ) ⟶ 2 1 − p ( s → ∞ ) ,
so that C p ( Ψ ( A ) ) > C p ( A ) C_p(\Psi(A)) > C_p(A) C p ( Ψ ( A )) > C p ( A ) for all large s s s , for every p ∈ ( 0 , 1 ) p \in(0,1) p ∈ ( 0 , 1 ) .
(ii) Unitality is necessary: for a non-scalar positive S S S with τ ( S ) = 1 \tau(S) = 1 τ ( S ) = 1 , the map X ↦ τ ( X ) S X \mapsto\tau(X)S X ↦ τ ( X ) S is positive and trace preserving and sends 1 1 1 , of contrast 0 0 0 , to S S S .
(iii) For matrices, Proposition 3.2.5 applies to pinchings, to mixtures of unitary conjugations and to the normalised partial trace d 2 − 1 Tr 2 : M d 1 ⊗ M d 2 → M d 1 d_2^{-1}\operatorname{Tr}_2: M_{d_1} \otimes M_{d_2} \to M_{d_1} d 2 − 1 Tr 2 : M d 1 ⊗ M d 2 → M d 1 . Since C p C_p C p is invariant under scaling, for a positive definite density matrix ρ \rho ρ on C d 1 ⊗ C d 2 \mathbb{C}^{d_1} \otimes\mathbb{C}^{d_2} C d 1 ⊗ C d 2 this gives
C p ( Tr 2 ρ ) ≤ C p ( ρ ) , OLC ( Tr 2 ρ ) ≤ OLC ( ρ ) ( 1 ≤ p < ∞ ) C_p(\operatorname{Tr}_2 \rho) \le C_p(\rho), \qquad\operatorname{OLC}(\operatorname{Tr}_2 \rho) \le\operatorname{OLC}(\rho) \qquad(1 \le p < \infty) C p ( Tr 2 ρ ) ≤ C p ( ρ ) , OLC ( Tr 2 ρ ) ≤ OLC ( ρ ) ( 1 ≤ p < ∞ )
(the second inequality also follows from Chapter 1): the contrast of a marginal does not exceed the contrast of the joint state. We do not pursue quantum-information interpretations here; in particular we make no claim that the defect of § 6.7 is a measure of entanglement.
Positivity of the contrast characterises traces
In [1 ] the Michelson contrast was related to inequalities that characterise tracial functionals, in the tradition of Gardner’s theorem [25 ] . The Schatten contrast gives a characterisation of the same kind. Let A \mathcal{A} A be a unital C ∗ C^* C ∗ -algebra and φ \varphi φ a positive linear functional on A \mathcal{A} A . For invertible A ∈ A A \in\mathcal{A} A ∈ A and p ∈ ( 0 , ∞ ) p \in(0,\infty) p ∈ ( 0 , ∞ ) put
κ p φ ( A ) = ( φ ( ∣ A ∣ p ) φ ( ∣ A − 1 ∣ p ) ) 1 / p . \kappa_p^\varphi(A) = \left(\varphi(|A|^p)\varphi(|A^{-1}|^p)\right)^{1/p}. κ p φ ( A ) = ( φ ( ∣ A ∣ p ) φ ( ∣ A − 1 ∣ p ) ) 1/ p .
If φ = τ \varphi= \tau φ = τ is a tracial state this is e 2 C p ( A ) e^{2C_p(A)} e 2 C p ( A ) .
Theorem 3.3.1. Let φ ≠ 0 \varphi\ne0 φ = 0 be a positive linear functional on a unital C ∗ C^* C ∗ -algebra A \mathcal{A} A and p ∈ ( 0 , ∞ ) p \in(0,\infty) p ∈ ( 0 , ∞ ) . The following are equivalent:
(a) φ \varphi φ is tracial, i.e. φ ( X Y ) = φ ( Y X ) \varphi(XY) = \varphi(YX) φ ( X Y ) = φ ( Y X ) for all X , Y ∈ A X,Y \in\mathcal{A} X , Y ∈ A ;
(b) κ p φ ( A ) ≥ κ p φ ( 1 ) \kappa_p^\varphi(A) \ge\kappa_p^\varphi(1) κ p φ ( A ) ≥ κ p φ ( 1 ) , i.e. φ ( ∣ A ∣ p ) φ ( ∣ A − 1 ∣ p ) ≥ φ ( 1 ) 2 \varphi(|A|^p)\varphi(|A^{-1}|^p) \ge\varphi(1)^2 φ ( ∣ A ∣ p ) φ ( ∣ A − 1 ∣ p ) ≥ φ ( 1 ) 2 , for all invertible A ∈ A A \in\mathcal{A} A ∈ A ;
(c) κ p φ ( A ∗ ) = κ p φ ( A ) \kappa_p^\varphi(A^*) = \kappa_p^\varphi(A) κ p φ ( A ∗ ) = κ p φ ( A ) for all invertible A ∈ A A \in\mathcal{A} A ∈ A .
Proof. Let A A A be invertible with polar decomposition A = U P A = UP A = U P , P = ∣ A ∣ P = |A| P = ∣ A ∣ , U = A P − 1 U = AP^{-1} U = A P − 1 unitary. Then
∣ A ∣ = P , ∣ A − 1 ∣ = U P − 1 U ∗ , ∣ A ∗ ∣ = U P U ∗ , ∣ ( A ∗ ) − 1 ∣ = P − 1 . (3) |A| = P, \qquad|A^{-1}| = UP^{-1}U^*, \qquad|A^*| = UPU^*, \qquad|(A^*)^{-1}| = P^{-1}. \tag*{(3)} ∣ A ∣ = P , ∣ A − 1 ∣ = U P − 1 U ∗ , ∣ A ∗ ∣ = U P U ∗ , ∣ ( A ∗ ) − 1 ∣ = P − 1 . (3)
(a)⇒ \Rightarrow ⇒ (b),(c). If φ \varphi φ is tracial, then by (3) κ p φ ( A ) p = φ ( P p ) φ ( P − p ) = κ p φ ( A ∗ ) p \kappa_p^\varphi(A)^p = \varphi(P^p)\varphi(P^{-p}) = \kappa_p^\varphi(A^*)^p κ p φ ( A ) p = φ ( P p ) φ ( P − p ) = κ p φ ( A ∗ ) p , and by the Cauchy–Schwarz inequality for φ \varphi φ , φ ( 1 ) 2 = φ ( P p / 2 P − p / 2 ) 2 ≤ φ ( P p ) φ ( P − p ) \varphi(1)^2 = \varphi(P^{p/2}P^{-p/2})^2 \le\varphi(P^p)\varphi(P^{-p}) φ ( 1 ) 2 = φ ( P p /2 P − p /2 ) 2 ≤ φ ( P p ) φ ( P − p ) .
(b)⇒ \Rightarrow ⇒ (a). Fix X = X ∗ ∈ A X = X^* \in\mathcal{A} X = X ∗ ∈ A and a unitary U ∈ A U \in\mathcal{A} U ∈ A . For real ε \varepsilon ε with ∣ ε ∣ ∥ X ∥ < 1 |\varepsilon|\|X\| < 1 ∣ ε ∣∥ X ∥ < 1 let P ε = ( 1 + ε X ) 1 / p P_\varepsilon= (1+\varepsilon X)^{1/p} P ε = ( 1 + εX ) 1/ p and A ε = U P ε A_\varepsilon= UP_\varepsilon A ε = U P ε . By (3) ,
f ( ε ) : = κ p φ ( A ε ) p = φ ( 1 + ε X ) φ ( U ( 1 + ε X ) − 1 U ∗ ) = φ ( 1 ) 2 + ε φ ( 1 ) ( φ ( X ) − φ ( U X U ∗ ) ) + O ( ε 2 ) , \begin{aligned}
f(\varepsilon) := \kappa_p^\varphi(A_\varepsilon)^p
&= \varphi(1+\varepsilon X)\varphi\left(U(1+\varepsilon X)^{-1}U^*\right) \\
&= \varphi(1)^2+\varepsilon\varphi(1)\left(\varphi(X)-\varphi(UXU^*)\right)+O(\varepsilon^2),
\end{aligned} f ( ε ) := κ p φ ( A ε ) p = φ ( 1 + εX ) φ ( U ( 1 + εX ) − 1 U ∗ ) = φ ( 1 ) 2 + εφ ( 1 ) ( φ ( X ) − φ ( U X U ∗ ) ) + O ( ε 2 ) ,
because ( 1 + ε X ) − 1 = 1 − ε X + O ( ε 2 ) (1+\varepsilon X)^{-1}=1-\varepsilon X+O(\varepsilon^2) ( 1 + εX ) − 1 = 1 − εX + O ( ε 2 ) in norm. By (b), f f f has a minimum at ε = 0 \varepsilon=0 ε = 0 , so the linear term vanishes; since φ ( 1 ) > 0 \varphi(1)>0 φ ( 1 ) > 0 , φ ( U X U ∗ ) = φ ( X ) \varphi(UXU^*)=\varphi(X) φ ( U X U ∗ ) = φ ( X ) . By linearity this holds for all X ∈ A X \in\mathcal{A} X ∈ A ; replacing X X X by X U XU X U gives φ ( U X ) = φ ( X U ) \varphi(UX)=\varphi(XU) φ ( U X ) = φ ( X U ) , and since the unitaries span A \mathcal{A} A , φ \varphi φ is tracial.
(c)⇒ \Rightarrow ⇒ (a). With the same notation, by (3) ,
0 = κ p φ ( A ε ) p − κ p φ ( A ε ∗ ) p = 2 ε φ ( 1 ) ( φ ( X ) − φ ( U X U ∗ ) ) + O ( ε 2 ) , 0 = \kappa_p^\varphi(A_\varepsilon)^p-\kappa_p^\varphi(A_\varepsilon^*)^p = 2\varepsilon\varphi(1)\left(\varphi(X)-\varphi(UXU^*)\right)+O(\varepsilon^2), 0 = κ p φ ( A ε ) p − κ p φ ( A ε ∗ ) p = 2 εφ ( 1 ) ( φ ( X ) − φ ( U X U ∗ ) ) + O ( ε 2 ) ,
and we conclude as before. ▫
Thus, for a non-tracial state, the normalised contrast 1 2 p ln ( φ ( ∣ A ∣ p ) φ ( ∣ A − 1 ∣ p ) ) \frac{1}{2p}\ln\bigl(\varphi(|A|^p)\varphi(|A^{-1}|^p)\bigr) 2 p 1 ln ( φ ( ∣ A ∣ p ) φ ( ∣ A − 1 ∣ p ) ) takes negative values and distinguishes A A A from A ∗ A^* A ∗ : nonnegativity of the contrast and the symmetry C p ( A ) = C p ( A ∗ ) C_p(A)=C_p(A^*) C p ( A ) = C p ( A ∗ ) are each equivalent to the trace property.
Remark 3.3.2. Many trace inequalities (Cauchy–Schwarz, Hölder, Young, Golden–Thompson, Peierls–Bogoliubov, inequalities for the modulus and others) are known to characterise tracial functionals among positive ones, often already when restricted to projections; see [25 , 43 , 44 , 45 , 26 , 27 , 28 , 46 ] and the references in [1 ] . Condition (b) involves the two operators ∣ A ∣ p |A|^p ∣ A ∣ p and ∣ A − 1 ∣ p = U ∣ A ∣ − p U ∗ |A^{-1}|^p=U|A|^{-p}U^* ∣ A − 1 ∣ p = U ∣ A ∣ − p U ∗ , which do not commute in general; only for a tracial φ \varphi φ does it reduce to the Cauchy–Schwarz inequality used in the proof. We have not found (b) or (c) in this product form in the literature. The proof shows that it suffices to require (b) or (c) for invertible elements in an arbitrarily small norm neighbourhood of the unitary group.
The defect of a finite von Neumann algebra and the type of a factor
Let ( M , τ ) (\mathcal{M},\tau) ( M , τ ) be a von Neumann algebra with a faithful normal tracial state, GL ( M ) \operatorname{GL}(\mathcal{M}) GL ( M ) its group of invertible elements, and
ε M = inf { τ ( e ) : e ∈ M a nonzero projection } ∈ [ 0 , 1 ] , \varepsilon_{\mathcal{M}}=\inf\{\tau(e):e\in\mathcal{M}\text{ a nonzero projection}\}\in[0,1], ε M = inf { τ ( e ) : e ∈ M a nonzero projection } ∈ [ 0 , 1 ] ,
D p ( M ) = sup { C p ( P 2 ) − 2 C p ( P ) : P ∈ GL ( M ) , P > 0 } ∈ [ 0 , ∞ ] . D_p(\mathcal{M})=\sup\{C_p(P^2)-2C_p(P):P\in\operatorname{GL}(\mathcal{M}),\ P>0\}\in[0,\infty]. D p ( M ) = sup { C p ( P 2 ) − 2 C p ( P ) : P ∈ GL ( M ) , P > 0 } ∈ [ 0 , ∞ ] .
For M = M n ( C ) \mathcal{M}=M_n(\mathbb{C}) M = M n ( C ) , ε M = 1 n \varepsilon_{\mathcal{M}}=\frac{1}{n} ε M = n 1 . More generally, for a finite-dimensional algebra M = ⨁ k M n k ( C ) \mathcal{M}=\bigoplus_k M_{n_k}(\mathbb{C}) M = ⨁ k M n k ( C ) with the tracial state τ = ∑ k w k 1 n k Tr \tau=\sum_k w_k\frac{1}{n_k}\operatorname{Tr} τ = ∑ k w k n k 1 Tr , where w k > 0 w_k>0 w k > 0 and ∑ k w k = 1 \sum_k w_k=1 ∑ k w k = 1 , one has ε M = min k w k / n k \varepsilon_{\mathcal{M}}=\min_k w_k/n_k ε M = min k w k / n k .
Theorem 3.4.1. Let ( M , τ ) (\mathcal{M},\tau) ( M , τ ) be as above and p ∈ ( 0 , ∞ ) p\in(0,\infty) p ∈ ( 0 , ∞ ) .
(a) For all A , B ∈ GL ( M ) A,B\in\operatorname{GL}(\mathcal{M}) A , B ∈ GL ( M ) , C p ( A B ) ≤ C p ( A ) + C p ( B ) + 1 p ln 1 ε M C_p(AB)\leq C_p(A)+C_p(B)+\frac{1}{p}\ln\frac{1}{\varepsilon_{\mathcal{M}}} C p ( A B ) ≤ C p ( A ) + C p ( B ) + p 1 ln ε M 1 (with no content if ε M = 0 \varepsilon_{\mathcal{M}}=0 ε M = 0 ).
(b) If M ≠ C \mathcal{M}\neq\mathbb{C} M = C is a factor, then D p ( M ) = 1 p ln 1 ε M D_p(\mathcal{M})=\frac{1}{p}\ln\frac{1}{\varepsilon_{\mathcal{M}}} D p ( M ) = p 1 ln ε M 1 . Thus D p ( M ) = ln n p D_p(\mathcal{M})=\frac{\ln n}{p} D p ( M ) = p l n n if M \mathcal{M} M is of type I n \mathrm{I}_n I n , and D p ( M ) = + ∞ D_p(\mathcal{M})=+\infty D p ( M ) = + ∞ if M \mathcal{M} M is of type I I 1 \mathrm{II}_1 II 1 .
Proof. (a) We may assume ε M > 0 \varepsilon_{\mathcal{M}}>0 ε M > 0 . Let X ∈ M X\in\mathcal{M} X ∈ M , X ≠ 0 X\neq0 X = 0 , and 0 < η < ∥ X ∥ 0<\eta<\lVert X\rVert 0 < η < ∥ X ∥ . The spectral projection e η = 1 ( ∥ X ∥ − η , ∥ X ∥ ] ( ∣ X ∣ ) e_\eta=\mathbf{1}_{(\lVert X\rVert-\eta,\lVert X\rVert]}(|X|) e η = 1 (∥ X ∥ − η , ∥ X ∥] ( ∣ X ∣ ) is nonzero and ∣ X ∣ p ≥ ( ∥ X ∥ − η ) p e η |X|^p\geq(\lVert X\rVert-\eta)^p e_\eta ∣ X ∣ p ≥ (∥ X ∥ − η ) p e η , so τ ( ∣ X ∣ p ) ≥ ( ∥ X ∥ − η ) p ε M \tau(|X|^p)\geq(\lVert X\rVert-\eta)^p\varepsilon_{\mathcal{M}} τ ( ∣ X ∣ p ) ≥ (∥ X ∥ − η ) p ε M . Hence ∥ X ∥ ≤ ε M − 1 / p ∥ X ∥ p \lVert X\rVert\leq\varepsilon_{\mathcal{M}}^{-1/p}\lVert X\rVert_p ∥ X ∥ ≤ ε M − 1/ p ∥ X ∥ p . Together with ∥ X Y ∥ p ≤ ∥ X ∥ ∥ Y ∥ p \lVert XY\rVert_p\leq\lVert X\rVert\lVert Y\rVert_p ∥ X Y ∥ p ≤ ∥ X ∥ ∥ Y ∥ p [31 ] this gives ∥ A B ∥ p ≤ ε M − 1 / p ∥ A ∥ p ∥ B ∥ p \lVert AB\rVert_p\leq\varepsilon_{\mathcal{M}}^{-1/p}\lVert A\rVert_p\lVert B\rVert_p ∥ A B ∥ p ≤ ε M − 1/ p ∥ A ∥ p ∥ B ∥ p , and the same for ( A B ) − 1 = B − 1 A − 1 (AB)^{-1}=B^{-1}A^{-1} ( A B ) − 1 = B − 1 A − 1 . Multiply and take half the logarithm.
(b) The upper bound is (a) with A = B = P A=B=P A = B = P . For the lower bound let δ ∈ ( 0 , 1 2 ] \delta\in(0,\frac{1}{2}] δ ∈ ( 0 , 2 1 ] be such that there are orthogonal projections e , f ∈ M e,f\in\mathcal{M} e , f ∈ M with τ ( e ) = τ ( f ) = δ \tau(e)=\tau(f)=\delta τ ( e ) = τ ( f ) = δ , and put P t = t e + ( 1 − e − f ) + t − 1 f P_t=te+(1-e-f)+t^{-1}f P t = t e + ( 1 − e − f ) + t − 1 f , t > 1 t>1 t > 1 . For every q > 0 q>0 q > 0 , τ ( P t q ) = τ ( P t − q ) = δ t q + ( 1 − 2 δ ) + δ t − q \tau(P_t^q)=\tau(P_t^{-q})=\delta t^q+(1-2\delta)+\delta t^{-q} τ ( P t q ) = τ ( P t − q ) = δ t q + ( 1 − 2 δ ) + δ t − q , so C q ( P t ) = 1 q ln ( δ t q + 1 − 2 δ + δ t − q ) = ln t + 1 q ln δ + o ( 1 ) C_q(P_t)=\frac{1}{q}\ln(\delta t^q+1-2\delta+\delta t^{-q})=\ln t+\frac{1}{q}\ln\delta+o(1) C q ( P t ) = q 1 ln ( δ t q + 1 − 2 δ + δ t − q ) = ln t + q 1 ln δ + o ( 1 ) as t → ∞ t\to\infty t → ∞ . By the power law of Proposition 3.2.3 ,
C p ( P t 2 ) − 2 C p ( P t ) = 2 C 2 p ( P t ) − 2 C p ( P t ) ⟶ 1 p ln δ − 2 p ln δ = 1 p ln 1 δ . C_p(P_t^2)-2C_p(P_t)=2C_{2p}(P_t)-2C_p(P_t)\longrightarrow\frac{1}{p}\ln\delta-\frac{2}{p}\ln\delta=\frac{1}{p}\ln\frac{1}{\delta}. C p ( P t 2 ) − 2 C p ( P t ) = 2 C 2 p ( P t ) − 2 C p ( P t ) ⟶ p 1 ln δ − p 2 ln δ = p 1 ln δ 1 .
In the factor of type I n \mathrm{I}_n I n , n ≥ 2 n\geq2 n ≥ 2 , one can take δ = 1 n = ε M \delta=\frac{1}{n}=\varepsilon_{\mathcal{M}} δ = n 1 = ε M ; in a factor of type I I 1 \mathrm{II}_1 II 1 every δ ∈ ( 0 , 1 2 ] \delta\in(0,\frac{1}{2}] δ ∈ ( 0 , 2 1 ] is admissible and ε M = 0 \varepsilon_{\mathcal{M}}=0 ε M = 0 .
Corollary 3.4.2 (The type of a semifinite factor). Let M \mathcal{M} M be a factor with a faithful normal semifinite trace τ \tau τ , normalised by τ ( 1 ) = 1 \tau(1)=1 τ ( 1 ) = 1 if τ \tau τ is finite, and let p ∈ ( 0 , ∞ ) p\in(0,\infty) p ∈ ( 0 , ∞ ) . For M = C \mathcal{M}=\mathbb{C} M = C we have D p ( M ) = 0 D_p(\mathcal{M})=0 D p ( M ) = 0 , and (a) holds with n = 1 n=1 n = 1 .
(a) M \mathcal{M} M is of type I n \mathrm{I}_n I n , n < ∞ n < \infty n < ∞ , if and only if ∥ A ∥ p ∥ A − 1 ∥ p < ∞ \lVert A\rVert_p\lVert A^{-1}\rVert_p < \infty ∥ A ∥ p ∥ A − 1 ∥ p < ∞ on GL ( M ) \operatorname{GL}(\mathcal{M}) GL ( M ) and D p ( M ) < ∞ D_p(\mathcal{M}) < \infty D p ( M ) < ∞ ; in this case n = exp ( p D p ( M ) ) n = \exp(pD_p(\mathcal{M})) n = exp ( p D p ( M )) .
(b) M \mathcal{M} M is of type I I 1 \mathrm{II}_1 II 1 if and only if ∥ A ∥ p ∥ A − 1 ∥ p < ∞ \lVert A\rVert_p\lVert A^{-1}\rVert_p < \infty ∥ A ∥ p ∥ A − 1 ∥ p < ∞ on GL ( M ) \operatorname{GL}(\mathcal{M}) GL ( M ) and D p ( M ) = ∞ D_p(\mathcal{M}) = \infty D p ( M ) = ∞ .
(c) M \mathcal{M} M is of type I ∞ \mathrm{I}_\infty I ∞ or I I ∞ \mathrm{II}_\infty II ∞ if and only if ∥ A ∥ p ∥ A − 1 ∥ p = ∞ \lVert A\rVert_p\lVert A^{-1}\rVert_p = \infty ∥ A ∥ p ∥ A − 1 ∥ p = ∞ for all A ∈ GL ( M ) A \in\operatorname{GL}(\mathcal{M}) A ∈ GL ( M ) . The two types are distinguished by their corners: for a nonzero projection e e e with τ ( e ) < ∞ \tau(e) < \infty τ ( e ) < ∞ , the finite factor e M e e\mathcal{M}e e M e with the tracial state τ / τ ( e ) \tau/\tau(e) τ / τ ( e ) has D p ( e M e ) = 1 p ln rank ( e ) < ∞ D_p(e\mathcal{M}e) = \frac{1}{p}\ln\operatorname{rank}(e) < \infty D p ( e M e ) = p 1 ln rank ( e ) < ∞ in type I ∞ \mathrm{I}_\infty I ∞ and D p ( e M e ) = ∞ D_p(e\mathcal{M}e) = \infty D p ( e M e ) = ∞ in type I I ∞ \mathrm{II}_\infty II ∞ .
Factors of type I I I \mathrm{III} III carry no normal semifinite trace, and the tracial Schatten contrast used here is unavailable; the operator-norm contrast OLC \operatorname{OLC} OLC remains defined on bounded invertible elements.
Proof. If τ ( 1 ) = 1 \tau(1) = 1 τ ( 1 ) = 1 and A ∈ GL ( M ) A \in\operatorname{GL}(\mathcal{M}) A ∈ GL ( M ) , then ∥ A ∥ p ∥ A − 1 ∥ p ≤ ∥ A ∥ ∥ A − 1 ∥ < ∞ \lVert A\rVert_p\lVert A^{-1}\rVert_p \leq\lVert A\rVert\lVert A^{-1}\rVert< \infty ∥ A ∥ p ∥ A − 1 ∥ p ≤ ∥ A ∥ ∥ A − 1 ∥ < ∞ ; if τ ( 1 ) = ∞ \tau(1) = \infty τ ( 1 ) = ∞ , Proposition 3.2.1 applies. The rest follows from Theorem 3.4.1 and the classification of semifinite factors [47 ] ; in type I ∞ \mathrm{I}_\infty I ∞ the corner e M e e\mathcal{M}e e M e is isomorphic to M r ( C ) M_r(\mathbb{C}) M r ( C ) , r = rank ( e ) r = \operatorname{rank}(e) r = rank ( e ) , and in type I I ∞ \mathrm{II}_\infty II ∞ it is of type I I 1 \mathrm{II}_1 II 1 .
In other words, for a finite factor the sharp defect is 1 p \frac{1}{p} p 1 times the logarithm of the reciprocal of the smallest Murray–von Neumann dimension τ ( e ) \tau(e) τ ( e ) of a nonzero projection e e e ; for matrices it is the constant ln n p \frac{\ln n}{p} p l n n .
Example 3.4.3 (The hyperfinite factor). Let R = ⨂ k ≥ 1 ( M 2 ( C ) , 1 2 Tr ) R = \bigotimes_{k\geq1}(M_2(\mathbb{C}),\frac{1}{2}\operatorname{Tr}) R = ⨂ k ≥ 1 ( M 2 ( C ) , 2 1 Tr ) be the hyperfinite factor of type I I 1 \mathrm{II}_1 II 1 , with its increasing sequence of subfactors R k ≅ M 2 k ( C ) R_k \cong M_{2^k}(\mathbb{C}) R k ≅ M 2 k ( C ) , whose union is dense, and the trace-preserving conditional expectations E k : R → R k E_k:R \to R_k E k : R → R k . By Theorem 3.4.1 ,
D p ( R k ) = k ln 2 p ⟶ ∞ = D p ( R ) ( k → ∞ ) : D_p(R_k) = \frac{k\ln2}{p} \longrightarrow\infty= D_p(R) \qquad(k \to\infty): D p ( R k ) = p k ln 2 ⟶ ∞ = D p ( R ) ( k → ∞ ) :
the defect grows linearly in k k k , that is, logarithmically in the matrix size, and becomes infinite in the limit. The contrast itself, on the other hand, is continuous along this approximation: for positive invertible A ∈ R A \in R A ∈ R and 1 ≤ p < ∞ 1 \leq p < \infty 1 ≤ p < ∞ ,
C p ( E k ( A ) ) ↑ C p ( A ) ( k → ∞ ) . C_p(E_k(A)) \mathrel{\uparrow} C_p(A) \qquad(k \to\infty). C p ( E k ( A )) ↑ C p ( A ) ( k → ∞ ) .
Indeed, E k = E k ∘ E k + 1 E_k = E_k \circ E_{k+1} E k = E k ∘ E k + 1 , so C p ( E k ( A ) ) ≤ C p ( E k + 1 ( A ) ) ≤ C p ( A ) C_p(E_k(A)) \leq C_p(E_{k+1}(A)) \leq C_p(A) C p ( E k ( A )) ≤ C p ( E k + 1 ( A )) ≤ C p ( A ) by Proposition 3.2.5 . Moreover E k ( A ) → A E_k(A) \to A E k ( A ) → A in ∥ ⋅ ∥ 2 \lVert\cdot\rVert_2 ∥ ⋅ ∥ 2 , all E k ( A ) E_k(A) E k ( A ) have spectrum in the interval [ m , M ] [m,M] [ m , M ] spanned by the spectrum of A A A , and for uniformly bounded self-adjoint X k → X X_k \to X X k → X in ∥ ⋅ ∥ 2 \lVert\cdot\rVert_2 ∥ ⋅ ∥ 2 one has τ ( g ( X k ) ) → τ ( g ( X ) ) \tau(g(X_k)) \to\tau(g(X)) τ ( g ( X k )) → τ ( g ( X )) for polynomials g g g , hence for continuous g g g on [ m , M ] [m,M] [ m , M ] ; with g ( t ) = t ± p g(t) = t^{\pm p} g ( t ) = t ± p this gives C p ( E k ( A ) ) → C p ( A ) C_p(E_k(A)) \to C_p(A) C p ( E k ( A )) → C p ( A ) . Since C p ( A ) = C p ( ∣ A ∣ ) C_p(A) = C_p(\lvert A\rvert) C p ( A ) = C p (∣ A ∣) for invertible A A A , every Schatten contrast C p C_p C p , p ≥ 1 p \geq1 p ≥ 1 , in R R R is a monotone limit of matrix contrasts, although no subadditivity estimate survives the limit.
Remark 3.4.4 (Symmetric operator spaces). The same argument works for a symmetric Banach function space E E E on ( 0 , 1 ) (0,1) ( 0 , 1 ) and the associated noncommutative space E ( M , τ ) E(\mathcal{M},\tau) E ( M , τ ) over a factor of type I I 1 \mathrm{II}_1 II 1 [48 ] . Recall that L ∞ ( 0 , 1 ) ⊂ E L^\infty(0,1) \subset E L ∞ ( 0 , 1 ) ⊂ E with ∥ x ∥ E ≤ ∥ 1 ∥ E ∥ x ∥ ∞ \lVert x\rVert_E \leq\lVert1\rVert_E\lVert x\rVert_\infty ∥ x ∥ E ≤ ∥ 1 ∥ E ∥ x ∥ ∞ . Let ϕ E ( t ) = ∥ e ∥ E \phi_E(t) = \lVert e\rVert_E ϕ E ( t ) = ∥ e ∥ E for a projection e e e with τ ( e ) = t \tau(e) = t τ ( e ) = t be the fundamental function, which is nondecreasing, and C E ( A ) = 1 2 ln ( ∥ A ∥ E ∥ A − 1 ∥ E / ϕ E ( 1 ) 2 ) C_E(A) = \frac{1}{2}\ln(\lVert A\rVert_E\lVert A^{-1}\rVert_E/\phi_E(1)^2) C E ( A ) = 2 1 ln (∥ A ∥ E ∥ A − 1 ∥ E / ϕ E ( 1 ) 2 ) . If X ≠ 0 X \ne0 X = 0 and 0 < η < ∥ X ∥ 0 < \eta< \lVert X\rVert 0 < η < ∥ X ∥ , the spectral projection e η e_\eta e η of ∣ X ∣ \lvert X\rvert ∣ X ∣ for ( ∥ X ∥ − η , ∥ X ∥ ] (\lVert X\rVert-\eta,\lVert X\rVert] (∥ X ∥ − η , ∥ X ∥] is nonzero and ∣ X ∣ ≥ ( ∥ X ∥ − η ) e η \lvert X\rvert\geq(\lVert X\rVert-\eta)e_\eta ∣ X ∣ ≥ (∥ X ∥ − η ) e η , so ∥ X ∥ E ≥ ( ∥ X ∥ − η ) ϕ E ( τ ( e η ) ) ≥ ( ∥ X ∥ − η ) ϕ E ( 0 + ) \lVert X\rVert_E \geq(\lVert X\rVert-\eta)\phi_E(\tau(e_\eta)) \geq(\lVert X\rVert-\eta)\phi_E(0+) ∥ X ∥ E ≥ (∥ X ∥ − η ) ϕ E ( τ ( e η )) ≥ (∥ X ∥ − η ) ϕ E ( 0 + ) . Hence ϕ E ( 0 + ) ∥ X ∥ ≤ ∥ X ∥ E ≤ ϕ E ( 1 ) ∥ X ∥ \phi_E(0+)\lVert X\rVert\leq\lVert X\rVert_E \leq\phi_E(1)\lVert X\rVert ϕ E ( 0 + ) ∥ X ∥ ≤ ∥ X ∥ E ≤ ϕ E ( 1 ) ∥ X ∥ , and together with ∥ X Y ∥ E ≤ ∥ X ∥ ∥ Y ∥ E \lVert XY\rVert_E \leq\lVert X\rVert\lVert Y\rVert_E ∥ X Y ∥ E ≤ ∥ X ∥ ∥ Y ∥ E this gives C E ( A B ) ≤ C E ( A ) + C E ( B ) + ln ϕ E ( 1 ) ϕ E ( 0 + ) C_E(AB) \leq C_E(A) + C_E(B) + \ln\frac{\phi_E(1)}{\phi_E(0+)} C E ( A B ) ≤ C E ( A ) + C E ( B ) + ln ϕ E ( 0 + ) ϕ E ( 1 ) . The constant is optimal: for the operators P t P_t P t above, ∥ P t ± k ∥ E = t ± k ϕ E ( δ ) ( 1 + o ( 1 ) ) \lVert P_t^{\pm k}\rVert_E = t^{\pm k}\phi_E(\delta)(1+o(1)) ∥ P t ± k ∥ E = t ± k ϕ E ( δ ) ( 1 + o ( 1 )) as t → ∞ t \to\infty t → ∞ , k = 1 , 2 k = 1,2 k = 1 , 2 , so C E ( P t 2 ) − 2 C E ( P t ) → ln ϕ E ( 1 ) ϕ E ( δ ) C_E(P_t^2) - 2C_E(P_t) \to\ln\frac{\phi_E(1)}{\phi_E(\delta)} C E ( P t 2 ) − 2 C E ( P t ) → ln ϕ E ( δ ) ϕ E ( 1 ) , and δ ↓ 0 \delta\downarrow0 δ ↓ 0 . Hence in type I I 1 \mathrm{II}_1 II 1 the defect is finite if and only if ϕ E ( 0 + ) > 0 \phi_E(0+) > 0 ϕ E ( 0 + ) > 0 , i.e. if and only if E = L ∞ E = L^\infty E = L ∞ with an equivalent norm, and it vanishes if and only if ϕ E \phi_E ϕ E is constant, i.e. ∥ ⋅ ∥ E \|\cdot\|_E ∥ ⋅ ∥ E is a multiple of the operator norm. Thus in a factor of type I I 1 \mathrm{II}_1 II 1 the operator norm is, up to equivalence, the only symmetric norm with finite defect. This is strictly more rigid than the matrix case: in M n ( C ) M_n(\mathbb{C}) M n ( C ) , n ≥ 2 n \ge2 n ≥ 2 , the same argument, whose inputs are now a rank-one projection e ≤ 1 { ∥ X ∥ } ( ∣ X ∣ ) e \le1_{\{\|X\|\}}(|X|) e ≤ 1 { ∥ X ∥ } ( ∣ X ∣ ) , so that ∣ X ∣ ≥ ∥ X ∥ e |X| \ge\|X\|e ∣ X ∣ ≥ ∥ X ∥ e , the monotonicity of a unitarily invariant norm ∥ ⋅ ∥ \lVert\cdot\rVert ∥ ⋅ ∥ on positive operators, and the inequality ∥ X Y ∥ ≤ ∥ X ∥ ∥ Y ∥ \lVert XY\rVert\le\|X\|\lVert Y\rVert ∥ X Y ∥ ≤ ∥ X ∥ ∥ Y ∥ [7 Ch. IV] , shows that ∥ ⋅ ∥ \lVert\cdot\rVert ∥ ⋅ ∥ , with the contrast 1 2 ln ( ∥ A ∥ ∥ A − 1 ∥ / ∥ 1 ∥ 2 ) \frac{1}{2}\ln(\lVert A\rVert\lVert A^{-1}\rVert/\lVert1\rVert^2) 2 1 ln (∥ A ∥ ∥ A − 1 ∥ / ∥ 1 ∥ 2 ) , has defect exactly ln ∥ 1 ∥ ∥ e ∥ \ln\frac{\lVert1\rVert}{\lVert e\rVert} ln ∥ e ∥ ∥ 1 ∥ , e e e a rank-one projection. This is at most ln n \ln n ln n , since ∥ 1 ∥ ≤ n ∥ e ∥ \lVert1\rVert\le n\lVert e\rVert ∥ 1 ∥ ≤ n ∥ e ∥ , and it vanishes if and only if ∥ 1 ∥ = ∥ e ∥ \lVert1\rVert= \lVert e\rVert ∥ 1 ∥ = ∥ e ∥ , i.e. (as ∥ e ∥ ∥ X ∥ ≤ ∥ X ∥ ≤ ∥ 1 ∥ ∥ X ∥ \lVert e\rVert\|X\| \le\lVert X\rVert\le\lVert1\rVert\|X\| ∥ e ∥ ∥ X ∥ ≤ ∥ X ∥ ≤ ∥ 1 ∥ ∥ X ∥ ) if and only if ∥ ⋅ ∥ \lVert\cdot\rVert ∥ ⋅ ∥ is a multiple of the operator norm. So in the matrix case every unitarily invariant norm has finite defect, and only the vanishing of the defect singles out the operator norm.
The rate of convergence to OLC and the dimension range
Theorem 3.5.1. Let ( M , τ ) (\mathcal{M},\tau) ( M , τ ) be a von Neumann algebra with a faithful normal tracial state.
(a) Let A ∈ GL ( M ) A \in\operatorname{GL}(\mathcal{M}) A ∈ GL ( M ) with ∣ A ∣ |A| ∣ A ∣ not scalar, let a < b a < b a < b be the minimum and maximum of the spectrum of ln ∣ A ∣ \ln|A| ln ∣ A ∣ , and e + = 1 { b } ( ln ∣ A ∣ ) e_+ = 1_{\{b\}}(\ln|A|) e + = 1 { b } ( ln ∣ A ∣ ) , e − = 1 { a } ( ln ∣ A ∣ ) e_- = 1_{\{a\}}(\ln|A|) e − = 1 { a } ( ln ∣ A ∣ ) . Then
lim p → ∞ 2 p ( OLC ( A ) − C p ( A ) ) = − ln τ ( e + ) − ln τ ( e − ) ∈ [ 2 ln 2 , ∞ ] , \lim_{p\to\infty} 2p\bigl(\operatorname{OLC}(A)-C_p(A)\bigr)=-\ln\tau(e_+)-\ln\tau(e_-)\in[2\ln2,\infty], p → ∞ lim 2 p ( OLC ( A ) − C p ( A ) ) = − ln τ ( e + ) − ln τ ( e − ) ∈ [ 2 ln 2 , ∞ ] , with the convention − ln 0 = ∞ -\ln0=\infty − ln 0 = ∞ .
(b) OLC ( A ) − C p ( A ) = O ( 1 / p ) \operatorname{OLC}(A)-C_p(A)=O(1/p) OLC ( A ) − C p ( A ) = O ( 1/ p ) as p → ∞ p\to\infty p → ∞ for every A ∈ GL ( M ) A\in\operatorname{GL}(\mathcal{M}) A ∈ GL ( M ) if and only if M \mathcal{M} M is finite-dimensional.
(c) If M ≅ M n ( C ) \mathcal{M}\cong M_n(\mathbb{C}) M ≅ M n ( C ) , the limits in (a) form the finite set { ln n 2 j k : j , k ≥ 1 , j + k ≤ n } \left\{\ln\frac{n^2}{jk}:j,k\ge1,\ j+k\le n\right\} { ln j k n 2 : j , k ≥ 1 , j + k ≤ n } . If M \mathcal{M} M is a factor of type I I 1 \mathrm{II}_1 II 1 , they fill the whole interval [ 2 ln 2 , ∞ ) [2\ln2,\infty) [ 2 ln 2 , ∞ ) .
Proof. (a) We have L A ( p ) − p b = ln ∫ e − p ( b − x ) d ν A ( x ) L_A(p)-pb=\ln\int e^{-p(b-x)}\,\mathrm{d}\nu_A(x) L A ( p ) − p b = ln ∫ e − p ( b − x ) d ν A ( x ) , and the integrand decreases to 1 { b } 1_{\{b\}} 1 { b } as p → ∞ p\to\infty p → ∞ , so L A ( p ) − p b → ln ν A ( { b } ) = ln τ ( e + ) L_A(p)-pb\to\ln\nu_A(\{b\})=\ln\tau(e_+) L A ( p ) − p b → ln ν A ({ b }) = ln τ ( e + ) by monotone convergence. Similarly L A ( − p ) + p a → ln τ ( e − ) L_A(-p)+pa\to\ln\tau(e_-) L A ( − p ) + p a → ln τ ( e − ) . Since 2 p OLC ( A ) = p ( b − a ) 2p\operatorname{OLC}(A)=p(b-a) 2 p OLC ( A ) = p ( b − a ) , the formula follows. As e + ⊥ e − e_+\perp e_- e + ⊥ e − , τ ( e + ) + τ ( e − ) ≤ 1 \tau(e_+)+\tau(e_-)\le1 τ ( e + ) + τ ( e − ) ≤ 1 and τ ( e + ) τ ( e − ) ≤ 1 4 \tau(e_+)\tau(e_-)\le\frac14 τ ( e + ) τ ( e − ) ≤ 4 1 .
(b) If M \mathcal{M} M is finite-dimensional, ln ∣ A ∣ \ln|A| ln ∣ A ∣ has finite spectrum, e ± ≠ 0 e_\pm\ne0 e ± = 0 , and the limit in (a) is finite. If M \mathcal{M} M is infinite-dimensional, it contains a sequence ( q k ) k ≥ 1 (q_k)_{k\ge1} ( q k ) k ≥ 1 of mutually orthogonal nonzero projections. Let H = ∑ k ≥ 1 ( 1 − 1 k ) q k H=\sum_{k\ge1}(1-\frac1k)q_k H = ∑ k ≥ 1 ( 1 − k 1 ) q k and A = e H ∈ GL ( M ) A=e^H\in\operatorname{GL}(\mathcal{M}) A = e H ∈ GL ( M ) . The spectrum of H H H has maximum 1 1 1 and minimum 0 0 0 , but 1 { 1 } ( H ) = 0 1_{\{1\}}(H)=0 1 { 1 } ( H ) = 0 , so the limit in (a) is + ∞ +\infty + ∞ .
(c) In M n ( C ) M_n(\mathbb{C}) M n ( C ) , τ ( e + ) = j n \tau(e_+)=\frac{j}{n} τ ( e + ) = n j and τ ( e − ) = k n \tau(e_-)=\frac{k}{n} τ ( e − ) = n k with j , k ≥ 1 j,k\ge1 j , k ≥ 1 , j + k ≤ n j+k\le n j + k ≤ n , and all such pairs occur. In a factor of type I I 1 \mathrm{II}_1 II 1 , for any s , t > 0 s,t>0 s , t > 0 with s + t ≤ 1 s+t\le1 s + t ≤ 1 there are orthogonal projections e , f e,f e , f with τ ( e ) = s \tau(e)=s τ ( e ) = s , τ ( f ) = t \tau(f)=t τ ( f ) = t , and A = exp ( e − f ) A=\exp(e-f) A = exp ( e − f ) has limit − ln ( s t ) -\ln(st) − ln ( s t ) ; these values fill [ 2 ln 2 , ∞ ) [2\ln2,\infty) [ 2 ln 2 , ∞ ) , and + ∞ +\infty + ∞ occurs by (b). □ \square □
Part (c) shows that the first-order correction to OLC reads off the range of the trace on projections, which is the classical invariant distinguishing type I n I_n I n from type I I 1 \mathrm{II}_1 II 1 . When an extreme spectral projection vanishes, the correction is of larger order, determined by the behaviour of ν A \nu_A ν A at the edges.
Proposition 3.5.2. Let A ∈ GL ( M ) A\in\operatorname{GL}(\mathcal{M}) A ∈ GL ( M ) and let a < b a<b a < b be the extreme points of supp ν A \operatorname{supp}\nu_A supp ν A . Suppose that, as u ↓ 0 u\downarrow0 u ↓ 0 ,
ν A ( ( b − u , b ] ) = c + u β + ( 1 + o ( 1 ) ) , ν A ( [ a , a + u ) ) = c − u β − ( 1 + o ( 1 ) ) , \nu_A((b-u,b])=c_+u^{\beta_+}(1+o(1)),\qquad\nu_A([a,a+u))=c_-u^{\beta_-}(1+o(1)), ν A (( b − u , b ]) = c + u β + ( 1 + o ( 1 )) , ν A ([ a , a + u )) = c − u β − ( 1 + o ( 1 )) , with c ± > 0 c_\pm>0 c ± > 0 and β ± > 0 \beta_\pm>0 β ± > 0 . Then, as p → ∞ p\to\infty p → ∞ ,
C p ( A ) = OLC ( A ) − β + + β − 2 ln p p + 1 2 p ln ( c + c − Γ ( 1 + β + ) Γ ( 1 + β − ) ) + o ( 1 p ) . C_p(A)=\operatorname{OLC}(A)-\frac{\beta_++\beta_-}{2}\frac{\ln p}{p}+\frac{1}{2p}\ln\bigl(c_+c_-\Gamma(1+\beta_+)\Gamma(1+\beta_-)\bigr)+o\left(\frac1p\right). C p ( A ) = OLC ( A ) − 2 β + + β − p ln p + 2 p 1 ln ( c + c − Γ ( 1 + β + ) Γ ( 1 + β − ) ) + o ( p 1 ) .
Proof. The hypothesis forces ν A ( { b } ) = ν A ( { a } ) = 0 \nu_A(\{b\})=\nu_A(\{a\})=0 ν A ({ b }) = ν A ({ a }) = 0 . Let F ( u ) = ν A ( ( b − u , b ] ) F(u)=\nu_A((b-u,b]) F ( u ) = ν A (( b − u , b ]) , so that F ( 0 ) = 0 F(0)=0 F ( 0 ) = 0 . Integrating by parts and substituting v = p u v=pu v = p u ,
e L A ( p ) − p b = ∫ [ 0 , ∞ ) e − p u d F ( u ) = ∫ 0 ∞ e − v F ( v / p ) d v . e^{L_A(p)-pb}=\int_{[0,\infty)}e^{-pu}\,dF(u)=\int_0^\infty e^{-v}F(v/p)\,dv. e L A ( p ) − p b = ∫ [ 0 , ∞ ) e − p u d F ( u ) = ∫ 0 ∞ e − v F ( v / p ) d v .
Choose u 0 > 0 u_0>0 u 0 > 0 with F ( u ) ≤ 2 c + u β + F(u)\le2c_+u^{\beta_+} F ( u ) ≤ 2 c + u β + for u ≤ u 0 u\le u_0 u ≤ u 0 ; since F ≤ 1 F\le1 F ≤ 1 , F ( u ) ≤ C u β + F(u)\le Cu^{\beta_+} F ( u ) ≤ C u β + for all u > 0 u>0 u > 0 with C = max { 2 c + , u 0 − β + } C=\max\{2c_+,u_0^{-\beta_+}\} C = max { 2 c + , u 0 − β + } . Hence p β + F ( v / p ) ≤ C v β + p^{\beta_+}F(v/p)\le Cv^{\beta_+} p β + F ( v / p ) ≤ C v β + , and by dominated convergence p β + e L A ( p ) − p b → c + ∫ 0 ∞ e − v v β + d v = c + Γ ( 1 + β + ) p^{\beta_+}e^{L_A(p)-pb}\to c_+\int_0^\infty e^{-v}v^{\beta_+}\,dv=c_+\Gamma(1+\beta_+) p β + e L A ( p ) − p b → c + ∫ 0 ∞ e − v v β + d v = c + Γ ( 1 + β + ) . The lower edge is treated in the same way, and 2 p C p ( A ) = L A ( p ) + L A ( − p ) 2pC_p(A)=L_A(p)+L_A(-p) 2 p C p ( A ) = L A ( p ) + L A ( − p ) , 2 p OLC ( A ) = p ( b − a ) 2p\operatorname{OLC}(A)=p(b-a) 2 p OLC ( A ) = p ( b − a ) .
Example 3.5.3 (Exponential of a semicircular element). Let s s s be a semicircular element of variance 1 1 1 in a factor of type I I 1 \mathrm{II}_1 II 1 [49 ] , with density 1 2 π 4 − x 2 \frac{1}{2\pi}\sqrt{4-x^2} 2 π 1 4 − x 2 on [ − 2 , 2 ] [-2,2] [ − 2 , 2 ] , and A = e s A=e^s A = e s . Then τ ( A p ) = τ ( A − p ) = I 1 ( 2 p ) / p \tau(A^p)=\tau(A^{-p})=I_1(2p)/p τ ( A p ) = τ ( A − p ) = I 1 ( 2 p ) / p , where I 1 I_1 I 1 is the modified Bessel function, and
C p ( A ) = 1 p ln I 1 ( 2 p ) p = 2 − 3 2 ln p p − ln ( 2 π ) p + o ( 1 p ) , OLC ( A ) = 2. C_p(A)=\frac{1}{p}\ln\frac{I_1(2p)}{p}=2-\frac{3}{2}\frac{\ln p}{p}-\frac{\ln(2\sqrt{\pi})}{p}+o\left(\frac{1}{p}\right),\qquad\operatorname{OLC}(A)=2. C p ( A ) = p 1 ln p I 1 ( 2 p ) = 2 − 2 3 p ln p − p ln ( 2 π ) + o ( p 1 ) , OLC ( A ) = 2.
in agreement with Proposition 3.5.2 , where β ± = 3 2 \beta_\pm=\frac{3}{2} β ± = 2 3 and c ± = 2 3 π c_\pm=\frac{2}{3\pi} c ± = 3 π 2 . No matrix has this behaviour: by Theorem 3.5.1(a),(b), in M n ( C ) M_n(\mathbb{C}) M n ( C ) the quantity 2 p ( OLC − C p ) 2p(\operatorname{OLC}-C_p) 2 p ( OLC − C p ) has a finite limit, so the correction is of order 1 p \frac{1}{p} p 1 .
Example 3.5.4 (Haar unitaries). Let u u u be a Haar unitary in ( M , τ ) (\mathcal{M},\tau) ( M , τ ) , i.e. τ ( u k ) = 0 \tau(u^k)=0 τ ( u k ) = 0 for k ≠ 0 k\ne0 k = 0 ; for instance a canonical generator of a free group factor L ( F r ) L(\mathbb{F}_r) L ( F r ) , or of L ( Z ) L(\mathbb{Z}) L ( Z ) . For 0 < c < 1 0<c<1 0 < c < 1 let A = 1 + c 2 ( u + u ∗ ) A=1+\frac{c}{2}(u+u^*) A = 1 + 2 c ( u + u ∗ ) . Then 1 2 ( u + u ∗ ) \frac{1}{2}(u+u^*) 2 1 ( u + u ∗ ) has the arcsine distribution on [ − 1 , 1 ] [-1,1] [ − 1 , 1 ] , A A A is positive and invertible, and
τ ( A ) = 1 , τ ( A − 1 ) = 1 1 − c 2 , \tau(A)=1,\qquad\tau(A^{-1})=\frac{1}{\sqrt{1-c^2}}, τ ( A ) = 1 , τ ( A − 1 ) = 1 − c 2 1 ,
C 1 ( A ) = − 1 4 ln ( 1 − c 2 ) , OLC ( A ) = 1 2 ln 1 + c 1 − c . C_1(A)=-\frac{1}{4}\ln(1-c^2),\qquad\operatorname{OLC}(A)=\frac{1}{2}\ln\frac{1+c}{1-c}. C 1 ( A ) = − 4 1 ln ( 1 − c 2 ) , OLC ( A ) = 2 1 ln 1 − c 1 + c .
The arcsine density has inverse square-root singularities at both edges, so Proposition 3.5.2 applies with β ± = 1 2 \beta_\pm=\frac{1}{2} β ± = 2 1 and c ± = 1 π 2 ( 1 ± c ) / c c_\pm=\frac{1}{\pi}\sqrt{2(1\pm c)/c} c ± = π 1 2 ( 1 ± c ) / c :
C p ( A ) = OLC ( A ) − 1 2 ln p p + 1 2 p ln 1 − c 2 2 π c + o ( 1 p ) . C_p(A)=\operatorname{OLC}(A)-\frac{1}{2}\frac{\ln p}{p}+\frac{1}{2p}\ln\frac{\sqrt{1-c^2}}{2\pi c}+o\left(\frac{1}{p}\right). C p ( A ) = OLC ( A ) − 2 1 p ln p + 2 p 1 ln 2 π c 1 − c 2 + o ( p 1 ) .
Together with Example 3.5.3 (β ± = 3 2 \beta_\pm=\frac{3}{2} β ± = 2 3 ) this shows that the coefficient of ln p p \frac{\ln p}{p} p l n p records the edge exponents of the spectral measure, here those of the two basic laws of free probability.
The second derivative of the profile: the gap at the edge
Theorem 3.5.1 concerns the first-order behaviour of the profile L A L_A L A at infinity, which sees the mass of the spectral measure at the edge. The second derivative L A ′ ′ ( t ) = Var ( ν A ( t ) ) L_A''(t)=\operatorname{Var}(\nu_A^{(t)}) L A ′′ ( t ) = Var ( ν A ( t ) ) sees whether the edge is isolated . The following statement consists of standard Laplace-type asymptotics [50 ] and is not new. We include it because together with Theorem 3.5.1 it gives the complete picture of the upper edge.
Proposition 3.5.5 (Edge exponent; classical). Let A ∈ GL ( M ) A\in\operatorname{GL}(\mathcal{M}) A ∈ GL ( M ) with ∣ A ∣ |A| ∣ A ∣ not scalar, let [ a , b ] [a,b] [ a , b ] be the smallest interval containing the spectrum of ln ∣ A ∣ \ln|A| ln ∣ A ∣ , and e + = 1 { b } ( ln ∣ A ∣ ) e_+=1_{\{b\}}(\ln|A|) e + = 1 { b } ( ln ∣ A ∣ ) .
(a) If b b b is an isolated point of the spectrum of ln ∣ A ∣ \ln|A| ln ∣ A ∣ , at distance δ > 0 \delta>0 δ > 0 from the rest of the spectrum, then
L A ′ ′ ( t ) = 1 + o ( 1 ) τ ( e + ) ∫ { x < b } ( b − x ) 2 e − t ( b − x ) d ν A ( x ) and − lim t → ∞ 1 t ln L A ′ ′ ( t ) = δ . L_A''(t)=\frac{1+o(1)}{\tau(e_+)}\int_{\{x<b\}}(b-x)^2e^{-t(b-x)}\,d\nu_A(x)\qquad\text{and}\qquad-\lim_{t\to\infty}\frac{1}{t}\ln L_A''(t)=\delta. L A ′′ ( t ) = τ ( e + ) 1 + o ( 1 ) ∫ { x < b } ( b − x ) 2 e − t ( b − x ) d ν A ( x ) and − t → ∞ lim t 1 ln L A ′′ ( t ) = δ .
(b) If b b b is not an isolated point of the spectrum, then − lim t → ∞ 1 t ln L A ′ ′ ( t ) = 0 -\lim_{t\to\infty}\frac{1}{t}\ln L_A^{\prime\prime}(t)=0 − lim t → ∞ t 1 ln L A ′′ ( t ) = 0 .
(c) If ν A ( ( b − u , b ] ) = c u β ( 1 + o ( 1 ) ) \nu_A((b-u,b])=cu^\beta(1+o(1)) ν A (( b − u , b ]) = c u β ( 1 + o ( 1 )) as u ↓ 0 u\downarrow0 u ↓ 0 with c , β > 0 c,\beta>0 c , β > 0 , then L A ′ ′ ( t ) = β t − 2 ( 1 + o ( 1 ) ) L_A^{\prime\prime}(t)=\beta t^{-2}(1+o(1)) L A ′′ ( t ) = β t − 2 ( 1 + o ( 1 )) .
In particular, L A ′ ′ L_A^{\prime\prime} L A ′′ decays exponentially if and only if the top of the spectrum is isolated, and then the rate of decay is the gap.
Proof. Write ν = ν A \nu=\nu_A ν = ν A and M k ( t ) = ∫ { x < b } ( b − x ) k e − t ( b − x ) d ν ( x ) M_k(t)=\int_{\{x<b\}}(b-x)^k e^{-t(b-x)}\,d\nu(x) M k ( t ) = ∫ { x < b } ( b − x ) k e − t ( b − x ) d ν ( x ) for k = 0 , 1 , 2 k=0,1,2 k = 0 , 1 , 2 . Since the variance is invariant under x ↦ b − x x\mapsto b-x x ↦ b − x ,
L A ′ ′ ( t ) = M 2 ( t ) τ ( e + ) + M 0 ( t ) − ( M 1 ( t ) τ ( e + ) + M 0 ( t ) ) 2 . (4) L_A^{\prime\prime}(t)=\frac{M_2(t)}{\tau(e_+)+M_0(t)}-\left(\frac{M_1(t)}{\tau(e_+)+M_0(t)}\right)^2. \tag*{(4)} L A ′′ ( t ) = τ ( e + ) + M 0 ( t ) M 2 ( t ) − ( τ ( e + ) + M 0 ( t ) M 1 ( t ) ) 2 . (4)
(a) Here τ ( e + ) > 0 \tau(e_+)>0 τ ( e + ) > 0 and δ ≤ b − x ≤ b − a \delta\le b-x\le b-a δ ≤ b − x ≤ b − a on the support of the restriction of ν \nu ν to { x < b } \{x<b\} { x < b } . Hence M 0 ( t ) ≤ e − δ t M_0(t)\le e^{-\delta t} M 0 ( t ) ≤ e − δ t , and by the Cauchy–Schwarz inequality M 1 ( t ) 2 ≤ M 2 ( t ) M 0 ( t ) ≤ M 2 ( t ) e − δ t M_1(t)^2\le M_2(t)M_0(t)\le M_2(t)e^{-\delta t} M 1 ( t ) 2 ≤ M 2 ( t ) M 0 ( t ) ≤ M 2 ( t ) e − δ t , so the second term in (4) is o ( M 2 ( t ) ) o(M_2(t)) o ( M 2 ( t )) and L A ′ ′ ( t ) = M 2 ( t ) ( 1 + o ( 1 ) ) / τ ( e + ) L_A^{\prime\prime}(t)=M_2(t)(1+o(1))/\tau(e_+) L A ′′ ( t ) = M 2 ( t ) ( 1 + o ( 1 )) / τ ( e + ) . Moreover M 2 ( t ) ≤ ( b − a ) 2 e − δ t M_2(t)\le(b-a)^2e^{-\delta t} M 2 ( t ) ≤ ( b − a ) 2 e − δ t , and for every ε > 0 \varepsilon>0 ε > 0 the set { b − δ − ε < x ≤ b − δ } \{b-\delta-\varepsilon<x\le b-\delta\} { b − δ − ε < x ≤ b − δ } has positive measure η \eta η , so M 2 ( t ) ≥ δ 2 η e − ( δ + ε ) t M_2(t)\ge\delta^2\eta e^{-(\delta+\varepsilon)t} M 2 ( t ) ≥ δ 2 η e − ( δ + ε ) t . Thus 1 t ln M 2 ( t ) → − δ \frac{1}{t}\ln M_2(t)\to-\delta t 1 ln M 2 ( t ) → − δ .
(b) Since ν ( t ) \nu^{(t)} ν ( t ) is supported in [ a , b ] [a,b] [ a , b ] , L A ′ ′ ( t ) ≤ 1 4 ( b − a ) 2 L_A^{\prime\prime}(t)\le\frac{1}{4}(b-a)^2 L A ′′ ( t ) ≤ 4 1 ( b − a ) 2 , so lim inf ( − 1 t ln L A ′ ′ ( t ) ) ≥ 0 \liminf(-\frac{1}{t}\ln L_A^{\prime\prime}(t))\ge0 lim inf ( − t 1 ln L A ′′ ( t )) ≥ 0 . For the converse let ε > 0 \varepsilon>0 ε > 0 . As b b b is not isolated, there are two distinct points of the spectrum in ( b − ε , b ] (b-\varepsilon,b] ( b − ε , b ] , hence disjoint intervals U , V ⊂ ( b − ε , b ] U,V\subset(b-\varepsilon,b] U , V ⊂ ( b − ε , b ] at positive distance d d d with ν ( U ) , ν ( V ) > 0 \nu(U),\nu(V)>0 ν ( U ) , ν ( V ) > 0 . Since ∫ e t x d ν ≤ e t b \int e^{tx}\,d\nu\le e^{tb} ∫ e t x d ν ≤ e t b , ν ( t ) ( U ) ≥ ν ( U ) e − ε t \nu^{(t)}(U)\ge\nu(U)e^{-\varepsilon t} ν ( t ) ( U ) ≥ ν ( U ) e − εt , and likewise for V V V . Therefore
L A ′ ′ ( t ) = 1 2 ∬ ( x − y ) 2 d ν ( t ) ( x ) d ν ( t ) ( y ) ≥ d 2 ν ( U ) ν ( V ) e − 2 ε t L_A^{\prime\prime}(t)=\frac{1}{2}\iint(x-y)^2\,d\nu^{(t)}(x)d\nu^{(t)}(y)\ge d^2\nu(U)\nu(V)e^{-2\varepsilon t} L A ′′ ( t ) = 2 1 ∬ ( x − y ) 2 d ν ( t ) ( x ) d ν ( t ) ( y ) ≥ d 2 ν ( U ) ν ( V ) e − 2 εt
and lim sup ( − 1 t ln L A ′ ′ ( t ) ) ≤ 2 ε \limsup(-\frac{1}{t}\ln L_A^{\prime\prime}(t))\le2\varepsilon lim sup ( − t 1 ln L A ′′ ( t )) ≤ 2 ε .
(c) Here τ ( e + ) = 0 \tau(e_+)=0 τ ( e + ) = 0 . Let F ( u ) = ν ( ( b − u , b ] ) F(u)=\nu((b-u,b]) F ( u ) = ν (( b − u , b ]) and G k ( u ) = ∫ [ 0 , u ] v k d F ( v ) = u k F ( u ) − k ∫ 0 u v k − 1 F ( v ) d v G_k(u)=\int_{[0,u]}v^k\,dF(v)=u^kF(u)-k\int_0^u v^{k-1}F(v)\,dv G k ( u ) = ∫ [ 0 , u ] v k d F ( v ) = u k F ( u ) − k ∫ 0 u v k − 1 F ( v ) d v , so that G k ( u ) = c β β + k u β + k ( 1 + o ( 1 ) ) G_k(u)=\frac{c\beta}{\beta+k}u^{\beta+k}(1+o(1)) G k ( u ) = β + k c β u β + k ( 1 + o ( 1 )) . The Abelian argument in the proof of Proposition 3.5.2 , applied to G k G_k G k , gives M k ( t ) = c β Γ ( β + k ) t − ( β + k ) ( 1 + o ( 1 ) ) M_k(t)=c\beta\Gamma(\beta+k)t^{-(\beta+k)}(1+o(1)) M k ( t ) = c β Γ ( β + k ) t − ( β + k ) ( 1 + o ( 1 )) . Hence M 1 / M 0 ∼ β / t M_1/M_0\sim\beta/t M 1 / M 0 ∼ β / t and M 2 / M 0 ∼ β ( β + 1 ) / t 2 M_2/M_0\sim\beta(\beta+1)/t^2 M 2 / M 0 ∼ β ( β + 1 ) / t 2 , and (4) gives L A ′ ′ ( t ) ∼ β / t 2 L_A^{\prime\prime}(t)\sim\beta/t^2 L A ′′ ( t ) ∼ β / t 2 .
Combining Theorem 3.5.1 and Proposition 3.5.5 , there are three regimes at the top of the spectrum of ∣ A ∣ |A| ∣ A ∣ , read off from the top-edge term L A ( p ) − p b L_A(p)-pb L A ( p ) − p b of 2 p ( O L C − C p ) = ( p b − L A ( p ) ) + ( − p a − L A ( − p ) ) 2p(\mathrm{OLC}-C_p)=(pb-L_A(p))+(-pa-L_A(-p)) 2 p ( OLC − C p ) = ( p b − L A ( p )) + ( − p a − L A ( − p )) . If the top is an isolated eigenvalue, then L A ( p ) − p b L_A(p)-pb L A ( p ) − p b has the finite limit ln τ ( e + ) \ln\tau(e_+) ln τ ( e + ) and L A ′ ′ L_A^{\prime\prime} L A ′′ decays exponentially at the rate given by the gap. If it is an eigenvalue which is not isolated, the limit is still finite but L A ′ ′ L_A^{\prime\prime} L A ′′ decays subexponentially. If it is not an eigenvalue, then L A ( p ) − p b → − ∞ L_A(p)-pb\to-\infty L A ( p ) − p b → − ∞ , and for a regular edge with exponent β \beta β one has L A ( p ) − p b = − β ln p + O ( 1 ) L_A(p)-pb=-\beta\ln p+O(1) L A ( p ) − p b = − β ln p + O ( 1 ) , a contribution β 2 p \frac{\beta}{2p} 2 p β to O L C − C p \mathrm{OLC}-C_p OLC − C p , and L A ′ ′ ( t ) ∼ β / t 2 L_A^{\prime\prime}(t)\sim\beta/t^2 L A ′′ ( t ) ∼ β / t 2 ; for instance L A ′ ′ ( t ) ∼ 1 2 t 2 L_A^{\prime\prime}(t)\sim\frac{1}{2t^2} L A ′′ ( t ) ∼ 2 t 2 1 for the exponential of a semicircular element and 1 2 t 2 \frac{1}{2t^2} 2 t 2 1 for 1 + c 2 ( u + u ∗ ) 1+\frac{c}{2}(u+u^*) 1 + 2 c ( u + u ∗ ) with a Haar unitary u u u . The same applies to the bottom edge. Only the first regime occurs in finite dimension; the second requires an eigenvalue that is a limit of other spectral points, and the third a spectral edge without an eigenvalue.
Araki–Woods factors and sequences of contrasts
The invariants of § 6.7 and § 6.8 depend only on the range of the trace on projections and say nothing about factors of type III. For hyperfinite factors the natural object is not the contrast of one operator but the sequence of contrasts of the component states of an infinite tensor product.
Infinite tensor products of matrix algebras
Let n k ≥ 2 n_k\ge2 n k ≥ 2 be integers and ρ k ∈ M n k ( C ) \rho_k\in M_{n_k}(\mathbb{C}) ρ k ∈ M n k ( C ) positive definite density matrices, with states φ k = Tr ( ρ k ⋅ ) \varphi_k=\operatorname{Tr}(\rho_k\cdot) φ k = Tr ( ρ k ⋅ ) . On the algebraic infinite tensor product A = ⨂ k ≥ 1 M n k ( C ) \mathcal{A}=\bigotimes_{k\ge1}M_{n_k}(\mathbb{C}) A = ⨂ k ≥ 1 M n k ( C ) , the union of the finite tensor products A N = M n 1 ( C ) ⊗ ⋯ ⊗ M n N ( C ) \mathcal{A}_N=M_{n_1}(\mathbb{C})\otimes\cdots\otimes M_{n_N}(\mathbb{C}) A N = M n 1 ( C ) ⊗ ⋯ ⊗ M n N ( C ) , consider the product state φ = ⨂ k φ k \varphi=\bigotimes_k\varphi_k φ = ⨂ k φ k and its GNS representation π φ \pi_\varphi π φ . The von Neumann algebra
M = ⨂ k ≥ 1 ( M n k ( C ) , ρ k ) : = π φ ( A ) ′ ′ \mathcal{M}=\bigotimes_{k\geq1}(M_{n_k}(\mathbb{C}),\rho_k):=\pi_\varphi(\mathcal{A})^{\prime\prime} M = k ≥ 1 ⨂ ( M n k ( C ) , ρ k ) := π φ ( A ) ′′
is a hyperfinite factor, called an Araki–Woods or ITPFI factor [39 ] ; it is called ITPFI2 _2 2 if all n k = 2 n_k=2 n k = 2 . The cyclic vector defines a faithful normal state on M \mathcal{M} M whose modular group is ⨂ k Ad ρ k i t \bigotimes_k\operatorname{Ad}\rho_k^{it} ⨂ k Ad ρ k i t [40 ] . For tracial components ρ k = 1 n k \rho_k=\frac{1}{n_k} ρ k = n k 1 one obtains the hyperfinite factor R R R of type I I 1 \mathrm{II}_1 II 1 , and for ρ k = 1 1 + λ diag ( 1 , λ ) \rho_k=\frac{1}{1+\lambda}\operatorname{diag}(1,\lambda) ρ k = 1 + λ 1 diag ( 1 , λ ) , 0 < λ < 1 0<\lambda<1 0 < λ < 1 , the Powers factor R λ R_\lambda R λ [38 ] . The restriction of φ \varphi φ to A N \mathcal{A}_N A N has the density matrix
ρ ( N ) = ρ 1 ⊗ ⋯ ⊗ ρ N ∈ M d N ( C ) , d N = n 1 ⋯ n N . \rho^{(N)}=\rho_1\otimes\cdots\otimes\rho_N\in M_{d_N}(\mathbb{C}),\qquad d_N=n_1\cdots n_N. ρ ( N ) = ρ 1 ⊗ ⋯ ⊗ ρ N ∈ M d N ( C ) , d N = n 1 ⋯ n N .
For a positive definite matrix ρ \rho ρ we write ν ρ \nu_\rho ν ρ for its logarithmic spectral measure with respect to the normalised trace, L ρ L_\rho L ρ for its cumulant generating function and V ( ρ ) = Var ( ν ρ ) V(\rho)=\operatorname{Var}(\nu_\rho) V ( ρ ) = Var ( ν ρ ) for the variance of its logarithmic spectrum.
Lemma 3.6.1. For every N N N and every p ∈ ( 0 , ∞ ) p\in(0,\infty) p ∈ ( 0 , ∞ ) ,
OLC ( ρ ( N ) ) = ∑ k ≤ N OLC ( ρ k ) , C p ( ρ ( N ) ) = ∑ k ≤ N C p ( ρ k ) , \operatorname{OLC}(\rho^{(N)})=\sum_{k\leq N}\operatorname{OLC}(\rho_k),\qquad C_p(\rho^{(N)})=\sum_{k\leq N}C_p(\rho_k), OLC ( ρ ( N ) ) = k ≤ N ∑ OLC ( ρ k ) , C p ( ρ ( N ) ) = k ≤ N ∑ C p ( ρ k ) ,
V ( ρ ( N ) ) = ∑ k ≤ N V ( ρ k ) . V(\rho^{(N)})=\sum_{k\leq N}V(\rho_k). V ( ρ ( N ) ) = k ≤ N ∑ V ( ρ k ) .
Proof. The eigenvalues of a tensor product are the products of the eigenvalues, so ν ρ ( N ) = ν ρ 1 ∗ ⋯ ∗ ν ρ N \nu_{\rho^{(N)}}=\nu_{\rho_1}*\cdots*\nu_{\rho_N} ν ρ ( N ) = ν ρ 1 ∗ ⋯ ∗ ν ρ N . Supports and variances add under convolution, and C p C_p C p is additive by Proposition 3.2.3 . (For OLC this is also (Nikolaeva and Novikov, arXiv:2609.06997v2, Prop. 4.7).) □
Thus all contrasts of the partial products are nondecreasing in N N N , and the question is which of them stay bounded and how fast the others grow. For the Powers factor with OLC ( ρ k ) = z \operatorname{OLC}(\rho_k)=z OLC ( ρ k ) = z one finds OLC ( ρ ( N ) ) = N z \operatorname{OLC}(\rho^{(N)})=Nz OLC ( ρ ( N ) ) = N z and C p ( ρ ( N ) ) = N p ln cosh ( p z ) C_p(\rho^{(N)})=\frac{N}{p}\ln\cosh(pz) C p ( ρ ( N ) ) = p N ln cosh ( p z ) : all contrasts grow linearly in N N N , that is, logarithmically in the dimension d N = 2 N d_N=2^N d N = 2 N .
Components of size two: the Araki–Woods criteria in contrast form
Let all n k = 2 n_k=2 n k = 2 and put
z n = OLC ( ρ n ) , a n = C 1 ( ρ n ) , b n = C 1 ( ρ n ) − OLC ( ρ n ) + ln 2. z_n=\operatorname{OLC}(\rho_n),\qquad a_n=C_1(\rho_n),\qquad b_n=C_1(\rho_n)-\operatorname{OLC}(\rho_n)+\ln2. z n = OLC ( ρ n ) , a n = C 1 ( ρ n ) , b n = C 1 ( ρ n ) − OLC ( ρ n ) + ln 2.
For a positive definite 2 × 2 2\times2 2 × 2 density matrix ρ \rho ρ one has C 1 ( ρ ) = ln cosh OLC ( ρ ) C_1(\rho)=\ln\cosh\operatorname{OLC}(\rho) C 1 ( ρ ) = ln cosh OLC ( ρ ) , and therefore
max { 0 , OLC ( ρ ) − ln 2 } ≤ C 1 ( ρ ) ≤ OLC ( ρ ) (5) \max\{0,\operatorname{OLC}(\rho)-\ln2\}\leq C_1(\rho)\leq\operatorname{OLC}(\rho) \tag*{(5)} max { 0 , OLC ( ρ ) − ln 2 } ≤ C 1 ( ρ ) ≤ OLC ( ρ ) (5)
(in dimension n n n the same holds with ln n \ln n ln n in place of ln 2 \ln2 ln 2 , by Proposition 3.2.3 (a),(c) and τ ( ρ ) τ ( ρ − 1 ) ≥ ∥ ρ ∥ n ⋅ ∥ ρ − 1 ∥ n \tau(\rho)\tau(\rho^{-1})\geq\frac{\|\rho\|}{n}\cdot\frac{\|\rho^{-1}\|}{n} τ ( ρ ) τ ( ρ − 1 ) ≥ n ∥ ρ ∥ ⋅ n ∥ ρ − 1 ∥ ), so a n ≥ 0 a_n\geq0 a n ≥ 0 and b n ≥ 0 b_n\geq0 b n ≥ 0 measure how far ρ n \rho_n ρ n is from saturating the two lower bounds for the contrast: a n = 0 a_n=0 a n = 0 for the tracial state and b n → 0 b_n\to0 b n → 0 as ρ n \rho_n ρ n approaches a pure state. Explicitly, a n = ln cosh z n a_n=\ln\cosh z_n a n = ln cosh z n and b n = ln ( 1 + e − 2 z n ) b_n=\ln(1+e^{-2z_n}) b n = ln ( 1 + e − 2 z n ) .
The following statement is not new: it is a reformulation of the criteria of Araki–Woods and of Connes’ formula for the invariant T T T . We include it because in the language of contrasts these classical results take a strikingly simple form, in which the types I I 1 \mathrm{II}_1 II 1 and I correspond exactly to the two lower bounds in (5) , and the Michelson and the logarithmic contrast play the roles of amplitude and frequency.
Proposition 3.6.2 (Araki–Woods and Connes, reformulated). Let M = ⨂ n ( M 2 ( C ) , ρ n ) \mathcal{M} = \bigotimes_{n}(M_{2}(\mathbb{C}),\rho_{n}) M = ⨂ n ( M 2 ( C ) , ρ n ) . With the above notation:
(a) M \mathcal{M} M is of type I I 1 \mathrm{II}_{1} II 1 if and only if ∑ n a n < ∞ \sum_{n} a_{n} < \infty ∑ n a n < ∞ , equivalently ∑ n OMC ( ρ n ) 2 < ∞ \sum_{n} \operatorname{OMC}(\rho_{n})^{2} < \infty ∑ n OMC ( ρ n ) 2 < ∞ .
(b) M \mathcal{M} M is of type I \mathrm{I} I if and only if ∑ n b n < ∞ \sum_{n} b_{n} < \infty ∑ n b n < ∞ , equivalently ∑ n ( 1 − OMC ( ρ n ) ) < ∞ \sum_{n}(1-\operatorname{OMC}(\rho_{n})) < \infty ∑ n ( 1 − OMC ( ρ n )) < ∞ , equivalently ∑ n e − 2 OLC ( ρ n ) < ∞ \sum_{n} e^{-2\operatorname{OLC}(\rho_{n})} < \infty ∑ n e − 2 OLC ( ρ n ) < ∞ .
(c) M \mathcal{M} M is of type I I I \mathrm{III} III if and only if ∑ n min { a n , b n } = ∞ \sum_{n}\min\{a_{n},b_{n}\}=\infty ∑ n min { a n , b n } = ∞ . Consequently M \mathcal{M} M is of type I I ∞ \mathrm{II}_{\infty} II ∞ if and only if ∑ n a n = ∑ n b n = ∞ \sum_{n}a_{n}=\sum_{n}b_{n}=\infty ∑ n a n = ∑ n b n = ∞ and ∑ n min { a n , b n } < ∞ \sum_{n}\min\{a_{n},b_{n}\}<\infty ∑ n min { a n , b n } < ∞ .
(d) Connes’ invariant is T ( M ) = { t ∈ R : ∑ n ≥ 1 ( 1 − OMC ( ρ n ) 2 ) sin 2 ( t OLC ( ρ n ) ) < ∞ } T(\mathcal{M})=\left\{t\in\mathbb{R}:\sum_{n\geq1}(1-\operatorname{OMC}(\rho_{n})^{2})\sin^{2}(t\operatorname{OLC}(\rho_{n}))<\infty\right\} T ( M ) = { t ∈ R : ∑ n ≥ 1 ( 1 − OMC ( ρ n ) 2 ) sin 2 ( t OLC ( ρ n )) < ∞ } .
Moreover, by Lemma 3.6.1,
∑ n ≤ N a n = C 1 ( ρ ( N ) ) , ∑ n ≤ N b n = C 1 ( ρ ( N ) ) − ( OLC ( ρ ( N ) ) − ln d N ) , \sum_{n\leq N}a_{n}=C_{1}(\rho^{(N)}),\qquad\sum_{n\leq N}b_{n}=C_{1}(\rho^{(N)})-\left(\operatorname{OLC}(\rho^{(N)})-\ln d_{N}\right), n ≤ N ∑ a n = C 1 ( ρ ( N ) ) , n ≤ N ∑ b n = C 1 ( ρ ( N ) ) − ( OLC ( ρ ( N ) ) − ln d N ) , so that M \mathcal{M} M is of type I I 1 \mathrm{II}_{1} II 1 (resp. of type I \mathrm{I} I ) if and only if C 1 ( ρ ( N ) ) C_{1}(\rho^{(N)}) C 1 ( ρ ( N ) ) stays within bounded distance of the lower bound 0 0 0 (resp. of the lower bound OLC ( ρ ( N ) ) − ln d N \operatorname{OLC}(\rho^{(N)})-\ln d_{N} OLC ( ρ ( N ) ) − ln d N ).
Proof. The eigenvalues of ρ n \rho_{n} ρ n are λ n = e z n / ( 2 cosh z n ) \lambda_{n}=e^{z_{n}}/(2\cosh z_{n}) λ n = e z n / ( 2 cosh z n ) and 1 − λ n 1-\lambda_{n} 1 − λ n , with ratio κ n = e 2 z n \kappa_{n}=e^{2z_{n}} κ n = e 2 z n , and OMC ( ρ n ) = tanh z n \operatorname{OMC}(\rho_{n})=\tanh z_{n} OMC ( ρ n ) = tanh z n . We use [39 Lemma 2.14] .
(a) M \mathcal{M} M is of type I I 1 \mathrm{II}_{1} II 1 iff ∑ n ( ∣ λ n 1 / 2 − 2 − 1 / 2 ∣ 2 + ∣ ( 1 − λ n ) 1 / 2 − 2 − 1 / 2 ∣ 2 ) < ∞ \sum_{n}(|\lambda_{n}^{1/2}-2^{-1/2}|^{2}+|(1-\lambda_{n})^{1/2}-2^{-1/2}|^{2})<\infty ∑ n ( ∣ λ n 1/2 − 2 − 1/2 ∣ 2 + ∣ ( 1 − λ n ) 1/2 − 2 − 1/2 ∣ 2 ) < ∞ . The n n n -th term equals 2 ( 1 − cosh ( z n / 2 ) / cosh z n ) 2(1-\cosh(z_{n}/2)/\sqrt{\cosh z_{n}}) 2 ( 1 − cosh ( z n /2 ) / cosh z n ) . This function of z n z_{n} z n , the function ln cosh z n \ln\cosh z_{n} ln cosh z n and the function tanh 2 z n \tanh^{2}z_{n} tanh 2 z n are increasing in z n ≥ 0 z_{n}\geq0 z n ≥ 0 , vanish only at z n = 0 z_{n}=0 z n = 0 , and are asymptotic to constant multiples of z n 2 z_{n}^{2} z n 2 as z n → 0 z_{n}\to0 z n → 0 . Convergence of any of the three series forces its terms, hence z n z_{n} z n , to tend to 0 0 0 , and then the terms are comparable; hence the three series converge simultaneously.
(b) M \mathcal{M} M is of type I iff ∑ n ( 1 − λ n ) < ∞ \sum_{n}(1-\lambda_{n})<\infty ∑ n ( 1 − λ n ) < ∞ , and 1 − λ n = u n / ( 1 + u n ) 1-\lambda_{n}=u_{n}/(1+u_{n}) 1 − λ n = u n / ( 1 + u n ) with u n = e − 2 z n ∈ ( 0 , 1 ] u_{n}=e^{-2z_{n}}\in(0,1] u n = e − 2 z n ∈ ( 0 , 1 ] , while b n = ln ( 1 + u n ) b_{n}=\ln(1+u_{n}) b n = ln ( 1 + u n ) and 1 − tanh z n = 2 u n / ( 1 + u n ) 1-\tanh z_{n}=2u_{n}/(1+u_{n}) 1 − tanh z n = 2 u n / ( 1 + u n ) ; all three are between 1 2 u n \frac{1}{2}u_{n} 2 1 u n and 2 u n 2u_{n} 2 u n .
(c) Since λ n ≥ 1 2 \lambda_{n}\geq\frac{1}{2} λ n ≥ 2 1 is the larger eigenvalue, the criterion of [39 Lemma 2.14] says that M \mathcal{M} M is of type III iff ∑ n ( 1 − λ n ) min { ( 1 − κ n − 1 ) 2 , C } = ∞ \sum_{n}(1-\lambda_{n})\min\{(1-\kappa_{n}^{-1})^{2},C\}=\infty ∑ n ( 1 − λ n ) min {( 1 − κ n − 1 ) 2 , C } = ∞ for some (equivalently, every) constant C > 0 C>0 C > 0 ; as ( 1 − κ n − 1 ) 2 ≤ 1 (1-\kappa_{n}^{-1})^{2}\leq1 ( 1 − κ n − 1 ) 2 ≤ 1 , we may take C = 1 C=1 C = 1 , and the minimum is then ( 1 − κ n − 1 ) 2 (1-\kappa_{n}^{-1})^{2} ( 1 − κ n − 1 ) 2 . The n n n -th term is a continuous positive function of z n > 0 z_{n}>0 z n > 0 which is asymptotic to 2 z n 2 2z_{n}^{2} 2 z n 2 as z n → 0 z_{n}\to0 z n → 0 and to e − 2 z n e^{-2z_{n}} e − 2 z n as z n → ∞ z_{n}\to\infty z n → ∞ . The same holds, with the constants 1 2 \frac{1}{2} 2 1 and 1 1 1 , for min { ln cosh z n , ln ( 1 + e − 2 z n ) } \min\{\ln\cosh z_{n},\ln(1+e^{-2z_{n}})\} min { ln cosh z n , ln ( 1 + e − 2 z n )} , so the two are bounded by constant multiples of each other (terms with z n = 0 z_{n}=0 z n = 0 vanish in both series). The statement on type I I ∞ \mathrm{II}_{\infty} II ∞ follows from (a)–(c), since an ITPFI factor is of one of the types I, I I 1 \mathrm{II}_{1} II 1 , I I ∞ \mathrm{II}_{\infty} II ∞ , III.
(d) By [40 Cor. 1.3.9] , T ( M ) = { t : ∑ n ( 1 − ∣ λ n 1 + i t + ( 1 − λ n ) 1 + i t ∣ ) < ∞ } T(\mathcal{M})=\{t:\sum_{n}(1-|\lambda_{n}^{1+it}+(1-\lambda_{n})^{1+it}|)<\infty\} T ( M ) = { t : ∑ n ( 1 − ∣ λ n 1 + i t + ( 1 − λ n ) 1 + i t ∣ ) < ∞ } , and
∣ λ n 1 + i t + ( 1 − λ n ) 1 + i t ∣ 2 = 1 − 4 λ n ( 1 − λ n ) sin 2 ( t z n ) = 1 − ( 1 − tanh 2 z n ) sin 2 ( t z n ) , |\lambda_{n}^{1+it}+(1-\lambda_{n})^{1+it}|^{2}=1-4\lambda_{n}(1-\lambda_{n})\sin^{2}(tz_{n})=1-(1-\tanh^{2}z_{n})\sin^{2}(tz_{n}), ∣ λ n 1 + i t + ( 1 − λ n ) 1 + i t ∣ 2 = 1 − 4 λ n ( 1 − λ n ) sin 2 ( t z n ) = 1 − ( 1 − tanh 2 z n ) sin 2 ( t z n ) ,
while u 2 ≤ 1 − 1 − u ≤ u \frac{u}{2}\leq1-\sqrt{1-u}\leq u 2 u ≤ 1 − 1 − u ≤ u for u ∈ [ 0 , 1 ] u\in[0,1] u ∈ [ 0 , 1 ] .
Corollary 3.6.3. If sup n OLC ( ρ n ) < ∞ \sup_{n}\operatorname{OLC}(\rho_{n})<\infty sup n OLC ( ρ n ) < ∞ , then ⨂ n ( M 2 ( C ) , ρ n ) \bigotimes_{n}(M_{2}(\mathbb{C}),\rho_{n}) ⨂ n ( M 2 ( C ) , ρ n ) is of type I I 1 \mathrm{II}_{1} II 1 when ∑ n OLC ( ρ n ) 2 < ∞ \sum_{n}\operatorname{OLC}(\rho_{n})^{2}<\infty ∑ n OLC ( ρ n ) 2 < ∞ and of type III otherwise.
Proof. If z n ≤ Z z_{n}\leq Z z n ≤ Z for all n n n , then b n ≥ ln ( 1 + e − 2 Z ) > 0 b_{n}\geq\ln(1+e^{-2Z})>0 b n ≥ ln ( 1 + e − 2 Z ) > 0 , so min { a n , b n } \min\{a_{n},b_{n}\} min { a n , b n } is comparable with min { a n , 1 } \min\{a_{n},1\} min { a n , 1 } , and ∑ n min { a n , 1 } < ∞ \sum_{n}\min\{a_{n},1\}<\infty ∑ n min { a n , 1 } < ∞ iff ∑ n a n < ∞ \sum_{n}a_{n}<\infty ∑ n a n < ∞ iff ∑ n z n 2 < ∞ \sum_{n}z_{n}^{2}<\infty ∑ n z n 2 < ∞ .
Components of arbitrary size with bounded contrast
For components of arbitrary size the criterion for type I I 1 \mathrm{II}_{1} II 1 can still be expressed through the contrasts of the partial products, provided the contrasts of the components are uniformly bounded.
Theorem 3.6.4. Let M = ⨂ k ( M n k ( C ) , ρ k ) \mathcal{M} = \bigotimes_k (M_{n_k}(\mathbb{C}), \rho_k) M = ⨂ k ( M n k ( C ) , ρ k ) with n k < ∞ n_k < \infty n k < ∞ arbitrary and sup k OLC ( ρ k ) < ∞ \sup_k \operatorname{OLC}(\rho_k) < \infty sup k OLC ( ρ k ) < ∞ . The following are equivalent:
(a) M \mathcal{M} M is of type I I 1 \mathrm{II}_1 II 1 ;
(b) sup N C p ( ρ ( N ) ) = ∑ k C p ( ρ k ) < ∞ \sup_N C_p(\rho^{(N)}) = \sum_k C_p(\rho_k) < \infty sup N C p ( ρ ( N ) ) = ∑ k C p ( ρ k ) < ∞ for some, equivalently for every, p ∈ ( 0 , ∞ ) p \in(0,\infty) p ∈ ( 0 , ∞ ) ;
(c) sup N V ( ρ ( N ) ) = ∑ k V ( ρ k ) < ∞ \sup_N V(\rho^{(N)}) = \sum_k V(\rho_k) < \infty sup N V ( ρ ( N ) ) = ∑ k V ( ρ k ) < ∞ .
Moreover, under the hypothesis sup k OLC ( ρ k ) < ∞ \sup_k \operatorname{OLC}(\rho_k) < \infty sup k OLC ( ρ k ) < ∞ the factor M \mathcal{M} M is never of type I.
Proof. Let Z = sup k OLC ( ρ k ) Z = \sup_k \operatorname{OLC}(\rho_k) Z = sup k OLC ( ρ k ) , let μ k , j = n k λ k , j \mu_{k,j} = n_k\lambda_{k,j} μ k , j = n k λ k , j be the eigenvalues of h k = n k ρ k h_k = n_k\rho_k h k = n k ρ k , and let X k X_k X k be a random variable with law ν h k \nu_{h_k} ν h k , i.e. taking the values ln μ k , j \ln\mu_{k,j} ln μ k , j with equal probabilities. Then E e X k = τ ( h k ) = 1 \mathbb{E}e^{X_k} = \tau(h_k) = 1 E e X k = τ ( h k ) = 1 , so the smallest and the largest of the μ k , j \mu_{k,j} μ k , j lie on different sides of 1 1 1 , and since their ratio is at most e 2 Z e^{2Z} e 2 Z , ∣ X k ∣ ≤ 2 Z \lvert X_k\rvert\le2Z ∣ X k ∣ ≤ 2 Z . By [39 Lemma 2.14] , M \mathcal{M} M is of type I I 1 \mathrm{II}_1 II 1 if and only if ∑ k S k < ∞ \sum_k S_k < \infty ∑ k S k < ∞ , where
S k = ∑ j ∣ λ k , j 1 / 2 − n k − 1 / 2 ∣ 2 = 1 n k ∑ j ( μ k , j 1 / 2 − 1 ) 2 = E ( e X k / 2 − 1 ) 2 . S_k = \sum_j \left\lvert\lambda_{k,j}^{1/2} - n_k^{-1/2}\right\rvert^2 = \frac{1}{n_k}\sum_j(\mu_{k,j}^{1/2}-1)^2 = \mathbb{E}(e^{X_k/2}-1)^2. S k = j ∑ λ k , j 1/2 − n k − 1/2 2 = n k 1 j ∑ ( μ k , j 1/2 − 1 ) 2 = E ( e X k /2 − 1 ) 2 .
We show that S k S_k S k , V k = V ( ρ k ) = Var X k V_k = V(\rho_k) = \operatorname{Var} X_k V k = V ( ρ k ) = Var X k and C p ( ρ k ) C_p(\rho_k) C p ( ρ k ) are bounded by constant multiples of each other, with constants depending only on Z Z Z and p p p ; together with Lemma 3.6.1 this proves the equivalences.
Since ( e x / 2 − 1 ) 2 (e^{x/2}-1)^2 ( e x /2 − 1 ) 2 is bounded above and below by constant multiples of x 2 x^2 x 2 on [ − 2 Z , 2 Z ] [-2Z,2Z] [ − 2 Z , 2 Z ] , S k S_k S k is comparable with E X k 2 = V k + ( E X k ) 2 \mathbb{E}X_k^2 = V_k + (\mathbb{E}X_k)^2 E X k 2 = V k + ( E X k ) 2 . Put m = E X k m = \mathbb{E}X_k m = E X k and Y = X k − m Y = X_k-m Y = X k − m , so ∣ Y ∣ ≤ 4 Z \lvert Y\rvert\le4Z ∣ Y ∣ ≤ 4 Z . By Jensen’s inequality e m ≤ E e X k = 1 e^m \le\mathbb{E}e^{X_k} = 1 e m ≤ E e X k = 1 , i.e. m ≤ 0 m \le0 m ≤ 0 , and from e y ≤ 1 + y + 1 2 e 4 Z y 2 e^y \le1+y+\frac{1}{2}e^{4Z}y^2 e y ≤ 1 + y + 2 1 e 4 Z y 2 for ∣ y ∣ ≤ 4 Z \lvert y\rvert\le4Z ∣ y ∣ ≤ 4 Z we get 1 = e m E e Y ≤ e m ( 1 + 1 2 e 4 Z V k ) 1=e^m\mathbb{E}e^Y \le e^m(1+\frac{1}{2}e^{4Z}V_k) 1 = e m E e Y ≤ e m ( 1 + 2 1 e 4 Z V k ) , hence 0 ≤ − m ≤ 1 2 e 4 Z V k 0 \le-m \le\frac{1}{2}e^{4Z}V_k 0 ≤ − m ≤ 2 1 e 4 Z V k . As V k ≤ 4 Z 2 V_k \le4Z^2 V k ≤ 4 Z 2 , this gives m 2 ≤ Z 2 e 8 Z V k m^2 \le Z^2e^{8Z}V_k m 2 ≤ Z 2 e 8 Z V k , and E X k 2 \mathbb{E}X_k^2 E X k 2 is comparable with V k V_k V k .
For the contrast, let X k ′ X'_k X k ′ be an independent copy of X k X_k X k . Then e 2 p C p ( ρ k ) = E cosh ( p ( X k − X k ′ ) ) e^{2pC_p(\rho_k)} = \mathbb{E}\cosh(p(X_k-X'_k)) e 2 p C p ( ρ k ) = E cosh ( p ( X k − X k ′ )) , E ( X k − X k ′ ) 2 = 2 V k \mathbb{E}(X_k-X'_k)^2 = 2V_k E ( X k − X k ′ ) 2 = 2 V k and ∣ X k − X k ′ ∣ ≤ 4 Z \lvert X_k-X'_k\rvert\le4Z ∣ X k − X k ′ ∣ ≤ 4 Z , so
1 + p 2 V k ≤ e 2 p C p ( ρ k ) ≤ 1 + p 2 cosh ( 4 p Z ) V k , 1+p^2V_k \le e^{2pC_p(\rho_k)} \le1+p^2\cosh(4pZ)V_k, 1 + p 2 V k ≤ e 2 p C p ( ρ k ) ≤ 1 + p 2 cosh ( 4 pZ ) V k ,
and ln ( 1 + u ) \ln(1+u) ln ( 1 + u ) is comparable with u u u for 0 ≤ u ≤ 4 p 2 Z 2 cosh ( 4 p Z ) 0 \le u \le4p^2Z^2\cosh(4pZ) 0 ≤ u ≤ 4 p 2 Z 2 cosh ( 4 pZ ) . Finally, λ k , 1 ≤ e 2 Z λ k , n k \lambda_{k,1} \le e^{2Z}\lambda_{k,n_k} λ k , 1 ≤ e 2 Z λ k , n k and 1 − λ k , 1 ≥ λ k , n k 1-\lambda_{k,1} \ge\lambda_{k,n_k} 1 − λ k , 1 ≥ λ k , n k give λ k , 1 ≤ ( 1 + e − 2 Z ) − 1 \lambda_{k,1} \le(1+e^{-2Z})^{-1} λ k , 1 ≤ ( 1 + e − 2 Z ) − 1 , so ∑ k ( 1 − λ k , 1 ) = ∞ \sum_k(1-\lambda_{k,1})=\infty ∑ k ( 1 − λ k , 1 ) = ∞ and M \mathcal{M} M is not of type I [39 Lemma 2.14] . □
When the conditions of Theorem 3.6.4 fail, M \mathcal{M} M is of type I I ∞ \mathrm{II}_\infty II ∞ or III, and the theorem does not decide between the two; for components of size two Corollary 3.6.3 shows that the type is III.
In words: an Araki–Woods factor with uniformly bounded component contrasts is of type I I 1 \mathrm{II}_1 II 1 exactly when the Schatten contrasts of the partial products ρ ( N ) \rho^{(N)} ρ ( N ) stay bounded, or equivalently when the logarithmic spectra of the ρ ( N ) \rho^{(N)} ρ ( N ) have bounded variance, while their operator-norm contrasts OLC ( ρ ( N ) ) = ∑ k ≤ N OLC ( ρ k ) \operatorname{OLC}(\rho^{(N)}) = \sum_{k\le N}\operatorname{OLC}(\rho_k) OLC ( ρ ( N ) ) = ∑ k ≤ N OLC ( ρ k ) may well diverge.
Arbitrary component contrasts. Without the bound on the component contrasts, the contrasts of the partial products still give a sufficient condition for type I I 1 \mathrm{II}_1 II 1 , but only for p ≥ 1 p \ge1 p ≥ 1 .
Proposition 3.6.5. Let M = ⨂ k ( M n k ( C ) , ρ k ) \mathcal{M} = \bigotimes_k(M_{n_k}(\mathbb{C}),\rho_k) M = ⨂ k ( M n k ( C ) , ρ k ) with arbitrary n k n_k n k . If ∑ k C p ( ρ k ) < ∞ \sum_k C_p(\rho_k) < \infty ∑ k C p ( ρ k ) < ∞ for some p ≥ 1 p \ge1 p ≥ 1 (it is enough to assume this for p = 1 p=1 p = 1 ), then M \mathcal{M} M is of type I I 1 \mathrm{II}_1 II 1 .
Proof. Let h k = n k ρ k h_k = n_k\rho_k h k = n k ρ k , so that τ ( h k ) = 1 \tau(h_k)=1 τ ( h k ) = 1 . As in the proof of Theorem 3.6.4 , M \mathcal{M} M is of type I I 1 \mathrm{II}_1 II 1 if ∑ k S k < ∞ \sum_k S_k < \infty ∑ k S k < ∞ , where S k = ∑ j ∣ λ k , j 1 / 2 − n k − 1 / 2 ∣ 2 = 2 − 2 τ ( h k 1 / 2 ) S_k = \sum_j\lvert\lambda_{k,j}^{1/2}-n_k^{-1/2}\rvert^2 = 2-2\tau(h_k^{1/2}) S k = ∑ j ∣ λ k , j 1/2 − n k − 1/2 ∣ 2 = 2 − 2 τ ( h k 1/2 ) . Write 0 = θ ⋅ 1 2 + ( 1 − θ ) ( − p ) 0=\theta\cdot\frac{1}{2}+(1-\theta)(-p) 0 = θ ⋅ 2 1 + ( 1 − θ ) ( − p ) with θ = 2 p 1 + 2 p \theta=\frac{2p}{1+2p} θ = 1 + 2 p 2 p . By Hölder’s inequality (log-convexity of s ↦ τ ( h s ) s\mapsto\tau(h^s) s ↦ τ ( h s ) ), 1 = τ ( h k p ) ≤ τ ( h k 1 / 2 ) θ τ ( h k − p ) 1 − θ 1=\tau(h_k^p)\leq\tau(h_k^{1/2})^\theta\tau(h_k^{-p})^{1-\theta} 1 = τ ( h k p ) ≤ τ ( h k 1/2 ) θ τ ( h k − p ) 1 − θ , i.e. ln τ ( h k 1 / 2 ) ≥ − 1 2 p ln τ ( h k − p ) \ln\tau(h_k^{1/2})\geq-\frac{1}{2p}\ln\tau(h_k^{-p}) ln τ ( h k 1/2 ) ≥ − 2 p 1 ln τ ( h k − p ) . By Jensen’s inequality τ ( h k p ) ≥ τ ( h k ) p = 1 \tau(h_k^p)\geq\tau(h_k)^p=1 τ ( h k p ) ≥ τ ( h k ) p = 1 for p ≥ 1 p\geq1 p ≥ 1 , so 1 2 p ln τ ( h k − p ) ≤ C p ( h k ) = C p ( ρ k ) \frac{1}{2p}\ln\tau(h_k^{-p})\leq C_p(h_k)=C_p(\rho_k) 2 p 1 ln τ ( h k − p ) ≤ C p ( h k ) = C p ( ρ k ) . Hence τ ( h k 1 / 2 ) ≥ e − C p ( ρ k ) \tau(h_k^{1/2})\geq e^{-C_p(\rho_k)} τ ( h k 1/2 ) ≥ e − C p ( ρ k ) and S k ≤ 2 ( 1 − e − C p ( ρ k ) ) ≤ 2 C p ( ρ k ) S_k\leq2(1-e^{-C_p(\rho_k)})\leq2C_p(\rho_k) S k ≤ 2 ( 1 − e − C p ( ρ k ) ) ≤ 2 C p ( ρ k ) .
Put p c ( ρ ∙ ) = sup { p > 0 : ∑ k C p ( ρ k ) < ∞ } p_c(\rho_\bullet)=\sup\{p>0:\sum_k C_p(\rho_k)<\infty\} p c ( ρ ∙ ) = sup { p > 0 : ∑ k C p ( ρ k ) < ∞ } (with sup ∅ = 0 ) ; s i n c e \sup\varnothing=0); since sup ∅ = 0 ) ; s in ce C_p$ is nondecreasing in p p p , the sum is finite on ( 0 , p c ) (0,p_c) ( 0 , p c ) and infinite on ( p c , ∞ ) (p_c,\infty) ( p c , ∞ ) . By Theorem 3.6.4 , if sup k OLC ( ρ k ) < ∞ \sup_k\operatorname{OLC}(\rho_k)<\infty sup k OLC ( ρ k ) < ∞ then p c ∈ { 0 , ∞ } p_c\in\{0,\infty\} p c ∈ { 0 , ∞ } , and p c = ∞ p_c=\infty p c = ∞ exactly for type I I 1 \mathrm{II}_1 II 1 . In general the threshold is not determined by the type, and the condition p ≥ 1 p\geq1 p ≥ 1 in Proposition 3.6.5 cannot be dropped. In the following example h k = n k ρ k h_k=n_k\rho_k h k = n k ρ k and S k S_k S k are as in the proof above, and we use that C p ( ρ k ) = C p ( h k ) C_p(\rho_k)=C_p(h_k) C p ( ρ k ) = C p ( h k ) .
Example 3.6.6. Put t k = 2 − k t_k=2^{-k} t k = 2 − k .
(a) Type I I 1 \mathrm{II}_1 II 1 with p c = 2 p_c=2 p c = 2 . Let n k = 4 k n_k=4^k n k = 4 k and h k = c k diag ( 2 k , 1 , … , 1 ) h_k=c_k\operatorname{diag}(2^k,1,\ldots,1) h k = c k diag ( 2 k , 1 , … , 1 ) , c k = 4 k / ( 2 k + 4 k − 1 ) c_k=4^k/(2^k+4^k-1) c k = 4 k / ( 2 k + 4 k − 1 ) . Then c k ≥ 1 − t k c_k\geq1-t_k c k ≥ 1 − t k , so τ ( h k 1 / 2 ) ≥ c k 1 / 2 ( 1 − 4 − k ) ≥ 1 − 2 t k \tau(h_k^{1/2})\geq c_k^{1/2}(1-4^{-k})\geq1-2t_k τ ( h k 1/2 ) ≥ c k 1/2 ( 1 − 4 − k ) ≥ 1 − 2 t k and S k ≤ 4 t k S_k\leq4t_k S k ≤ 4 t k : M \mathcal{M} M is of type I I 1 \mathrm{II}_1 II 1 . Further
2 p C p ( ρ k ) = ln ( 1 + ( 2 p k − 1 ) 4 − k ) + ln ( 1 − ( 1 − 2 − p k ) 4 − k ) . 2pC_p(\rho_k)=\ln(1+(2^{pk}-1)4^{-k})+\ln(1-(1-2^{-pk})4^{-k}). 2 p C p ( ρ k ) = ln ( 1 + ( 2 p k − 1 ) 4 − k ) + ln ( 1 − ( 1 − 2 − p k ) 4 − k ) .
For p < 2 p<2 p < 2 this is at most 2 ( p − 2 ) k 2^{(p-2)k} 2 ( p − 2 ) k , so ∑ k C p ( ρ k ) < ∞ \sum_k C_p(\rho_k)<\infty ∑ k C p ( ρ k ) < ∞ . For p = 2 p=2 p = 2 and k ≥ 1 k\geq1 k ≥ 1 the product of the two arguments is at least 7 4 ⋅ 3 4 > 1 \frac{7}{4}\cdot\frac{3}{4}>1 4 7 ⋅ 4 3 > 1 , so C 2 ( ρ k ) C_2(\rho_k) C 2 ( ρ k ) is bounded below by a positive constant and ∑ k C 2 ( ρ k ) = ∞ \sum_k C_2(\rho_k)=\infty ∑ k C 2 ( ρ k ) = ∞ .
(b) Not of type I I 1 \mathrm{II}_1 II 1 , with p c = 1 p_c=1 p c = 1 . Let n k = 2 k n_k=2^k n k = 2 k and let h k h_k h k have the eigenvalue 1 2 t k \frac{1}{2t_k} 2 t k 1 once and the eigenvalue 1 2 ( 1 − t k ) \frac{1}{2(1-t_k)} 2 ( 1 − t k ) 1 with multiplicity 2 k − 1 2^k-1 2 k − 1 . Then τ ( h k 1 / 2 ) = t k / 2 + ( 1 − t k ) / 2 → 2 − 1 / 2 \tau(h_k^{1/2})=\sqrt{t_k/2}+\sqrt{(1-t_k)/2}\to2^{-1/2} τ ( h k 1/2 ) = t k /2 + ( 1 − t k ) /2 → 2 − 1/2 , so S k → 2 − 2 S_k\to2-\sqrt{2} S k → 2 − 2 and ∑ k S k = ∞ \sum_k S_k=\infty ∑ k S k = ∞ . For 0 < p < 1 0<p<1 0 < p < 1 ,
τ ( h k p ) τ ( h k − p ) = ( t k 1 − p + ( 1 − t k ) 1 − p ) ( t k 1 + p + ( 1 − t k ) 1 + p ) ≤ ( 1 + t k 1 − p ) ( 1 + t k 1 + p ) ≤ 1 + 3 t k 1 − p , \begin{aligned}
\tau(h_k^p)\tau(h_k^{-p})=(t_k^{1-p}+(1-t_k)^{1-p})(t_k^{1+p}+(1-t_k)^{1+p}) \\
&\leq(1+t_k^{1-p})(1+t_k^{1+p})\leq1+3t_k^{1-p},
\end{aligned} τ ( h k p ) τ ( h k − p ) = ( t k 1 − p + ( 1 − t k ) 1 − p ) ( t k 1 + p + ( 1 − t k ) 1 + p ) ≤ ( 1 + t k 1 − p ) ( 1 + t k 1 + p ) ≤ 1 + 3 t k 1 − p ,
so C p ( ρ k ) ≤ 3 2 2 − ( 1 − p ) k C_p(\rho_k)\leq\frac{3}{2}2^{-(1-p)k} C p ( ρ k ) ≤ 2 3 2 − ( 1 − p ) k and the sum converges; for p = 1 p=1 p = 1 and k ≥ 2 k\geq2 k ≥ 2 , τ ( h k ) τ ( h k − 1 ) = 2 − 4 t k + 4 t k 2 ≥ 5 4 \tau(h_k)\tau(h_k^{-1})=2-4t_k+4t_k^2\geq\frac{5}{4} τ ( h k ) τ ( h k − 1 ) = 2 − 4 t k + 4 t k 2 ≥ 4 5 , so ∑ k C 1 ( ρ k ) = ∞ \sum_k C_1(\rho_k)=\infty ∑ k C 1 ( ρ k ) = ∞ . In view of Proposition 3.6.5 this example is sharp.
(c) Type I I 1 \mathrm{II}_1 II 1 with p c = 0 p_c=0 p c = 0 . Let n k = 2 k n_k=2^k n k = 2 k and let h k h_k h k have the eigenvalue ε k = exp ( − e k ) \varepsilon_k=\exp(-e^k) ε k = exp ( − e k ) once and the eigenvalue η k = ( 1 − t k ε k ) / ( 1 − t k ) ≥ 1 \eta_k=(1-t_k\varepsilon_k)/(1-t_k)\geq1 η k = ( 1 − t k ε k ) / ( 1 − t k ) ≥ 1 with multiplicity 2 k − 1 2^k-1 2 k − 1 . Then τ ( h k 1 / 2 ) ≥ ( 1 − t k ) η k 1 / 2 ≥ 1 − t k \tau(h_k^{1/2})\geq(1-t_k)\eta_k^{1/2}\geq1-t_k τ ( h k 1/2 ) ≥ ( 1 − t k ) η k 1/2 ≥ 1 − t k , so S k ≤ 2 t k S_k\leq2t_k S k ≤ 2 t k and M \mathcal{M} M is of type I I 1 \mathrm{II}_1 II 1 ; but for every p > 0 p>0 p > 0 ,
2 p C p ( ρ k ) ≥ ln ( ( 1 − t k ) η k p ) + ln ( t k ε k − p ) ≥ − ( k + 1 ) ln 2 + p e k → ∞ , 2pC_p(\rho_k)\geq\ln((1-t_k)\eta_k^p)+\ln(t_k\varepsilon_k^{-p})\geq-(k+1)\ln2+pe^k\to\infty, 2 p C p ( ρ k ) ≥ ln (( 1 − t k ) η k p ) + ln ( t k ε k − p ) ≥ − ( k + 1 ) ln 2 + p e k → ∞ ,
so ∑ k C p ( ρ k ) = ∞ \sum_k C_p(\rho_k)=\infty ∑ k C p ( ρ k ) = ∞ .
Thus, without the bound on the component contrasts, finiteness of ∑ k C p ( ρ k ) \sum_k C_p(\rho_k) ∑ k C p ( ρ k ) for some p ≥ 1 p\geq1 p ≥ 1 implies type I I 1 \mathrm{II}_1 II 1 , finiteness for p < 1 p<1 p < 1 does not, and a factor of type I I 1 \mathrm{II}_1 II 1 may have p c = 2 p_c=2 p c = 2 or p c = 0 p_c=0 p c = 0 .
Remark 3.6.7. The contrasts of the partial products are contrasts of modular operators. For a positive definite density matrix ρ ∈ M d ( C ) \rho\in M_d(\mathbb{C}) ρ ∈ M d ( C ) , let M d ( C ) M_d(\mathbb{C}) M d ( C ) act by left multiplication on itself with the inner product ⟨ x , y ⟩ = Tr ( x ∗ y ) \langle x,y\rangle=\operatorname{Tr}(x^*y) ⟨ x , y ⟩ = Tr ( x ∗ y ) and cyclic vector ρ 1 / 2 \rho^{1/2} ρ 1/2 ; the modular operator of Tr ( ρ ⋅ ) \operatorname{Tr}(\rho\cdot) Tr ( ρ ⋅ ) is then Δ ρ ( x ) = ρ x ρ − 1 \Delta_\rho(x)=\rho x\rho^{-1} Δ ρ ( x ) = ρ x ρ − 1 , with eigenvalues λ i / λ j \lambda_i/\lambda_j λ i / λ j and the matrix units as eigenvectors. With the normalised trace φ \varphi φ on B ( M d ( C ) ) ≅ M d 2 ( C ) B(M_d(\mathbb{C}))\cong M_{d^2}(\mathbb{C}) B ( M d ( C )) ≅ M d 2 ( C ) its logarithmic spectral measure is ν ρ ∗ ν ˇ ρ \nu_\rho*\check{\nu}_\rho ν ρ ∗ ν ˇ ρ , so C p ( Δ ρ ) = 2 C p ( ρ ) C_p(\Delta_\rho)=2C_p(\rho) C p ( Δ ρ ) = 2 C p ( ρ ) by Proposition 3.2.3 (e), and OLC ( Δ ρ ) = 2 OLC ( ρ ) \operatorname{OLC}(\Delta_\rho)=2\operatorname{OLC}(\rho) OLC ( Δ ρ ) = 2 OLC ( ρ ) because diameters of supports add under convolution. Hence ∑ k ≤ N C p ( ρ k ) = 1 2 C p ( Δ ρ ( N ) ) \sum_{k\leq N}C_p(\rho_k)=\frac{1}{2}C_p(\Delta_{\rho^{(N)}}) ∑ k ≤ N C p ( ρ k ) = 2 1 C p ( Δ ρ ( N ) ) , where Δ ρ ( N ) \Delta_{\rho^{(N)}} Δ ρ ( N ) is the modular operator of the restriction of φ \varphi φ to A N \mathcal{A}_N A N . These quantities are defined through the chosen tensor product decomposition; we do not know whether they are determined by ( M , φ ) (\mathcal{M},\varphi) ( M , φ ) .
Connes’ invariant and the spectral profile
For an arbitrary ITPFI factor Connes’ formula [40, Cor. 1.3.9 reads T ( M ) = { t : ∑ k ( 1 − ∣ Tr ρ k 1 + i t ∣ ) < ∞ } T(\mathcal{M})=\{t:\sum_k(1-\lvert\operatorname{Tr}\rho_k^{1+it}\rvert)<\infty\} T ( M ) = { t : ∑ k ( 1 − ∣ Tr ρ k 1 + i t ∣) < ∞ } . Since Tr ρ 1 + i t = τ ( ρ 1 + i t ) / τ ( ρ ) \operatorname{Tr}\rho^{1+it}=\tau(\rho^{1+it})/\tau(\rho) Tr ρ 1 + i t = τ ( ρ 1 + i t ) / τ ( ρ ) is the characteristic function of the tilted logarithmic spectral measure ν ρ k ( 1 ) \nu_{\rho_k}^{(1)} ν ρ k ( 1 ) ,
T ( M ) = { t ∈ R : ∑ k ( 1 − ∣ χ k ( t ) ∣ ) < ∞ } , χ k ( t ) = ∫ e i t x d ν ρ k ( 1 ) ( x ) = τ ( ρ k 1 + i t ) τ ( ρ k ) . T(\mathcal{M})=\left\{t\in\mathbb{R}:\sum_k(1-\lvert\chi_k(t)\rvert)<\infty\right\},\qquad\chi_k(t)=\int e^{itx}\,d\nu_{\rho_k}^{(1)}(x)=\frac{\tau(\rho_k^{1+it})}{\tau(\rho_k)}. T ( M ) = { t ∈ R : k ∑ ( 1 − ∣ χ k ( t )∣) < ∞ } , χ k ( t ) = ∫ e i t x d ν ρ k ( 1 ) ( x ) = τ ( ρ k ) τ ( ρ k 1 + i t ) .
In other words, the invariant T T T is determined by the analytic continuation of the functions e L ρ k ( s ) = τ ( ρ k s ) e^{L_{\rho_k}(s)}=\tau(\rho_k^s) e L ρ k ( s ) = τ ( ρ k s ) , i.e. of the logarithmic spectral profiles of the components, to the line Re s = 1 \operatorname{Re}s=1 Re s = 1 , just as the Schatten contrasts are determined by them on the real axis. For components of size two this is Proposition 3.6.2(d). Since 1 − ∣ χ k ( t ) ∣ ≤ 1 − ∣ χ k ( t ) ∣ 2 ≤ t 2 Var ( ν ρ k ( 1 ) ) 1-\lvert\chi_k(t)\rvert\leq1-\lvert\chi_k(t)\rvert^2\leq t^2\operatorname{Var}(\nu_{\rho_k}^{(1)}) 1 − ∣ χ k ( t )∣ ≤ 1 − ∣ χ k ( t ) ∣ 2 ≤ t 2 Var ( ν ρ k ( 1 ) ) , summability of these variances (of the tilted measures ν ρ k ( 1 ) \nu_{\rho_k}^{(1)} ν ρ k ( 1 ) , which are comparable with the variances V ( ρ k ) V(\rho_k) V ( ρ k ) when the contrasts are bounded) implies T ( M ) = R T(\mathcal{M})=\mathbb{R} T ( M ) = R , in accordance with Theorem 3.6.4. We also note that Connes’ invariant S ( M ) S(\mathcal{M}) S ( M ) of an ITPFI factor coincides with the asymptotic ratio set r ∞ ( M ) r_\infty(\mathcal{M}) r ∞ ( M ) of Araki and Woods [39 , 40 ] . We do not know whether S ( M ) S(\mathcal{M}) S ( M ) can be read off from the logarithmic spectral profiles of the tail products ⨂ k ≥ m ρ k \bigotimes_{k\geq m}\rho_k ⨂ k ≥ m ρ k , m → ∞ m\to\infty m → ∞ .
Recall that T ( M ) = R T(\mathcal{M})=\mathbb{R} T ( M ) = R for semifinite factors, T ( M ) = 2 π ∣ ln λ ∣ Z T(\mathcal{M})=\frac{2\pi}{\lvert\ln\lambda\rvert}\mathbb{Z} T ( M ) = ∣ l n λ ∣ 2 π Z in type I I I λ \mathrm{III}_\lambda III λ , 0 < λ < 1 0<\lambda<1 0 < λ < 1 , and T ( M ) = { 0 } T(\mathcal{M})=\{0\} T ( M ) = { 0 } in type I I I 1 \mathrm{III}_1 III 1 [40 ] .
Example 3.6.8 (Growth classes of the contrast). Let all n k = 2 n_k=2 n k = 2 and z n = OLC ( ρ n ) z_n=\operatorname{OLC}(\rho_n) z n = OLC ( ρ n ) .
Bounded contrast. If z n = z > 0 z_n=z>0 z n = z > 0 for all n n n , then M \mathcal{M} M is the Powers factor of type I I I λ \mathrm{III}_\lambda III λ with λ = e − 2 z = κ ( ρ n ) − 1 \lambda=e^{-2z}=\kappa(\rho_n)^{-1} λ = e − 2 z = κ ( ρ n ) − 1 , and T ( M ) = π z Z T(\mathcal{M})=\frac{\pi}{z}\mathbb{Z} T ( M ) = z π Z .
Polynomial condition numbers. If κ ( ρ n ) = n α \kappa(\rho_n)=n^\alpha κ ( ρ n ) = n α , i.e. z n = α 2 ln n z_n=\frac{\alpha}{2}\ln n z n = 2 α ln n , then b n b_n b n is comparable with n − α n^{-\alpha} n − α and a n → ∞ a_n\to\infty a n → ∞ . Hence M \mathcal{M} M is of type I for α > 1 \alpha>1 α > 1 and of type III for 0 < α ≤ 1 0<\alpha\leq1 0 < α ≤ 1 . In the latter case the series in Proposition 3.6.2(d) is comparable with ∑ n − α sin 2 ( t α 2 ln n ) \sum n^{-\alpha}\sin^2\left(\frac{t\alpha}{2}\ln n\right) ∑ n − α sin 2 ( 2 t α ln n ) , which diverges for every t ≠ 0 t\neq0 t = 0 ; hence T ( M ) = { 0 } T(\mathcal{M})=\{0\} T ( M ) = { 0 } , and M \mathcal{M} M is not of type I I I λ \mathrm{III}_\lambda III λ for any λ ∈ ( 0 , 1 ) \lambda\in(0,1) λ ∈ ( 0 , 1 ) . The invariant T T T alone does not decide between types I I I 0 \mathrm{III}_0 III 0 and I I I 1 \mathrm{III}_1 III 1 here.
Factorial contrasts. Let the contrast z j = ln 2 2 j ! z_j=\frac{\ln2}{2}j! z j = 2 l n 2 j ! occur with multiplicity 2 j ! 2^{j!} 2 j ! , j ≥ 1 j\geq1 j ≥ 1 ; factors of this kind appear in [51 ] . Since b b b is comparable with e − 2 z j = 2 − j ! e^{-2z_j}=2^{-j!} e − 2 z j = 2 − j ! on the j j j -th block, ∑ min { a n , b n } = ∞ \sum\min\{a_n,b_n\}=\infty ∑ min { a n , b n } = ∞ and M \mathcal{M} M is of type III. Moreover T ( M ) T(\mathcal{M}) T ( M ) is the set of t t t with ∑ j sin 2 ( t ln 2 2 j ! ) < ∞ \sum_j\sin^2\left(\frac{t\ln2}{2}j!\right)<\infty ∑ j sin 2 ( 2 t l n 2 j ! ) < ∞ , which contains 2 π ln 2 Q \frac{2\pi}{\ln2}\mathbb{Q} l n 2 2 π Q because r j ! ∈ Z rj!\in\mathbb{Z} r j ! ∈ Z for every r ∈ Q r\in\mathbb{Q} r ∈ Q and all large j j j . Thus T ( M ) ≠ { 0 } T(\mathcal{M})\neq\{0\} T ( M ) = { 0 } , and T ( M ) T(\mathcal{M}) T ( M ) is not of the form 2 π ∣ ln λ ∣ Z \frac{2\pi}{\lvert\ln\lambda\rvert}\mathbb{Z} ∣ l n λ ∣ 2 π Z since it contains a dense subgroup; by the values of T T T recalled above, M \mathcal{M} M is of type I I I 0 \mathrm{III}_0 III 0 .
Thus bounded contrasts give type I I 1 \mathrm{II}_1 II 1 or III (Corollary 3.6.3), with type I I I λ \mathrm{III}_\lambda III λ for constant contrast; polynomially growing condition numbers give type I or a type III factor with trivial T T T -invariant; and the factorially growing contrasts of Example 3.6.8(iii), taken with matching multiplicities, give type I I I 0 \mathrm{III}_0 III 0 . The type is, however, not a function of the growth rate alone: the arithmetic of the sequence of contrasts enters through the factor sin 2 ( t OLC ( ρ n ) ) \sin^2(t\operatorname{OLC}(\rho_n)) sin 2 ( t OLC ( ρ n )) in Proposition 3.6.2(d). In Example 3.6.8(iii) it is the fact that r j ! rj! r j ! is eventually an integer for every rational r r r , and not the size of j ! j! j ! , that makes T ( M ) T(\mathcal{M}) T ( M ) dense.
Concluding remarks
In a tracial von Neumann algebra the Schatten contrast turns out to measure three things: whether the functional is a trace, how small projections can be, and how the spectral measure of ln ∣ A ∣ \ln\lvert A\rvert ln ∣ A ∣ behaves at its edges; for infinite tensor products, the sequence of contrasts of the components reproduces the Araki–Woods type criteria for components of size two, and the type I I 1 \mathrm{II}_1 II 1 criterion for components of bounded contrast. We mention open questions.
Finer invariants. The invariants of § 6.7 and § 6.8 depend only on the range of the trace on projections, and therefore cannot separate non-isomorphic factors of type I I 1 \mathrm{II}_1 II 1 . Do joint contrasts of several noncommuting operators, as in the equality problem of Nikolaeva and Novikov [2 ] , carry finer information?
Type III. Is there an analogue of Theorem 3.3.1 for weights on factors of type III, with Haagerup L p L^p L p -spaces in place of L p ( M , τ ) L^p(\mathcal{M},\tau) L p ( M , τ ) ? Which growth and arithmetic conditions on the contrasts and on the profiles L ρ k L_{\rho_k} L ρ k of the component states characterise the types I I I 0 \mathrm{III}_0 III 0 , I I I λ \mathrm{III}_{\lambda} III λ , I I I 1 \mathrm{III}_1 III 1 of Araki–Woods factors with components of arbitrary size?
Infinite contrasts and finite sections
Introduction
Let A A A be a bounded invertible operator with condition number κ ( A ) = ∥ A ∥ ∥ A − 1 ∥ \kappa(A)=\lVert A\rVert\lVert A^{-1}\rVert κ ( A ) = ∥ A ∥ ∥ A − 1 ∥ . Its operator Michelson contrast and its operator logarithmic contrast [1 , 2 ] are
OMC ( A ) = κ ( A ) − 1 κ ( A ) + 1 and OLC ( A ) = 1 2 ln κ ( A ) = artanh OMC ( A ) ; \operatorname{OMC}(A)=\frac{\kappa(A)-1}{\kappa(A)+1}
\qquad\text{and}\qquad
\operatorname{OLC}(A)=\frac{1}{2}\ln\kappa(A)=\operatorname{artanh}\operatorname{OMC}(A); OMC ( A ) = κ ( A ) + 1 κ ( A ) − 1 and OLC ( A ) = 2 1 ln κ ( A ) = artanh OMC ( A ) ;
for positive A A A , OMC ( A ) = ( sup σ ( A ) − inf σ ( A ) ) / ( sup σ ( A ) + inf σ ( A ) ) \operatorname{OMC}(A)=(\sup\sigma(A)-\inf\sigma(A))/(\sup\sigma(A)+\inf\sigma(A)) OMC ( A ) = ( sup σ ( A ) − inf σ ( A )) / ( sup σ ( A ) + inf σ ( A )) . Replacing the operator norm by the Schatten norms of a von Neumann algebra with a faithful normal tracial state τ \tau τ gives the Schatten contrasts
C p ( A ) = 1 2 ln ( ∥ A ∥ p ∥ A − 1 ∥ p ) , ∥ X ∥ p = τ ( ∣ X ∣ p ) 1 / p . C_p(A)=\frac{1}{2}\ln\bigl(\lVert A\rVert_p\lVert A^{-1}\rVert_p\bigr),
\qquad
\lVert X\rVert_p=\tau\bigl(\lvert X\rvert^p\bigr)^{1/p}. C p ( A ) = 2 1 ln ( ∥ A ∥ p ∥ A − 1 ∥ p ) , ∥ X ∥ p = τ ( ∣ X ∣ p ) 1/ p .
All these quantities are finite for bounded invertible operators. If A A A or A − 1 A^{-1} A − 1 is unbounded, then OLC ( A ) = ∞ \operatorname{OLC}(A)=\infty OLC ( A ) = ∞ and OMC ( A ) = 1 \operatorname{OMC}(A)=1 OMC ( A ) = 1 . The question of this chapter is how such a contrast is infinite: which spectral information survives, and in which invariant form. We give three answers, one for operators affiliated with a finite von Neumann algebra (through the function p ↦ C p ( A ) p\mapsto C_p(A) p ↦ C p ( A ) ), one for unbounded operators on a Hilbert space (through the growth of the contrasts of finite sections, and through upper and lower limits of that growth), and one for bounded operators whose spectrum is read modulo compact operators.
The finite-section phase transition and the bounds for oscillating Weyl constants are the main quantitative results. The parts on moment thresholds and essential spectra organise classical Laplace-transform and quotient-algebra facts in the same notation. Precise assumptions and proofs are given in the corresponding sections.
Preliminaries
Throughout, τ \tau τ is a faithful normal tracial state on a von Neumann algebra M \mathcal{M} M , and M ~ \widetilde{\mathcal{M}} M denotes the ∗ * ∗ -algebra of closed, densely defined operators affiliated with M \mathcal{M} M ; as τ \tau τ is finite, each of them is τ \tau τ -measurable [30 , 31 ] . Let A ∈ M ~ A\in\widetilde{\mathcal{M}} A ∈ M be injective. Its right support equals 1 1 1 , and its left support, being equivalent to the right one in the finite algebra M \mathcal{M} M , equals 1 1 1 as well. Hence A A A has dense range, A − 1 ∈ M ~ A^{-1}\in\widetilde{\mathcal{M}} A − 1 ∈ M , the partial isometry U U U in the polar decomposition A = U ∣ A ∣ A=U\lvert A\rvert A = U ∣ A ∣ is unitary, and ∣ A − 1 ∣ = U ∣ A ∣ − 1 U ∗ \lvert A^{-1}\rvert=U\lvert A\rvert^{-1}U^* ∣ A − 1 ∣ = U ∣ A ∣ − 1 U ∗ . We call the Borel probability measure
ν A ( E ) = τ ( 1 E ( ln ∣ A ∣ ) ) \nu_A(E)=\tau(\mathbf{1}_E(\ln\lvert A\rvert)) ν A ( E ) = τ ( 1 E ( ln ∣ A ∣))
on R \mathbb{R} R the logarithmic spectral measure of A A A ; thus τ ( ∣ A ∣ p ) = ∫ e p x d ν A \tau(\lvert A\rvert^p)=\int e^{px}\,d\nu_A τ (∣ A ∣ p ) = ∫ e p x d ν A and τ ( ∣ A − 1 ∣ p ) = ∫ e − p x d ν A \tau(\lvert A^{-1}\rvert^p)=\int e^{-px}\,d\nu_A τ (∣ A − 1 ∣ p ) = ∫ e − p x d ν A , both in ( 0 , ∞ ] (0,\infty] ( 0 , ∞ ] . Each Borel probability measure on R \mathbb{R} R is of the form ν A \nu_A ν A , already for M = L ∞ [ 0 , 1 ] \mathcal{M}=L^\infty[0,1] M = L ∞ [ 0 , 1 ] . For a Borel probability measure ν \nu ν on R \mathbb{R} R let L ν ( t ) = ln ∫ e t x d ν ( x ) ∈ ( − ∞ , ∞ ] L_{\nu}(t)=\ln\int e^{tx}\,d\nu(x)\in(-\infty,\infty] L ν ( t ) = ln ∫ e t x d ν ( x ) ∈ ( − ∞ , ∞ ] , and for 0 < p < ∞ 0<p<\infty 0 < p < ∞ set
C p ( ν ) = L ν ( p ) + L ν ( − p ) 2 p ∈ [ 0 , ∞ ] , C ∞ ( ν ) = 1 2 diam supp ν ∈ [ 0 , ∞ ] . C_p(\nu)=\frac{L_{\nu}(p)+L_{\nu}(-p)}{2p}\in[0,\infty],\qquad C_{\infty}(\nu)=\frac{1}{2}\operatorname{diam}\operatorname{supp}\nu\in[0,\infty]. C p ( ν ) = 2 p L ν ( p ) + L ν ( − p ) ∈ [ 0 , ∞ ] , C ∞ ( ν ) = 2 1 diam supp ν ∈ [ 0 , ∞ ] .
For injective A ∈ M ~ A\in\widetilde{\mathcal{M}} A ∈ M we put C p ( A ) = C p ( ν A ) C_p(A)=C_p(\nu_A) C p ( A ) = C p ( ν A ) and OLC ( A ) = C ∞ ( ν A ) \operatorname{OLC}(A)=C_{\infty}(\nu_A) OLC ( A ) = C ∞ ( ν A ) ; with L A = L ν A L_A=L_{\nu_A} L A = L ν A this reads 2 p C p ( A ) = L A ( p ) + L A ( − p ) 2pC_p(A)=L_A(p)+L_A(-p) 2 p C p ( A ) = L A ( p ) + L A ( − p ) . The critical exponent of ν \nu ν is p c ( ν ) = sup { p > 0 : C p ( ν ) < ∞ } p_c(\nu)=\sup\{p>0:C_p(\nu)<\infty\} p c ( ν ) = sup { p > 0 : C p ( ν ) < ∞ } , where sup ∅ = 0 \sup\varnothing=0 sup ∅ = 0 .
Proposition 3.2.3 applies without a bounded-support assumption: it gives positivity, the exponential-moment criterion, monotonicity in p p p , the bounded-support estimate, and the dilation and convolution laws. We use those properties below for extended-valued contrasts.
Remark 4.2.1. At a fixed exponent, C p C_p C p admits no subadditivity estimate on injective elements of M ~ \widetilde{\mathcal{M}} M : for a positive A A A with p < p c ( ν A ) < 2 p p<p_c(\nu_A)<2p p < p c ( ν A ) < 2 p we have C p ( A ) < ∞ C_p(A)<\infty C p ( A ) < ∞ but C p ( A 2 ) = 2 C 2 p ( A ) = ∞ C_p(A^2)=2C_{2p}(A)=\infty C p ( A 2 ) = 2 C 2 p ( A ) = ∞ by Proposition 3.2.3 (e),(c) and the definition of p c p_c p c . If the exponents are combined harmonically, subadditivity holds with no additive correction term (Proposition 4.3.7 ).
Unbounded operators: critical exponents and growth
In this section A A A is an injective element of M ~ \widetilde{\mathcal{M}} M , possibly unbounded and with unbounded inverse, and C p ( A ) = C p ( ν A ) C_p(A)=C_p(\nu_A) C p ( A ) = C p ( ν A ) as in § 7.2 . We first treat the case OLC ( A ) = ∞ \operatorname{OLC}(A)=\infty OLC ( A ) = ∞ , p c ( ν A ) = ∞ p_c(\nu_A)=\infty p c ( ν A ) = ∞ , where the rate at which C p ( A ) C_p(A) C p ( A ) diverges is a quantitative substitute for the contrast, governed by the tails of ν A \nu_A ν A , and give first examples with finite p c p_c p c . In Subsections 4.3.2–4.3.4 we then consider the whole contrast function, its threshold p c p_c p c , and the transition at p c p_c p c .
Growth of the contrast, and first examples
Theorem 4.3.1 (Growth of the contrast). Let ν \nu ν be a Borel probability measure on R \mathbb{R} R with ∫ e p ∣ x ∣ d ν < ∞ \int e^{p|x|}\,d\nu<\infty ∫ e p ∣ x ∣ d ν < ∞ for all p > 0 p>0 p > 0 . Put M ( p ) = ln ∫ e p ∣ x ∣ d ν M(p)=\ln\int e^{p|x|}\,d\nu M ( p ) = ln ∫ e p ∣ x ∣ d ν and Θ ( u ) = ν ( { x : ∣ x ∣ ≥ u } ) \Theta(u)=\nu(\{x:|x|\ge u\}) Θ ( u ) = ν ({ x : ∣ x ∣ ≥ u }) .
(a) For all p > 0 p>0 p > 0 ,
M ( p ) − ln 2 2 p − 1 2 ∫ ∣ x ∣ d ν ≤ C p ( ν ) ≤ M ( p ) p . \frac{M(p)-\ln2}{2p}-\frac{1}{2}\int|x|\,d\nu\le C_p(\nu)\le\frac{M(p)}{p}. 2 p M ( p ) − ln 2 − 2 1 ∫ ∣ x ∣ d ν ≤ C p ( ν ) ≤ p M ( p ) . (b) Suppose there are q > 1 q>1 q > 1 , c 1 ≥ c 2 > 0 c_1\ge c_2>0 c 1 ≥ c 2 > 0 and u 0 ≥ 0 u_0\ge0 u 0 ≥ 0 such that e − c 1 u q ≤ Θ ( u ) ≤ e − c 2 u q e^{-c_1u^q}\le\Theta(u)\le e^{-c_2u^q} e − c 1 u q ≤ Θ ( u ) ≤ e − c 2 u q for all u ≥ u 0 u\ge u_0 u ≥ u 0 . Then, with η ( c ) = ( q − 1 ) q − q / ( q − 1 ) c − 1 / ( q − 1 ) \eta(c)=(q-1)q^{-q/(q-1)}c^{-1/(q-1)} η ( c ) = ( q − 1 ) q − q / ( q − 1 ) c − 1/ ( q − 1 ) ,
1 2 η ( c 1 ) ≤ lim inf p → ∞ C p ( ν ) p 1 / ( q − 1 ) ≤ lim sup p → ∞ C p ( ν ) p 1 / ( q − 1 ) ≤ η ( c 2 ) . \frac{1}{2}\eta(c_1)\le\liminf_{p\to\infty}\frac{C_p(\nu)}{p^{1/(q-1)}}\le\limsup_{p\to\infty}\frac{C_p(\nu)}{p^{1/(q-1)}}\le\eta(c_2). 2 1 η ( c 1 ) ≤ p → ∞ lim inf p 1/ ( q − 1 ) C p ( ν ) ≤ p → ∞ lim sup p 1/ ( q − 1 ) C p ( ν ) ≤ η ( c 2 ) .
Proof. (a) Since ∫ e ± p x d ν ≤ ∫ e p ∣ x ∣ d ν \int e^{\pm px}\,d\nu\le\int e^{p|x|}\,d\nu ∫ e ± p x d ν ≤ ∫ e p ∣ x ∣ d ν , the upper bound is clear. For the lower bound, e M ( p ) ≤ ∫ e p x d ν + ∫ e − p x d ν ≤ 2 max ( ∫ e p x d ν , ∫ e − p x d ν ) e^{M(p)}\le\int e^{px}\,d\nu+\int e^{-px}\,d\nu\le2\max\left(\int e^{px}\,d\nu,\int e^{-px}\,d\nu\right) e M ( p ) ≤ ∫ e p x d ν + ∫ e − p x d ν ≤ 2 max ( ∫ e p x d ν , ∫ e − p x d ν ) , while by Jensen’s inequality each of the two integrals is at least e − p ∫ ∣ x ∣ d ν e^{-p\int|x|\,d\nu} e − p ∫ ∣ x ∣ d ν . Hence ∫ e p x d ν ∫ e − p x d ν ≥ 1 2 e M ( p ) e − p ∫ ∣ x ∣ d ν \int e^{px}\,d\nu\int e^{-px}\,d\nu\ge\frac{1}{2}e^{M(p)}e^{-p\int|x|\,d\nu} ∫ e p x d ν ∫ e − p x d ν ≥ 2 1 e M ( p ) e − p ∫ ∣ x ∣ d ν .
(b) Let q ′ = q / ( q − 1 ) q'=q/(q-1) q ′ = q / ( q − 1 ) and S c ( p ) = sup u ≥ 0 ( p u − c u q ) = η ( c ) p q ′ S_c(p)=\sup_{u\ge0}(pu-cu^q)=\eta(c)p^{q'} S c ( p ) = sup u ≥ 0 ( p u − c u q ) = η ( c ) p q ′ , attained at u p = ( p / ( c q ) ) 1 / ( q − 1 ) u_p=(p/(cq))^{1/(q-1)} u p = ( p / ( c q ) ) 1/ ( q − 1 ) . For the lower bound, by Chebyshev’s inequality e M ( p ) ≥ e p u Θ ( u ) e^{M(p)}\ge e^{pu}\Theta(u) e M ( p ) ≥ e p u Θ ( u ) for every u ≥ 0 u\ge0 u ≥ 0 ; taking u = u p u=u_p u = u p with c = c 1 c=c_1 c = c 1 , which exceeds u 0 u_0 u 0 for large p p p , gives M ( p ) ≥ η ( c 1 ) p q ′ M(p)\ge\eta(c_1)p^{q'} M ( p ) ≥ η ( c 1 ) p q ′ . For the upper bound, by Fubini’s theorem
e M ( p ) = 1 + p ∫ 0 ∞ e p u ν ( { ∣ x ∣ > u } ) d u ≤ e p u 0 + p ∫ 0 ∞ e p u − c 2 u q d u . e^{M(p)}=1+p\int_0^\infty e^{pu}\nu(\{|x|>u\})\,du\le e^{pu_0}+p\int_0^\infty e^{pu-c_2u^q}\,du. e M ( p ) = 1 + p ∫ 0 ∞ e p u ν ({ ∣ x ∣ > u }) d u ≤ e p u 0 + p ∫ 0 ∞ e p u − c 2 u q d u .
Let U = ( 2 p / c 2 ) 1 / ( q − 1 ) U=(2p/c_2)^{1/(q-1)} U = ( 2 p / c 2 ) 1/ ( q − 1 ) . For u > U u>U u > U we have c 2 u q > 2 p u c_2u^q>2pu c 2 u q > 2 p u , so the integrand is at most e − p u e^{-pu} e − p u and this part of the integral is at most 1 / p 1/p 1/ p . For u ≤ U u\le U u ≤ U the integrand is at most e S c 2 ( p ) e^{S_{c_2}(p)} e S c 2 ( p ) . Hence e M ( p ) ≤ e p u 0 + 1 + p U e η ( c 2 ) p q ′ e^{M(p)}\le e^{pu_0}+1+pUe^{\eta(c_2)p^{q'}} e M ( p ) ≤ e p u 0 + 1 + p U e η ( c 2 ) p q ′ , and since q ′ > 1 q'>1 q ′ > 1 , M ( p ) ≤ η ( c 2 ) p q ′ + O ( ln p ) M(p)\le\eta(c_2)p^{q'}+O(\ln p) M ( p ) ≤ η ( c 2 ) p q ′ + O ( ln p ) . Inserting both estimates into (a) and using q ′ − 1 = 1 / ( q − 1 ) q'-1=1/(q-1) q ′ − 1 = 1/ ( q − 1 ) gives the claim. ∎
Thus a logarithmic spectrum with tails of order e − c u q e^{-cu^{q}} e − c u q gives C p ≍ p 1 / ( q − 1 ) C_p \asymp p^{1/(q-1)} C p ≍ p 1/ ( q − 1 ) : the lighter the tails, the slower the growth, and bounded spectra (q = ∞ q=\infty q = ∞ ) give bounded contrast. Sharper asymptotics for regularly varying tail exponents follow from Kasahara’s Tauberian theorem [50 Section 4.12] .
Example 4.3.2 (Log-normal operators). If ν A \nu_A ν A is the nondegenerate Gaussian law N ( m , σ 2 ) N(m,\sigma^2) N ( m , σ 2 ) , σ > 0 \sigma>0 σ > 0 , then L A ( t ) = m t + 1 2 σ 2 t 2 L_A(t)=mt+\frac{1}{2}\sigma^2t^2 L A ( t ) = m t + 2 1 σ 2 t 2 and
C p ( A ) = σ 2 p 2 ( 0 < p < ∞ ) , OLC ( A ) = ∞ . C_p(A)=\frac{\sigma^2p}{2}\qquad(0<p<\infty),\qquad\operatorname{OLC}(A)=\infty. C p ( A ) = 2 σ 2 p ( 0 < p < ∞ ) , OLC ( A ) = ∞.
Here A A A and A − 1 A^{-1} A − 1 are unbounded, all contrasts are finite, and the growth is linear. This is consistent with Theorem 4.3.1 : for q = 2 q=2 q = 2 the tail hypothesis holds with any c 1 > 1 2 σ 2 > c 2 c_1>\frac{1}{2\sigma^2}>c_2 c 1 > 2 σ 2 1 > c 2 , and η ( 1 2 σ 2 ) = σ 2 2 \eta\left(\frac{1}{2\sigma^2}\right)=\frac{\sigma^2}{2} η ( 2 σ 2 1 ) = 2 σ 2 , so the theorem gives σ 2 4 ≤ lim inf p C p / p ≤ lim sup p C p / p ≤ σ 2 2 \frac{\sigma^2}{4}\leq\liminf_p C_p/p\leq\limsup_p C_p/p\leq\frac{\sigma^2}{2} 4 σ 2 ≤ lim inf p C p / p ≤ lim sup p C p / p ≤ 2 σ 2 ; the upper bound is attained.
Example 4.3.3 (Finite critical exponent). If ν A \nu_A ν A has density 1 2 b e − ∣ x ∣ / b \frac{1}{2b}e^{-|x|/b} 2 b 1 e − ∣ x ∣/ b , then ∫ e ± p x d ν A = ( 1 − b 2 p 2 ) − 1 \int e^{\pm px}\,\mathrm{d}\nu_A=(1-b^2p^2)^{-1} ∫ e ± p x d ν A = ( 1 − b 2 p 2 ) − 1 for p < 1 / b p<1/b p < 1/ b , so C p ( A ) = − 1 p ln ( 1 − b 2 p 2 ) C_p(A)=-\frac{1}{p}\ln(1-b^2p^2) C p ( A ) = − p 1 ln ( 1 − b 2 p 2 ) for p < 1 / b p<1/b p < 1/ b and p c ( ν A ) = 1 / b p_c(\nu_A)=1/b p c ( ν A ) = 1/ b .
Example 4.3.4 (Free Poisson elements). Let A ∈ M A\in\mathcal{M} A ∈ M be positive with the Marchenko–Pastur distribution (a scaled free Poisson law) of parameter y ∈ ( 0 , 1 ] y\in(0,1] y ∈ ( 0 , 1 ] [52 , 49 ] , i.e. with density ( b − λ ) ( λ − a ) 2 π y λ \frac{\sqrt{(b-\lambda)(\lambda-a)}}{2\pi y\lambda} 2 π y λ ( b − λ ) ( λ − a ) on [ a , b ] [a,b] [ a , b ] , a = ( 1 − y ) 2 a=(1-\sqrt{y})^2 a = ( 1 − y ) 2 , b = ( 1 + y ) 2 b=(1+\sqrt{y})^2 b = ( 1 + y ) 2 . This is the limiting spectral distribution of sample covariance matrices 1 N X X ∗ \frac{1}{N}XX^* N 1 X X ∗ with X X X of size n × N n\times N n × N , n / N → y n/N\to y n / N → y . For y < 1 y<1 y < 1 , using τ ( A ) = 1 \tau(A)=1 τ ( A ) = 1 , τ ( A 2 ) = 1 + y \tau(A^2)=1+y τ ( A 2 ) = 1 + y , τ ( A − 1 ) = ( 1 − y ) − 1 \tau(A^{-1})=(1-y)^{-1} τ ( A − 1 ) = ( 1 − y ) − 1 and τ ( A − 2 ) = ( 1 − y ) − 3 \tau(A^{-2})=(1-y)^{-3} τ ( A − 2 ) = ( 1 − y ) − 3 ,
C 1 ( A ) = − 1 2 ln ( 1 − y ) , C 2 ( A ) = 1 4 ln 1 + y ( 1 − y ) 3 , OLC ( A ) = ln 1 + y 1 − y . C_1(A)=-\frac{1}{2}\ln(1-y),\qquad C_2(A)=\frac{1}{4}\ln\frac{1+y}{(1-y)^3},\qquad\operatorname{OLC}(A)=\ln\frac{1+\sqrt{y}}{1-\sqrt{y}}. C 1 ( A ) = − 2 1 ln ( 1 − y ) , C 2 ( A ) = 4 1 ln ( 1 − y ) 3 1 + y , OLC ( A ) = ln 1 − y 1 + y .
As y ↑ 1 y\uparrow1 y ↑ 1 all three diverge, at the different rates 1 2 ∣ ln ( 1 − y ) ∣ \frac{1}{2}|\ln(1-y)| 2 1 ∣ ln ( 1 − y ) ∣ , 3 4 ∣ ln ( 1 − y ) ∣ \frac{3}{4}|\ln(1-y)| 4 3 ∣ ln ( 1 − y ) ∣ and ∣ ln ( 1 − y ) ∣ |\ln(1-y)| ∣ ln ( 1 − y ) ∣ . At y = 1 y=1 y = 1 the operator A A A is injective and bounded with unbounded inverse, the density is 1 2 π ( 4 − λ ) / λ \frac{1}{2\pi}\sqrt{(4-\lambda)/\lambda} 2 π 1 ( 4 − λ ) / λ on ( 0 , 4 ] (0,4] ( 0 , 4 ] , and by the Beta integral and the duplication formula
τ ( A s ) = Γ ( 2 s + 1 ) Γ ( s + 1 ) Γ ( s + 2 ) ( s > − 1 2 ) , \tau(A^s)=\frac{\Gamma(2s+1)}{\Gamma(s+1)\Gamma(s+2)}\qquad\left(s>-\frac{1}{2}\right), τ ( A s ) = Γ ( s + 1 ) Γ ( s + 2 ) Γ ( 2 s + 1 ) ( s > − 2 1 ) ,
C p ( A ) = 1 2 p ln Γ ( 1 + 2 p ) Γ ( 1 − 2 p ) Γ ( 1 + p ) Γ ( 2 + p ) Γ ( 1 − p ) Γ ( 2 − p ) ( 0 < p < 1 2 ) . C_p(A)=\frac{1}{2p}\ln\frac{\Gamma(1+2p)\Gamma(1-2p)}{\Gamma(1+p)\Gamma(2+p)\Gamma(1-p)\Gamma(2-p)}\qquad\left(0<p<\frac{1}{2}\right). C p ( A ) = 2 p 1 ln Γ ( 1 + p ) Γ ( 2 + p ) Γ ( 1 − p ) Γ ( 2 − p ) Γ ( 1 + 2 p ) Γ ( 1 − 2 p ) ( 0 < p < 2 1 ) .
Hence OLC ( A ) = ∞ \operatorname{OLC}(A)=\infty OLC ( A ) = ∞ and p c ( ν A ) = 1 2 p_c(\nu_A)=\frac{1}{2} p c ( ν A ) = 2 1 : the critical exponent records the square-root singularity of the spectral density at the hard edge. More generally, for a bounded positive A A A whose spectral density behaves like λ α \lambda^\alpha λ α at 0 0 0 , α > − 1 \alpha>-1 α > − 1 , one has p c = α + 1 p_c=\alpha+1 p c = α + 1 . For y < 1 y<1 y < 1 the largest and the smallest eigenvalue of 1 N X X ∗ \frac{1}{N}XX^* N 1 X X ∗ converge almost surely to b b b and to a > 0 a>0 a > 0 under the moment assumptions of [53 , 54 ] , so the values above are the almost sure limits of the matrix contrasts.
The contrast function, and orders of growth and decay
If A A A or A − 1 A^{-1} A − 1 is unbounded, then OLC ( A ) = ∞ \operatorname{OLC}(A)=\infty OLC ( A ) = ∞ , but the whole function p ↦ C p ( A ) p\mapsto C_p(A) p ↦ C p ( A ) still carries spectral information. We collect this information in a few numbers defined through traces of powers only. Since C p ( A ) = C p ( ν A ) C_p(A)=C_p(\nu_A) C p ( A ) = C p ( ν A ) , they are characteristics of the probability measure ν A \nu_A ν A , and most of the facts below are classical properties of two-sided Laplace transforms [55 ] ; we state them because, read as statements about the contrast, they describe how an infinite contrast is infinite.
Definition 4.3.5. Let A ∈ M ~ A \in\widetilde{\mathcal M} A ∈ M be injective. The contrast function of A A A is p ↦ C p ( A ) ∈ [ 0 , ∞ ] p \mapsto C_p(A) \in[0,\infty] p ↦ C p ( A ) ∈ [ 0 , ∞ ] , p ∈ ( 0 , ∞ ] p \in(0,\infty] p ∈ ( 0 , ∞ ] . Put
p + ( A ) = sup { p ≥ 0 : τ ( ∣ A ∣ p ) < ∞ } , p − ( A ) = sup { p ≥ 0 : τ ( ∣ A − 1 ∣ p ) < ∞ } . p_+(A)=\sup\{p\ge0:\tau(|A|^p)<\infty\},\qquad p_-(A)=\sup\{p\ge0:\tau(|A^{-1}|^p)<\infty\}. p + ( A ) = sup { p ≥ 0 : τ ( ∣ A ∣ p ) < ∞ } , p − ( A ) = sup { p ≥ 0 : τ ( ∣ A − 1 ∣ p ) < ∞ } . and call ω + ( A ) = 1 / p + ( A ) \omega_+(A)=1/p_+(A) ω + ( A ) = 1/ p + ( A ) and ω − ( A ) = 1 / p − ( A ) \omega_-(A)=1/p_-(A) ω − ( A ) = 1/ p − ( A ) (with 1 / 0 = ∞ 1/0=\infty 1/0 = ∞ , 1 / ∞ = 0 1/\infty=0 1/∞ = 0 ) the order of growth and the order of decay of A A A .
Thus τ ( ∣ A ∣ s ) = ∫ e s x d ν A \tau(|A|^s)=\int e^{sx}\,\mathrm{d}\nu_A τ ( ∣ A ∣ s ) = ∫ e s x d ν A is finite for − p − ( A ) < s < p + ( A ) -p_-(A)<s<p_+(A) − p − ( A ) < s < p + ( A ) and infinite for s > p + ( A ) s>p_+(A) s > p + ( A ) and for s < − p − ( A ) s<-p_-(A) s < − p − ( A ) : the two orders are the reciprocal abscissae of convergence of the two-sided Laplace transform of ν A \nu_A ν A , i.e. of the Mellin transform of the spectral distribution of ∣ A ∣ |A| ∣ A ∣ . By Proposition 3.2.3(b),
p c ( ν A ) = min { p + ( A ) , p − ( A ) } , 1 p c ( ν A ) = max { ω + ( A ) , ω − ( A ) } . (6) p_c(\nu_A)=\min\{p_+(A),p_-(A)\},\qquad\frac{1}{p_c(\nu_A)}=\max\{\omega_+(A),\omega_-(A)\}. \tag*{(6)} p c ( ν A ) = min { p + ( A ) , p − ( A )} , p c ( ν A ) 1 = max { ω + ( A ) , ω − ( A )} . (6)
Bounded operators have ω + = 0 \omega_+=0 ω + = 0 , operators with bounded inverse have ω − = 0 \omega_-=0 ω − = 0 , and ω + ( A − 1 ) = ω − ( A ) \omega_+(A^{-1})=\omega_-(A) ω + ( A − 1 ) = ω − ( A ) , since ν A − 1 \nu_{A^{-1}} ν A − 1 is the reflection of ν A \nu_A ν A .
Proposition 4.3.6. Let A ∈ M ~ A \in\widetilde{\mathcal M} A ∈ M be injective.
(a) One has
p + ( A ) = lim inf x → ∞ − ln ν A ( ( x , ∞ ) ) x , p_+(A)=\liminf_{x\to\infty}\frac{-\ln\nu_A((x,\infty))}{x}, p + ( A ) = x → ∞ lim inf x − ln ν A (( x , ∞ )) , p − ( A ) = lim inf x → ∞ − ln ν A ( ( − ∞ , − x ) ) x , p_-(A)=\liminf_{x\to\infty}\frac{-\ln\nu_A((-\infty,-x))}{x}, p − ( A ) = x → ∞ lim inf x − ln ν A (( − ∞ , − x )) , where − ln 0 = + ∞ -\ln0=+\infty − ln 0 = + ∞ .
(b) For λ ∈ C ∖ { 0 } \lambda\in\mathbb C\setminus\{0\} λ ∈ C ∖ { 0 } , unitaries U , V ∈ M U,V\in\mathcal M U , V ∈ M and s > 0 s>0 s > 0 : p ± ( λ U A V ) = p ± ( A ) p_\pm(\lambda UAV)=p_\pm(A) p ± ( λ U A V ) = p ± ( A ) , p ± ( A − 1 ) = p ∓ ( A ) p_\pm(A^{-1})=p_\mp(A) p ± ( A − 1 ) = p ∓ ( A ) and p ± ( ∣ A ∣ s ) = p ± ( A ) / s p_\pm(|A|^s)=p_\pm(A)/s p ± ( ∣ A ∣ s ) = p ± ( A ) / s .
Proof. (a) This is the classical expression of the abscissa of convergence of a Laplace–Stieltjes transform through the decay of the tail [55 Ch. II] ; we include the short argument. Write ν = ν A \nu=\nu_A ν = ν A , M ( s ) = ∫ e s x d ν M(s)=\int e^{sx}\,\mathrm{d}\nu M ( s ) = ∫ e s x d ν and let σ \sigma σ be the lower limit in the first formula. If 0 < s < σ 0<s<\sigma 0 < s < σ , choose s < s ′ < σ s<s'<\sigma s < s ′ < σ and x 0 > 0 x_0>0 x 0 > 0 with ν ( ( x , ∞ ) ) ≤ e − s ′ x \nu((x,\infty))\le e^{-s'x} ν (( x , ∞ )) ≤ e − s ′ x for x ≥ x 0 x\ge x_0 x ≥ x 0 . Then
M ( s ) ≤ e s x 0 + ∫ ( x 0 , ∞ ) e s x d ν ( x ) = e s x 0 + e s x 0 ν ( ( x 0 , ∞ ) ) + ∫ x 0 ∞ s e s y ν ( ( y , ∞ ) ) d y < ∞ , M(s)\le e^{sx_0}+\int_{(x_0,\infty)}e^{sx}\,\mathrm{d}\nu(x)=e^{sx_0}+e^{sx_0}\nu((x_0,\infty))+\int_{x_0}^{\infty}se^{sy}\nu((y,\infty))\,\mathrm{d}y<\infty, M ( s ) ≤ e s x 0 + ∫ ( x 0 , ∞ ) e s x d ν ( x ) = e s x 0 + e s x 0 ν (( x 0 , ∞ )) + ∫ x 0 ∞ s e sy ν (( y , ∞ )) d y < ∞ ,
so p + ( A ) ≥ σ p_+(A)\ge\sigma p + ( A ) ≥ σ . Conversely, if M ( s ) < ∞ M(s)<\infty M ( s ) < ∞ , then e s x ν ( ( x , ∞ ) ) ≤ M ( s ) e^{sx}\nu((x,\infty))\le M(s) e s x ν (( x , ∞ )) ≤ M ( s ) for x > 0 x>0 x > 0 by Chebyshev’s inequality, hence − 1 x ln ν ( ( x , ∞ ) ) ≥ s − 1 x ln M ( s ) -\frac{1}{x}\ln\nu((x,\infty))\ge s-\frac{1}{x}\ln M(s) − x 1 ln ν (( x , ∞ )) ≥ s − x 1 ln M ( s ) and σ ≥ s \sigma\ge s σ ≥ s ; thus σ ≥ p + ( A ) \sigma\ge p_+(A) σ ≥ p + ( A ) . The second formula follows by applying the first to A − 1 A^{-1} A − 1 . (b) ∣ λ U A V ∣ = ∣ λ ∣ V ∗ ∣ A ∣ V |\lambda UAV|=|\lambda|V^*|A|V ∣ λ U A V ∣ = ∣ λ ∣ V ∗ ∣ A ∣ V has the spectral distribution of ∣ λ ∣ ∣ A ∣ |\lambda||A| ∣ λ ∣∣ A ∣ , which translates ν A \nu_A ν A ; ∣ A − 1 ∣ |A^{-1}| ∣ A − 1 ∣ has the spectral distribution of ∣ A ∣ − 1 |A|^{-1} ∣ A ∣ − 1 ; and ν ∣ A ∣ s \nu_{|A|^s} ν ∣ A ∣ s is the image of ν A \nu_A ν A under x ↦ s x x\mapsto sx x ↦ s x .
With the distribution function d A ( λ ) = τ ( 1 ( λ , ∞ ) ( ∣ A ∣ ) ) d_A(\lambda)=\tau(1_{(\lambda,\infty)}(|A|)) d A ( λ ) = τ ( 1 ( λ , ∞ ) ( ∣ A ∣ )) , part (a) says that
p + ( A ) = lim inf λ → ∞ ln ( 1 / d A ( λ ) ) ln λ . p_+(A)=\liminf_{\lambda\to\infty}\frac{\ln(1/d_A(\lambda))}{\ln\lambda}. p + ( A ) = λ → ∞ lim inf ln λ ln ( 1/ d A ( λ )) .
Inverting the distribution function one finds equivalently
ω + ( A ) = lim sup t → 0 ln μ t ( A ) ln ( 1 / t ) , ω − ( A ) = lim sup t → 0 ln μ t ( A − 1 ) ln ( 1 / t ) , \omega_+(A)=\limsup_{t\to0}\frac{\ln\mu_t(A)}{\ln(1/t)},\qquad\omega_-(A)=\limsup_{t\to0}\frac{\ln\mu_t(A^{-1})}{\ln(1/t)}, ω + ( A ) = t → 0 lim sup ln ( 1/ t ) ln μ t ( A ) , ω − ( A ) = t → 0 lim sup ln ( 1/ t ) ln μ t ( A − 1 ) ,
in terms of the generalised singular values μ t \mu_t μ t of [31 ] ; this explains the names.
Subadditivity with harmonic exponents
By Remark 4.2.1, C p C_p C p is not subadditive at a fixed exponent. It is subadditive, with no additive correction term, if the exponents are combined as in Hölder’s inequality.
Proposition 4.3.7. Let A , B ∈ M A, B \in\mathcal{M} A , B ∈ M be injective and p , q ∈ ( 0 , ∞ ] p, q \in(0,\infty] p , q ∈ ( 0 , ∞ ] , and define r r r by
1 r = 1 p + 1 q . \frac{1}{r}=\frac{1}{p}+\frac{1}{q}. r 1 = p 1 + q 1 . Then
C r ( A B ) ≤ C p ( A ) + C q ( B ) in [ 0 , ∞ ] , C_r(AB) \leq C_p(A)+C_q(B) \qquad\text{in } [0,\infty], C r ( A B ) ≤ C p ( A ) + C q ( B ) in [ 0 , ∞ ] , with C ∞ = O L C C_\infty=\mathrm{OLC} C ∞ = OLC . Equality holds if A ≥ 0 A\geq0 A ≥ 0 , B = A θ B=A^\theta B = A θ for some θ > 0 \theta>0 θ > 0 and q = p / θ q=p/\theta q = p / θ ; in particular if A = B ≥ 0 A=B\geq0 A = B ≥ 0 and p = q p=q p = q . Consequently
ω ± ( A B ) ≤ ω ± ( A ) + ω ± ( B ) , 1 p c ( ν A B ) ≤ 1 p c ( ν A ) + 1 p c ( ν B ) , \omega_\pm(AB)\leq\omega_\pm(A)+\omega_\pm(B), \qquad\frac{1}{p_c(\nu_{AB})}\leq\frac{1}{p_c(\nu_A)}+\frac{1}{p_c(\nu_B)}, ω ± ( A B ) ≤ ω ± ( A ) + ω ± ( B ) , p c ( ν A B ) 1 ≤ p c ( ν A ) 1 + p c ( ν B ) 1 , and, for p = q = ∞ p=q=\infty p = q = ∞ , O L C ( A B ) ≤ O L C ( A ) + O L C ( B ) \mathrm{OLC}(AB)\leq\mathrm{OLC}(A)+\mathrm{OLC}(B) OLC ( A B ) ≤ OLC ( A ) + OLC ( B ) .
Proof. For τ \tau τ -measurable X , Y X,Y X , Y and 0 < r < ∞ 0<r<\infty 0 < r < ∞ , the majorisation theorem of Fack and Kosaki [31 Theorem 4.2] , applied to f ( t ) = t r f(t)=t^r f ( t ) = t r , gives
τ ( ∣ X Y ∣ r ) = ∫ 0 1 μ t ( X Y ) r d t ≤ ∫ 0 1 μ t ( X ) r μ t ( Y ) r d t , \tau\left(|XY|^r\right)=\int_0^1\mu_t(XY)^r\,\mathrm{d}t\leq\int_0^1\mu_t(X)^r\mu_t(Y)^r\,\mathrm{d}t, τ ( ∣ X Y ∣ r ) = ∫ 0 1 μ t ( X Y ) r d t ≤ ∫ 0 1 μ t ( X ) r μ t ( Y ) r d t ,
and the classical Hölder inequality on ( 0 , 1 ) (0,1) ( 0 , 1 ) yields ∥ X Y ∥ r ≤ ∥ X ∥ p ∥ Y ∥ q \|XY\|_r\leq\|X\|_p\|Y\|_q ∥ X Y ∥ r ≤ ∥ X ∥ p ∥ Y ∥ q for all 0 < p , q ≤ ∞ 0<p,q\leq\infty 0 < p , q ≤ ∞ (see also [32 ] ); for p = q = ∞ p=q=\infty p = q = ∞ this is ∥ X Y ∥ ≤ ∥ X ∥ ∥ Y ∥ \|XY\|\leq\|X\|\|Y\| ∥ X Y ∥ ≤ ∥ X ∥∥ Y ∥ . Here A B AB A B is the strong product, which is injective with ( A B ) − 1 = B − 1 A − 1 (AB)^{-1}=B^{-1}A^{-1} ( A B ) − 1 = B − 1 A − 1 in M \mathcal{M} M . Applying the inequality to A B AB A B and to B − 1 A − 1 B^{-1}A^{-1} B − 1 A − 1 ,
2 C r ( A B ) = ln ∥ A B ∥ r + ln ∥ B − 1 A − 1 ∥ r ≤ ln ∥ A ∥ p + ln ∥ B ∥ q + ln ∥ B − 1 ∥ q + ln ∥ A − 1 ∥ p = 2 C p ( A ) + 2 C q ( B ) . \begin{aligned}
2C_r(AB)&=\ln\|AB\|_r+\ln\|B^{-1}A^{-1}\|_r\\
&\leq\ln\|A\|_p+\ln\|B\|_q+\ln\|B^{-1}\|_q+\ln\|A^{-1}\|_p=2C_p(A)+2C_q(B).
\end{aligned} 2 C r ( A B ) = ln ∥ A B ∥ r + ln ∥ B − 1 A − 1 ∥ r ≤ ln ∥ A ∥ p + ln ∥ B ∥ q + ln ∥ B − 1 ∥ q + ln ∥ A − 1 ∥ p = 2 C p ( A ) + 2 C q ( B ) .
If A ≥ 0 A\geq0 A ≥ 0 , B = A θ B=A^\theta B = A θ and q = p / θ q=p/\theta q = p / θ , then A B = A 1 + θ AB=A^{1+\theta} A B = A 1 + θ and r = p / ( 1 + θ ) r=p/(1+\theta) r = p / ( 1 + θ ) , and by Proposition 3.2.3 (e) both sides equal ( 1 + θ ) C p ( A ) (1+\theta)C_p(A) ( 1 + θ ) C p ( A ) . For the orders, if p < p + ( A ) p<p_+(A) p < p + ( A ) and q < p + ( B ) q<p_+(B) q < p + ( B ) then τ ( ∣ A B ∣ r ) < ∞ \tau(|AB|^r)<\infty τ ( ∣ A B ∣ r ) < ∞ , so p + ( A B ) ≥ r p_+(AB)\geq r p + ( A B ) ≥ r , that is, ω + ( A B ) ≤ 1 p + 1 q \omega_+(AB)\leq\frac{1}{p}+\frac{1}{q} ω + ( A B ) ≤ p 1 + q 1 ; letting p ↑ p + ( A ) p\uparrow p_+(A) p ↑ p + ( A ) , q ↑ p + ( B ) q\uparrow p_+(B) q ↑ p + ( B ) gives the inequality for ω + \omega_+ ω + . The one for ω − \omega_- ω − follows by applying it to B − 1 A − 1 B^{-1}A^{-1} B − 1 A − 1 , and the last one from (6) . □
The bound on the orders is the contrast form of the classical inclusion L p ⊂ L q ⊂ L r L^p\subset L^q\subset L^r L p ⊂ L q ⊂ L r . Thus ω ± \omega_\pm ω ± and max { ω + , ω − } = 1 / p c \max\{\omega_+,\omega_-\}=1/p_c max { ω + , ω − } = 1/ p c have the formal properties of OLC recorded in [2 ] : they are subadditive under products and homogeneous under powers of positive operators (Proposition 4.3.6 (b)), max { ω + , ω − } \max\{\omega_+,\omega_-\} max { ω + , ω − } is invariant under A ↦ A − 1 A\mapsto A^{-1} A ↦ A − 1 , and ω + \omega_+ ω + and ω − \omega_- ω − both vanish, i.e. max { ω + , ω − } = 0 \max\{\omega_+,\omega_-\}=0 max { ω + , ω − } = 0 , exactly when p c = ∞ p_c=\infty p c = ∞ , in particular on invertible elements of M \mathcal{M} M . For all injective A , B A,B A , B , the case p = q = 2 r p=q=2r p = q = 2 r gives C r ( A B ) ≤ C 2 r ( A ) + C 2 r ( B ) C_r(AB)\leq C_{2r}(A)+C_{2r}(B) C r ( A B ) ≤ C 2 r ( A ) + C 2 r ( B ) .
The transition at the threshold
Proposition 4.3.8. Let A ∈ M A\in\mathcal{M} A ∈ M be injective and p c = p c ( ν A ) p_c=p_c(\nu_A) p c = p c ( ν A ) .
(a) The function p ↦ C p ( A ) p\mapsto C_p(A) p ↦ C p ( A ) is finite and real-analytic on ( 0 , p c ) (0,p_c) ( 0 , p c ) , equal to + ∞ +\infty + ∞ on ( p c , ∞ ) (p_c,\infty) ( p c , ∞ ) , nondecreasing on ( 0 , ∞ ] (0,\infty] ( 0 , ∞ ] , and left-continuous at every p ∈ ( 0 , ∞ ] p\in(0,\infty] p ∈ ( 0 , ∞ ] as a [ 0 , ∞ ] [0,\infty] [ 0 , ∞ ] -valued function. If 0 < p c < ∞ 0<p_c<\infty 0 < p c < ∞ , the restriction of p ↦ C p ( A ) p\mapsto C_p(A) p ↦ C p ( A ) to ( 0 , p c ) (0,p_c) ( 0 , p c ) has no real-analytic extension to any interval ( p c − ε , p c + ε ) (p_c-\varepsilon,p_c+\varepsilon) ( p c − ε , p c + ε ) .
(b) If 0 < p c < ∞ 0<p_c<\infty 0 < p c < ∞ , exactly one of the following holds:
(i) blow-up: C p c ( A ) = ∞ C_{p_c}(A)=\infty C p c ( A ) = ∞ , and C p ( A ) ↑ ∞ C_p(A)\uparrow\infty C p ( A ) ↑ ∞ as p ↑ p c p\uparrow p_c p ↑ p c ;
(ii) jump: C p c ( A ) < ∞ C_{p_c}(A)<\infty C p c ( A ) < ∞ , while C p ( A ) = ∞ C_p(A)=\infty C p ( A ) = ∞ for every p > p c p>p_c p > p c .
The jump occurs if and only if ∣ A ∣ |A| ∣ A ∣ and ∣ A − 1 ∣ |A^{-1}| ∣ A − 1 ∣ both belong to L p c ( M , τ ) L^{p_c}(\mathcal{M},\tau) L p c ( M , τ ) .
(c) In the case of a jump, let k ≥ 1 k \ge1 k ≥ 1 . The function p ↦ C p ( A ) p \mapsto C_p(A) p ↦ C p ( A ) has finite left derivatives of all orders ≤ k \le k ≤ k at p c p_c p c if and only if ∫ ∣ x ∣ k e p c ∣ x ∣ d ν A ( x ) < ∞ \int|x|^k e^{p_c|x|}\,d\nu_A(x) < \infty ∫ ∣ x ∣ k e p c ∣ x ∣ d ν A ( x ) < ∞ . If k 0 k_0 k 0 is the smallest k k k for which this integral is infinite, then d k 0 d p k 0 C p ( A ) → + ∞ \frac{d^{k_0}}{dp^{k_0}}C_p(A) \to+\infty d p k 0 d k 0 C p ( A ) → + ∞ as p ↑ p c p \uparrow p_c p ↑ p c , and the left derivative of order k 0 k_0 k 0 at p c p_c p c is + ∞ +\infty + ∞ .
Proof. Write ν = ν A \nu=\nu_A ν = ν A , G ± ( p ) = ∫ e ± p x d ν G_{\pm}(p)=\int e^{\pm px}\,d\nu G ± ( p ) = ∫ e ± p x d ν and F = ln G + + ln G − F=\ln G_+ + \ln G_- F = ln G + + ln G − , so that C p ( A ) = F ( p ) / ( 2 p ) C_p(A)=F(p)/(2p) C p ( A ) = F ( p ) / ( 2 p ) .
(a) Finiteness on ( 0 , p c ) (0,p_c) ( 0 , p c ) , infiniteness on ( p c , ∞ ) (p_c,\infty) ( p c , ∞ ) and monotonicity are Proposition 3.2.3(b),(c) and the definition of p c p_c p c . The two-sided Laplace transform s ↦ ∫ e s x d ν s\mapsto\int e^{sx}\,d\nu s ↦ ∫ e s x d ν is holomorphic in the strip − p − ( A ) < Re s < p + ( A ) -p_-(A)<\operatorname{Re}s<p_+(A) − p − ( A ) < Re s < p + ( A ) and positive on its real points [55 Ch. VI] , so F F F is real-analytic on ( 0 , p c ) (0,p_c) ( 0 , p c ) . Moreover e F ( p ) = ∫ e p x d ν ∫ e − p y d ν = ∫ e p z d ( ν ∗ ν ˇ ) ( z ) e^{F(p)}=\int e^{px}\,d\nu\int e^{-py}\,d\nu=\int e^{pz}\,d(\nu*\check{\nu})(z) e F ( p ) = ∫ e p x d ν ∫ e − p y d ν = ∫ e p z d ( ν ∗ ν ˇ ) ( z ) , where ν ˇ \check{\nu} ν ˇ is the reflection of ν \nu ν , is the Laplace transform of the positive measure ν ∗ ν ˇ \nu*\check{\nu} ν ∗ ν ˇ , and its abscissa of convergence is p c p_c p c . By Landau’s theorem [55 Ch. II, Thm. 5b] , the real point of the abscissa of convergence of the Laplace transform of a positive measure is a singular point (apply it to the restriction of ν ∗ ν ˇ \nu*\check{\nu} ν ∗ ν ˇ to [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) ; the transform of the restriction to ( − ∞ , 0 ) (-\infty,0) ( − ∞ , 0 ) is holomorphic near p c p_c p c ); hence e F e^F e F , and with it C p = F / ( 2 p ) C_p=F/(2p) C p = F / ( 2 p ) , has no analytic continuation to a neighbourhood of p c p_c p c . For left-continuity at p 0 ∈ ( 0 , ∞ ) p_0\in(0,\infty) p 0 ∈ ( 0 , ∞ ) let p ↑ p 0 p\uparrow p_0 p ↑ p 0 : on { x > 0 } \{x>0\} { x > 0 } the integrand e p x e^{px} e p x increases to e p 0 x e^{p_0x} e p 0 x , and on { x ≤ 0 } \{x\le0\} { x ≤ 0 } it is bounded by 1 1 1 and converges; by monotone and dominated convergence G + ( p ) → G + ( p 0 ) ∈ ( 0 , ∞ ] G_+(p)\to G_+(p_0)\in(0,\infty] G + ( p ) → G + ( p 0 ) ∈ ( 0 , ∞ ] , and likewise for G − G_- G − . At p 0 = ∞ p_0=\infty p 0 = ∞ this is Proposition 3.2.3(c).
(b) This is (a) at p = p c p=p_c p = p c together with Proposition 3.2.3(b): C p c ( A ) < ∞ C_{p_c}(A)<\infty C p c ( A ) < ∞ iff ∫ e p c ∣ x ∣ d ν < ∞ \int e^{p_c|x|}\,d\nu<\infty ∫ e p c ∣ x ∣ d ν < ∞ iff τ ( ∣ A ∣ p c ) < ∞ \tau(|A|^{p_c})<\infty τ ( ∣ A ∣ p c ) < ∞ and τ ( ∣ A − 1 ∣ p c ) < ∞ \tau(|A^{-1}|^{p_c})<\infty τ ( ∣ A − 1 ∣ p c ) < ∞ .
(c) Assume C p c ( A ) < ∞ C_{p_c}(A)<\infty C p c ( A ) < ∞ and put I = [ p c / 2 , p c ] I=[p_c/2,p_c] I = [ p c /2 , p c ] . For p ∈ I p\in I p ∈ I , j ≥ 0 j\ge0 j ≥ 0 and all x x x , ∣ x ∣ j e ± p x ≤ ∣ x ∣ j ( 1 + e p c ∣ x ∣ ) |x|^j e^{\pm px}\le|x|^j(1+e^{p_c|x|}) ∣ x ∣ j e ± p x ≤ ∣ x ∣ j ( 1 + e p c ∣ x ∣ ) , and ∫ ∣ x ∣ j d ν < ∞ \int|x|^j\,d\nu<\infty ∫ ∣ x ∣ j d ν < ∞ because e p c ∣ x ∣ / 2 e^{p_c|x|/2} e p c ∣ x ∣/2 is integrable. Hence, if ∫ ∣ x ∣ k e p c ∣ x ∣ d ν < ∞ \int|x|^k e^{p_c|x|}\,d\nu<\infty ∫ ∣ x ∣ k e p c ∣ x ∣ d ν < ∞ , dominated convergence shows that G ± ( j ) ( p ) = ∫ ( ± x ) j e ± p x d ν G_{\pm}^{(j)}(p)=\int(\pm x)^j e^{\pm px}\,d\nu G ± ( j ) ( p ) = ∫ ( ± x ) j e ± p x d ν extend continuously to I I I for j ≤ k j\le k j ≤ k ; since G ± > 0 G_{\pm}>0 G ± > 0 on I I I , so do the derivatives of F F F and of F ( p ) / ( 2 p ) F(p)/(2p) F ( p ) / ( 2 p ) up to order k k k , and by the mean value theorem these limits are the left derivatives at p c p_c p c . Conversely, let k 0 k_0 k 0 be the smallest order with ∫ ∣ x ∣ k 0 e p c ∣ x ∣ d ν = ∞ \int|x|^{k_0}e^{p_c|x|}\,d\nu=\infty ∫ ∣ x ∣ k 0 e p c ∣ x ∣ d ν = ∞ . By the first part applied with k 0 − 1 k_0-1 k 0 − 1 , the derivatives of order < k 0 <k_0 < k 0 extend continuously to I I I . Split
G + ( k 0 ) ( p ) = ∫ x > 0 x k 0 e p x d ν + ∫ x < 0 x k 0 e p x d ν . G_+^{(k_0)}(p)=\int_{x>0}x^{k_0}e^{px}\,d\nu+\int_{x<0}x^{k_0}e^{px}\,d\nu. G + ( k 0 ) ( p ) = ∫ x > 0 x k 0 e p x d ν + ∫ x < 0 x k 0 e p x d ν .
G − ( k 0 ) ( p ) = ∫ x < 0 ∣ x ∣ k 0 e p ∣ x ∣ d ν + ∫ x > 0 ( − x ) k 0 e − p x d ν . G_-^{(k_0)}(p)=\int_{x<0}|x|^{k_0}e^{p|x|}\,d\nu+\int_{x>0}(-x)^{k_0}e^{-px}\,d\nu. G − ( k 0 ) ( p ) = ∫ x < 0 ∣ x ∣ k 0 e p ∣ x ∣ d ν + ∫ x > 0 ( − x ) k 0 e − p x d ν .
The second integrals are bounded on I I I by ∫ ∣ x ∣ k 0 d ν \int|x|^{k_0}\,d\nu ∫ ∣ x ∣ k 0 d ν . The first ones are nonnegative and nondecreasing in p p p , and by monotone convergence their sum tends, as p ↑ p c p\uparrow p_c p ↑ p c , to ∫ ∣ x ∣ k 0 e p c ∣ x ∣ d ν = ∞ \int|x|^{k_0}e^{p_c|x|}\,d\nu=\infty ∫ ∣ x ∣ k 0 e p c ∣ x ∣ d ν = ∞ . Now F ( k 0 ) = G + ( k 0 ) / G + + G − ( k 0 ) / G − + R F^{(k_0)}=G_+^{(k_0)}/G_+ + G_-^{(k_0)}/G_- + R F ( k 0 ) = G + ( k 0 ) / G + + G − ( k 0 ) / G − + R , where R R R is a polynomial in G ± ( j ) / G ± G_{\pm}^{(j)}/G_{\pm} G ± ( j ) / G ± , j < k 0 j<k_0 j < k 0 , and therefore bounded on I I I ; since G ± G_{\pm} G ± stay between positive constants on I I I , F ( k 0 ) → + ∞ F^{(k_0)}\to+\infty F ( k 0 ) → + ∞ . Further
d k 0 d p k 0 F ( p ) 2 p = F ( k 0 ) ( p ) 2 p + ∑ j < k 0 c j p − ( k 0 − j ) − 1 F ( j ) ( p ) \frac{d^{k_0}}{dp^{k_0}}\frac{F(p)}{2p}=\frac{F^{(k_0)}(p)}{2p}+\sum_{j<k_0}c_jp^{-(k_0-j)-1}F^{(j)}(p) d p k 0 d k 0 2 p F ( p ) = 2 p F ( k 0 ) ( p ) + j < k 0 ∑ c j p − ( k 0 − j ) − 1 F ( j ) ( p )
with bounded second term, so it tends to + ∞ +\infty + ∞ . Finally, the derivative of order k 0 − 1 k_0-1 k 0 − 1 extends continuously to p c p_c p c , the extension is its left derivative there, and by the mean value theorem its left difference quotients at p c p_c p c tend to + ∞ +\infty + ∞ . ▫
So, when 0 < p c < ∞ 0<p_c<\infty 0 < p c < ∞ , the passage from finite to infinite contrast happens at a single point, which is always a singularity, and comes in a graded scale: a blow-up, a jump with infinite slope, a jump with finite slope but infinite curvature, and so on. This is the behaviour of a power series with nonnegative coefficients, such as ∑ n n − δ z n \sum_n n^{-\delta}z^n ∑ n n − δ z n , at the positive point of its circle of convergence. In terms of noncommutative L p L^p L p -spaces: a jump means that ∣ A ∣ |A| ∣ A ∣ and ∣ A − 1 ∣ |A^{-1}| ∣ A − 1 ∣ lie in L p c L^{p_c} L p c , while for every p > p c p>p_c p > p c at least one of them is not in L p L^p L p ; a blow-up means that both lie in every L p L^p L p with p < p c p<p_c p < p c , while at least one of them is not in L p c L^{p_c} L p c . The relation between tail decay and singular behaviour at the abscissa is studied in detail for Laplace–Stieltjes transforms in [56 ] .
Example 4.3.9 (The scale of transitions). Let δ ≥ 0 \delta\ge0 δ ≥ 0 and let A ≥ 1 A \ge1 A ≥ 1 have the logarithmic spectral measure ν = ν A \nu=\nu_A ν = ν A on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) with ν ( ( x , ∞ ) ) = e − x ( 1 + x ) − δ \nu((x,\infty))=e^{-x}(1+x)^{-\delta} ν (( x , ∞ )) = e − x ( 1 + x ) − δ , x ≥ 0 x\ge0 x ≥ 0 (such A A A exists in every diffuse finite von Neumann algebra). Then A − 1 A^{-1} A − 1 is bounded, p − ( A ) = ∞ p_-(A)=\infty p − ( A ) = ∞ , and p c = p + ( A ) = 1 p_c=p_+(A)=1 p c = p + ( A ) = 1 by Proposition 4.3.6 (a), so ω + ( A ) = 1 \omega_+(A)=1 ω + ( A ) = 1 and ω − ( A ) = 0 \omega_-(A)=0 ω − ( A ) = 0 . Integrating by parts,
∫ e p x d ν = 1 + p ∫ 0 ∞ e − ( 1 − p ) x ( 1 + x ) − δ d x , \int e^{px}\,d\nu=1+p\int_0^\infty e^{-(1-p)x}(1+x)^{-\delta}\,dx, ∫ e p x d ν = 1 + p ∫ 0 ∞ e − ( 1 − p ) x ( 1 + x ) − δ d x ,
1 ≥ ∫ e − p x d ν ≥ ∫ e − x d ν > 0 ( 0 < p ≤ 1 ) . 1\ge\int e^{-px}\,d\nu\ge\int e^{-x}\,d\nu>0\qquad(0<p\le1). 1 ≥ ∫ e − p x d ν ≥ ∫ e − x d ν > 0 ( 0 < p ≤ 1 ) .
(a) If 0 ≤ δ < 1 0\le\delta<1 0 ≤ δ < 1 , the substitution x = y / ( 1 − p ) x=y/(1-p) x = y / ( 1 − p ) and monotone convergence give ∫ 0 ∞ e − ( 1 − p ) x ( 1 + x ) − δ d x ∼ Γ ( 1 − δ ) ( 1 − p ) δ − 1 \int_0^\infty e^{-(1-p)x}(1+x)^{-\delta}\,dx\sim\Gamma(1-\delta)(1-p)^{\delta-1} ∫ 0 ∞ e − ( 1 − p ) x ( 1 + x ) − δ d x ∼ Γ ( 1 − δ ) ( 1 − p ) δ − 1 , so there is a blow-up with C p ( A ) = 1 − δ 2 ln 1 1 − p + O ( 1 ) C_p(A)=\frac{1-\delta}{2}\ln\frac{1}{1-p}+O(1) C p ( A ) = 2 1 − δ ln 1 − p 1 + O ( 1 ) as p ↑ 1 p\uparrow1 p ↑ 1 .
(b) If δ = 1 \delta=1 δ = 1 , the integral is ln 1 1 − p + O ( 1 ) \ln\frac{1}{1-p}+O(1) ln 1 − p 1 + O ( 1 ) , and C p ( A ) = 1 2 ln ln 1 1 − p + O ( 1 ) C_p(A)=\frac{1}{2}\ln\ln\frac{1}{1-p}+O(1) C p ( A ) = 2 1 ln ln 1 − p 1 + O ( 1 ) : still a blow-up, but a slower one.
(c) If δ > 1 \delta>1 δ > 1 , there is a jump, with C 1 ( A ) = 1 2 ln δ δ − 1 + 1 2 ln ∫ e − x d ν < ∞ C_1(A)=\frac{1}{2}\ln\frac{\delta}{\delta-1}+\frac{1}{2}\ln\int e^{-x}\,d\nu<\infty C 1 ( A ) = 2 1 ln δ − 1 δ + 2 1 ln ∫ e − x d ν < ∞ . Since, integrating by parts, ∫ x k e x d ν = ∫ 0 ∞ ( k x k − 1 + x k ) ( 1 + x ) − δ d x \int x^k e^x\,d\nu=\int_0^\infty(kx^{k-1}+x^k)(1+x)^{-\delta}\,dx ∫ x k e x d ν = ∫ 0 ∞ ( k x k − 1 + x k ) ( 1 + x ) − δ d x for k ≥ 1 k\ge1 k ≥ 1 , Proposition 4.3.8 (c) shows that the left derivatives of order k k k at p c = 1 p_c=1 p c = 1 are finite exactly for k < δ − 1 k<\delta-1 k < δ − 1 .
In terms of generalised singular values, μ t ( A ) ∼ t − 1 ( ln 1 t ) − δ \mu_t(A)\sim t^{-1}(\ln\frac{1}{t})^{-\delta} μ t ( A ) ∼ t − 1 ( ln t 1 ) − δ as t → 0 t\to0 t → 0 : the power t − 1 t^{-1} t − 1 fixes the threshold, and the logarithmic correction decides between blow-up and jump and grades the jumps. In Example 4.3.4 with y = 1 y=1 y = 1 the factor Γ ( 1 − 2 p ) \Gamma(1-2p) Γ ( 1 − 2 p ) produces a blow-up at p c = 1 2 p_c=\frac{1}{2} p c = 2 1 with C p ( A ) = ln 1 1 2 − p + O ( 1 ) C_p(A)=\ln\frac{1}{\frac{1}{2}-p}+O(1) C p ( A ) = ln 2 1 − p 1 + O ( 1 ) .
Remark 4.3.10. Proposition 4.3.8 and Example 4.3.9 single out the following numerical characteristics of an operator with infinite contrast. They depend only on ν A \nu_A ν A , hence are unitary invariants of A A A , and are defined through traces of powers, without any truncation: the threshold p c p_c p c and the orders ω ± \omega_\pm ω ± ; the type of the transition at p c p_c p c ; in the case of a jump the value C p c ( A ) C_{p_c}(A) C p c ( A ) and the number of finite left derivatives at p c p_c p c ; in the case of a blow-up its order lim sup p ↑ p c 2 p C p ( A ) / ln 1 p c − p ∈ [ 0 , ∞ ] \limsup_{p\uparrow p_c}2pC_p(A)/\ln\frac{1}{p_c-p}\in[0,\infty] lim sup p ↑ p c 2 p C p ( A ) / ln p c − p 1 ∈ [ 0 , ∞ ] , which equals 1 − δ 1-\delta 1 − δ (0 ≤ δ ≤ 1 0\le\delta\le1 0 ≤ δ ≤ 1 ) in Example 4.3.9 and 1 1 1 in Example 4.3.4 with y = 1 y=1 y = 1 ; and, when p c = ∞ p_c=\infty p c = ∞ but OLC ( A ) = ∞ \operatorname{OLC}(A)=\infty OLC ( A ) = ∞ , the growth order lim sup p → ∞ ln C p ( A ) / ln p \limsup_{p\to\infty}\ln C_p(A)/\ln p lim sup p → ∞ ln C p ( A ) / ln p , which equals 1 q − 1 \frac{1}{q-1} q − 1 1 under the hypothesis of Theorem 4.3.1 (b).
Finite sections of unbounded operators: growth of the contrast in the dimension
In Section 4.3 the operator was unbounded but affiliated with a finite algebra. We now consider an unbounded operator on a Hilbert space, where the Schatten contrast with respect to the trace of B ( H ) B(\mathcal{H}) B ( H ) is not available (already for bounded invertible operators on an infinite-dimensional H \mathcal{H} H , since ∣ A − 1 ∣ |A^{-1}| ∣ A − 1 ∣ is unitarily equivalent to ∣ A ∣ − 1 |A|^{-1} ∣ A ∣ − 1 and Tr ( ∣ A ∣ p ) + Tr ( ∣ A − 1 ∣ p ) = Tr ( ∣ A ∣ p + ∣ A ∣ − p ) ≥ 2 Tr ( 1 ) = ∞ \operatorname{Tr}(|A|^p)+\operatorname{Tr}(|A^{-1}|^p)=\operatorname{Tr}(|A|^p+|A|^{-p})\ge2\operatorname{Tr}(1)=\infty Tr ( ∣ A ∣ p ) + Tr ( ∣ A − 1 ∣ p ) = Tr ( ∣ A ∣ p + ∣ A ∣ − p ) ≥ 2 Tr ( 1 ) = ∞ , so that one of the two positive traces is infinite and C p ( A ) = ∞ C_p(A)=\infty C p ( A ) = ∞ ), and study the contrasts of its finite sections as functions of their size n n n . Throughout, L L L is a positive self-adjoint operator on a separable Hilbert space with compact resolvent, with eigenvalues
0 < λ 1 ≤ λ 2 ≤ ⋯ , λ n → ∞ , 0<\lambda_1\le\lambda_2\le\cdots,\qquad\lambda_n\to\infty, 0 < λ 1 ≤ λ 2 ≤ ⋯ , λ n → ∞ ,
counted with multiplicity, and an orthonormal basis of eigenvectors ( e j ) (e_j) ( e j ) . The spectral section L n L_n L n is the restriction of L L L to span { e 1 , … , e n } \operatorname{span}\{e_1,\ldots,e_n\} span { e 1 , … , e n } , a positive invertible element of M n ( C ) M_n(\mathbb{C}) M n ( C ) with the normalised trace. Thus
OLC ( L n ) = 1 2 ln λ n λ 1 , κ ( L n ) = λ n λ 1 . (7) \operatorname{OLC}(L_n)=\frac{1}{2}\ln\frac{\lambda_n}{\lambda_1},\qquad\kappa(L_n)=\frac{\lambda_n}{\lambda_1}. \tag*{(7)} OLC ( L n ) = 2 1 ln λ 1 λ n , κ ( L n ) = λ 1 λ n . (7)
and the growth of the operator-norm contrast is the growth of the eigenvalues. The Schatten contrasts C p ( L n ) C_p(L_n) C p ( L n ) behave differently, and their behaviour separates spectra of power type, κ ( L n ) = O ( n α ) \kappa(L_n)=O(n^\alpha) κ ( L n ) = O ( n α ) , from lacunary spectra, λ j + 1 / λ j ≥ ϑ > 1 \lambda_{j+1}/\lambda_j\geq\vartheta>1 λ j + 1 / λ j ≥ ϑ > 1 , such as λ j = j ! \lambda_j=j! λ j = j ! with κ ( L n ) = n ! \kappa(L_n)=n! κ ( L n ) = n ! .
Spectral sections are optimal
We first show that (7) is, up to an additive constant, the slowest possible growth among all nested finite sections, so that the growth class of n ↦ OLC ( L n ) n\mapsto\operatorname{OLC}(L_n) n ↦ OLC ( L n ) is an invariant of L L L and not of the choice of sections. For a finite-dimensional subspace V V V of the form domain of L L L let L V L_V L V be the positive operator on V V V defined by ⟨ L V u , v ⟩ = ⟨ L 1 / 2 u , L 1 / 2 v ⟩ \langle L_Vu,v\rangle=\langle L^{1/2}u,L^{1/2}v\rangle ⟨ L V u , v ⟩ = ⟨ L 1/2 u , L 1/2 v ⟩ , u , v ∈ V u,v\in V u , v ∈ V .
Proposition 4.4.1. Let V 1 ⊂ V 2 ⊂ ⋯ V_1\subset V_2\subset\cdots V 1 ⊂ V 2 ⊂ ⋯ be subspaces of the form domain of L L L with dim V n = n \dim V_n=n dim V n = n , and let μ 1 = ∥ L 1 / 2 v ∥ 2 / ∥ v ∥ 2 \mu_1=\lVert L^{1/2}v\rVert^2/\lVert v\rVert^2 μ 1 = ∥ L 1/2 v ∥ 2 / ∥ v ∥ 2 for 0 ≠ v ∈ V 1 0\ne v\in V_1 0 = v ∈ V 1 . Then
OLC ( L V n ) ≥ 1 2 ln λ n μ 1 = OLC ( L n ) − 1 2 ln μ 1 λ 1 ( n ≥ 1 ) . \operatorname{OLC}(L_{V_n})\geq\frac{1}{2}\ln\frac{\lambda_n}{\mu_1}=\operatorname{OLC}(L_n)-\frac{1}{2}\ln\frac{\mu_1}{\lambda_1}\qquad(n\geq1). OLC ( L V n ) ≥ 2 1 ln μ 1 λ n = OLC ( L n ) − 2 1 ln λ 1 μ 1 ( n ≥ 1 ) . If ⋃ n V n \bigcup_n V_n ⋃ n V n is a form core for L L L , then lim inf n → ∞ ( OLC ( L V n ) − OLC ( L n ) ) ≥ 0 \liminf_{n\to\infty}(\operatorname{OLC}(L_{V_n})-\operatorname{OLC}(L_n))\geq0 lim inf n → ∞ ( OLC ( L V n ) − OLC ( L n )) ≥ 0 . There is no bound in the opposite direction.
Proof. The largest eigenvalue of L V n L_{V_n} L V n is the maximum of the Rayleigh quotient on V n V_n V n , which is at least λ n \lambda_n λ n by the min–max principle [57 ] . The smallest eigenvalue of L V n L_{V_n} L V n is the minimum of the Rayleigh quotient on V n V_n V n ; it is nonincreasing in n n n , at most μ 1 \mu_1 μ 1 , and tends to λ 1 \lambda_1 λ 1 if ⋃ n V n \bigcup_n V_n ⋃ n V n is a form core. For the last statement let V n = span { e σ ( 1 ) , … , e σ ( n ) } V_n=\operatorname{span}\{e_{\sigma(1)},\ldots,e_{\sigma(n)}\} V n = span { e σ ( 1 ) , … , e σ ( n ) } for a permutation σ \sigma σ of N \mathbb{N} N : then OLC ( L V n ) = 1 2 ln ( max j ≤ n λ σ ( j ) / min j ≤ n λ σ ( j ) ) \operatorname{OLC}(L_{V_n})=\frac{1}{2}\ln(\max_{j\leq n}\lambda_{\sigma(j)}/\min_{j\leq n}\lambda_{\sigma(j)}) OLC ( L V n ) = 2 1 ln ( max j ≤ n λ σ ( j ) / min j ≤ n λ σ ( j ) ) , which can grow arbitrarily fast. □ \square □
Growth of the contrast and Schatten classes
The growth rate of OLC ( L n ) \operatorname{OLC}(L_n) OLC ( L n ) on the logarithmic scale is the same thing as the abscissa of convergence of the spectral zeta function ζ L ( r ) = Tr L − r = ∑ j λ j − r \zeta_L(r)=\operatorname{Tr}L^{-r}=\sum_j\lambda_j^{-r} ζ L ( r ) = Tr L − r = ∑ j λ j − r , that is, as the membership of L − 1 L^{-1} L − 1 in the Schatten classes S r S_r S r .
Theorem 4.4.2. Let γ ∗ = lim inf n → ∞ 2 OLC ( L n ) ln n ∈ [ 0 , ∞ ] \gamma_*=\liminf_{n\to\infty}\frac{2\operatorname{OLC}(L_n)}{\ln n}\in[0,\infty] γ ∗ = lim inf n → ∞ l n n 2 OLC ( L n ) ∈ [ 0 , ∞ ] . Then
inf { r > 0 : L − 1 ∈ S r } = inf { r > 0 : ζ L ( r ) < ∞ } = 1 γ ∗ . \inf\{r>0:L^{-1}\in S_r\}=\inf\{r>0:\zeta_L(r)<\infty\}=\frac{1}{\gamma_*}. inf { r > 0 : L − 1 ∈ S r } = inf { r > 0 : ζ L ( r ) < ∞ } = γ ∗ 1 . (with 1 0 = ∞ \frac{1}{0}=\infty 0 1 = ∞ and 1 ∞ = 0 \frac{1}{\infty}=0 ∞ 1 = 0 ). In particular:
(a) if κ ( L n ) = O ( n α ) \kappa(L_n)=O(n^\alpha) κ ( L n ) = O ( n α ) , then L − 1 ∉ S r L^{-1}\notin S_r L − 1 ∈ / S r for r ≤ 1 / α r\leq1/\alpha r ≤ 1/ α ;
(b) if κ ( L n ) ≥ c n α \kappa(L_n)\geq cn^\alpha κ ( L n ) ≥ c n α for some c > 0 c>0 c > 0 and all large n n n , then L − 1 ∈ S r L^{-1}\in S_r L − 1 ∈ S r for all r > 1 / α r>1/\alpha r > 1/ α ;
(c) L − 1 ∈ ⋂ r > 0 S r L^{-1}\in\bigcap_{r>0}S_r L − 1 ∈ ⋂ r > 0 S r if and only if OLC ( L n ) / ln n → ∞ \operatorname{OLC}(L_n)/\ln n\to\infty OLC ( L n ) / ln n → ∞ ; this is the case for geometric and for factorial growth of κ ( L n ) \kappa(L_n) κ ( L n ) .
Proof. By (7) , γ ∗ = lim inf n ln λ n / ln n \gamma_*=\liminf_n\ln\lambda_n/\ln n γ ∗ = lim inf n ln λ n / ln n . If r > 1 / γ ∗ r>1/\gamma_* r > 1/ γ ∗ , choose γ ′ < γ ∗ \gamma'<\gamma_* γ ′ < γ ∗ with r γ ′ > 1 r\gamma'>1 r γ ′ > 1 ; then λ n ≥ n γ ′ \lambda_n\geq n^{\gamma'} λ n ≥ n γ ′ for large n n n and ∑ n λ n − r < ∞ \sum_n\lambda_n^{-r}<\infty ∑ n λ n − r < ∞ . Conversely, if ∑ j λ j − r = S < ∞ \sum_j\lambda_j^{-r}=S<\infty ∑ j λ j − r = S < ∞ , then, the sequence being nondecreasing, n λ n − r ≤ S n\lambda_n^{-r}\leq S n λ n − r ≤ S , so λ n ≥ ( n / S ) 1 / r \lambda_n\geq(n/S)^{1/r} λ n ≥ ( n / S ) 1/ r and γ ∗ ≥ 1 / r \gamma_*\geq1/r γ ∗ ≥ 1/ r . The singular values of L − 1 L^{-1} L − 1 are λ j − 1 \lambda_j^{-1} λ j − 1 , which gives the statement on Schatten classes, and (b), (c) are special cases. In (a), λ n ≤ C n α \lambda_n\leq Cn^\alpha λ n ≤ C n α gives ∑ n λ n − r ≥ C − r ∑ n n − α r = ∞ \sum_n\lambda_n^{-r}\geq C^{-r}\sum_n n^{-\alpha r}=\infty ∑ n λ n − r ≥ C − r ∑ n n − α r = ∞ for α r ≤ 1 \alpha r\leq1 α r ≤ 1 .
Theorem 4.4.2 is a statement on the logarithmic scale. The relation between the contrasts of the sections and the Schatten ideals is in fact exact, and it is convenient to state it for compact operators. Let X X X be a compact operator with singular values s 1 ≥ s 2 ≥ ⋯ > 0 s_1\geq s_2\geq\cdots>0 s 1 ≥ s 2 ≥ ⋯ > 0 , and let X n X_n X n be the section of ∣ X ∣ |X| ∣ X ∣ on the span of the first n n n singular vectors, so that κ ( X n ) = s 1 / s n \kappa(X_n)=s_1/s_n κ ( X n ) = s 1 / s n and OLC ( X n ) = 1 2 ln ( s 1 / s n ) \operatorname{OLC}(X_n)=\frac{1}{2}\ln(s_1/s_n) OLC ( X n ) = 2 1 ln ( s 1 / s n ) . Recall that the weak Schatten ideal S p , ∞ S_{p,\infty} S p , ∞ consists of the compact operators with s n = O ( n − 1 / p ) s_n=O(n^{-1/p}) s n = O ( n − 1/ p ) , and that S p ⊂ S p , ∞ ⊂ S q S_p\subset S_{p,\infty}\subset S_q S p ⊂ S p , ∞ ⊂ S q for q > p q>p q > p . In this section ∥ X ∥ p = ( Tr ∣ X ∣ p ) 1 / p \|X\|_p=(\operatorname{Tr}|X|^p)^{1/p} ∥ X ∥ p = ( Tr ∣ X ∣ p ) 1/ p denotes the unnormalised Schatten (quasi-)norm on B ( H ) B(\mathcal{H}) B ( H ) , whereas the contrasts C q C_q C q of the sections are always computed with the normalised trace of M n ( C ) M_n(\mathbb{C}) M n ( C ) .
Proposition 4.4.3. Let X X X be as above and 0 < p < ∞ 0<p<\infty 0 < p < ∞ .
(a)
( ∥ X ∥ p ∥ X ∥ ) p = ∑ n ≥ 1 κ ( X n ) − p = ∑ n ≥ 1 e − 2 p OLC ( X n ) . \left(\frac{\|X\|_p}{\|X\|}\right)^p=\sum_{n\geq1}\kappa(X_n)^{-p}=\sum_{n\geq1}e^{-2p\operatorname{OLC}(X_n)}. ( ∥ X ∥ ∥ X ∥ p ) p = n ≥ 1 ∑ κ ( X n ) − p = n ≥ 1 ∑ e − 2 p OLC ( X n ) . In particular X ∈ S p X\in S_p X ∈ S p if and only if this series converges.
(b) If X ∈ S p X\in S_p X ∈ S p , then for all n n n
OLC ( X n ) ≥ 1 2 p ln n − 1 2 ln ∥ X ∥ p ∥ X ∥ , \operatorname{OLC}(X_n)\geq\frac{1}{2p}\ln n-\frac{1}{2}\ln\frac{\|X\|_p}{\|X\|}, OLC ( X n ) ≥ 2 p 1 ln n − 2 1 ln ∥ X ∥ ∥ X ∥ p , and
OLC ( X n ) − 1 2 p ln n → + ∞ . \operatorname{OLC}(X_n)-\frac{1}{2p}\ln n\to+\infty. OLC ( X n ) − 2 p 1 ln n → + ∞. (c) X ∈ S p , ∞ X\in S_{p,\infty} X ∈ S p , ∞ if and only if inf n ( OLC ( X n ) − 1 2 p ln n ) > − ∞ \inf_n(\operatorname{OLC}(X_n)-\frac{1}{2p}\ln n)>-\infty inf n ( OLC ( X n ) − 2 p 1 ln n ) > − ∞ .
For an unbounded positive operator L L L with compact resolvent and X = L − 1 X=L^{-1} X = L − 1 one has κ ( X n ) = κ ( L n ) \kappa(X_n)=\kappa(L_n) κ ( X n ) = κ ( L n ) ; hence L − 1 ∈ S p L^{-1}\in S_p L − 1 ∈ S p if and only if ∑ n κ ( L n ) − p < ∞ \sum_n\kappa(L_n)^{-p}<\infty ∑ n κ ( L n ) − p < ∞ , with ζ L ( p ) = λ 1 − p ∑ n κ ( L n ) − p \zeta_L(p)=\lambda_1^{-p}\sum_n\kappa(L_n)^{-p} ζ L ( p ) = λ 1 − p ∑ n κ ( L n ) − p , and L − 1 ∈ S p , ∞ L^{-1}\in S_{p,\infty} L − 1 ∈ S p , ∞ if and only if OLC ( L n ) ≥ 1 2 p ln n − C \operatorname{OLC}(L_n)\geq\frac{1}{2p}\ln n-C OLC ( L n ) ≥ 2 p 1 ln n − C for some constant C C C .
Proof. (a) s n = s 1 / κ ( X n ) s_n=s_1/\kappa(X_n) s n = s 1 / κ ( X n ) and ∥ X ∥ = s 1 \|X\|=s_1 ∥ X ∥ = s 1 . (b) Since the s j s_j s j are nonincreasing, n s n p ≤ ∑ j ≤ n s j p ≤ ∥ X ∥ p p ns_n^p\leq\sum_{j\leq n}s_j^p\leq\|X\|_p^p n s n p ≤ ∑ j ≤ n s j p ≤ ∥ X ∥ p p , so κ ( X n ) = s 1 / s n ≥ n 1 / p ∥ X ∥ / ∥ X ∥ p \kappa(X_n)=s_1/s_n\geq n^{1/p}\|X\|/\|X\|_p κ ( X n ) = s 1 / s n ≥ n 1/ p ∥ X ∥/∥ X ∥ p ; moreover n 2 s n p ≤ ∑ n / 2 < j ≤ n s j p → 0 \frac{n}{2}s_n^p\leq\sum_{n/2<j\leq n}s_j^p\to0 2 n s n p ≤ ∑ n /2 < j ≤ n s j p → 0 , so n 1 / p s n → 0 n^{1/p}s_n\to0 n 1/ p s n → 0 . (c) s n ≤ C n − 1 / p s_n\leq Cn^{-1/p} s n ≤ C n − 1/ p is equivalent to 1 2 ln ( s 1 / s n ) ≥ 1 2 p ln n − 1 2 ln ( C / s 1 ) \frac{1}{2}\ln(s_1/s_n)\geq\frac{1}{2p}\ln n-\frac{1}{2}\ln(C/s_1) 2 1 ln ( s 1 / s n ) ≥ 2 p 1 ln n − 2 1 ln ( C / s 1 ) .
Thus a summable sequence of singular values forces the contrast of the sections to grow at least like 1 2 ln n \frac{1}{2}\ln n 2 1 ln n , a p p p -summable one at least like 1 2 p ln n \frac{1}{2p}\ln n 2 p 1 ln n , and the unbounded operators whose inverses lie in S p S_p S p are exactly those with ∑ n κ ( L n ) − p < ∞ \sum_n\kappa(L_n)^{-p}<\infty ∑ n κ ( L n ) − p < ∞ . Note the direction of the inequalities: membership in an ideal is a lower bound for the growth of the contrast, and it is the weak ideal S p , ∞ S_{p,\infty} S p , ∞ , not S p S_p S p , that corresponds exactly to the bound 1 2 p ln n − C \frac{1}{2p}\ln n-C 2 p 1 ln n − C . In Connes’ terminology [58 ] , X X X is an infinitesimal of order α \alpha α if s n = O ( n − α ) s_n=O(n^{-\alpha}) s n = O ( n − α ) , i.e. if OLC ( X n ) ≥ α 2 ln n − C \operatorname{OLC}(X_n)\geq\frac{\alpha}{2}\ln n-C OLC ( X n ) ≥ 2 α ln n − C : the slope of the contrast is half the order.
A dichotomy of the same kind appears at the abscissa. If σ = 1 / γ ∗ \sigma=1/\gamma_* σ = 1/ γ ∗ lies in ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) , then by Proposition 4.4.3 (a) and Theorem 4.4.2 either ∑ n κ ( L n ) − σ < ∞ \sum_n\kappa(L_n)^{-\sigma}<\infty ∑ n κ ( L n ) − σ < ∞ , i.e. L − 1 ∈ S σ L^{-1}\in S_\sigma L − 1 ∈ S σ , or the series diverges and L − 1 L^{-1} L − 1 lies in S r S_r S r only for r > σ r>\sigma r > σ . For λ n ∼ c n γ \lambda_n\sim c n^\gamma λ n ∼ c n γ one has κ ( L n ) − 1 / γ ∼ ( λ 1 / c ) 1 / γ n − 1 \kappa(L_n)^{-1/\gamma}\sim(\lambda_1/c)^{1/\gamma}n^{-1} κ ( L n ) − 1/ γ ∼ ( λ 1 / c ) 1/ γ n − 1 , so the series diverges like the harmonic series; for λ n = n γ ( 1 + ln n ) \lambda_n=n^\gamma(1+\ln n) λ n = n γ ( 1 + ln n ) the abscissa is still 1 / γ 1/\gamma 1/ γ , but ∑ n κ ( L n ) − 1 / γ = ∑ n 1 n ( 1 + ln n ) 2 < ∞ \sum_n\kappa(L_n)^{-1/\gamma}=\sum_n\frac{1}{n(1+\ln n)^2}<\infty ∑ n κ ( L n ) − 1/ γ = ∑ n n ( 1 + l n n ) 2 1 < ∞ . Unlike in Proposition 4.3.8 (b), the two cases are not separated by the contrasts of the sections themselves: C σ ( L n ) → ∞ C_\sigma(L_n)\to\infty C σ ( L n ) → ∞ in the convergent case by Proposition 4.4.4 (b), and for λ n ∼ c n γ \lambda_n\sim c n^\gamma λ n ∼ c n γ by Theorem 4.4.5 (c) (in the last example C 1 / γ ( L n ) = γ ln ln n + O ( 1 ) C_{1/\gamma}(L_n)=\gamma\ln\ln n+O(1) C 1/ γ ( L n ) = γ ln ln n + O ( 1 ) , compared with γ 2 ( ln ln n − ln 2 ) + o ( 1 ) \frac{\gamma}{2}(\ln\ln n-\ln2)+o(1) 2 γ ( ln ln n − ln 2 ) + o ( 1 ) in Theorem 4.4.5 (c) below).
The same comparison can be made between an ℓ p \ell_p ℓ p condition on the spectrum and a contrast with a different exponent q q q . The outcome is a threshold at q = p q=p q = p .
Proposition 4.4.4. Let X ∈ S p X\in S_p X ∈ S p , 0 < p < ∞ 0<p<\infty 0 < p < ∞ , have singular values s j > 0 s_j>0 s j > 0 , and let X n X_n X n be its sections.
(a) If q > p q>p q > p , then for all n n n
C q ( X n ) ≥ 1 2 ( 1 p − 1 q ) ln n − 1 2 ln ∥ X ∥ p ∥ X ∥ − 1 2 q ln ( 1 + q p ) . C_q(X_n)\geq\frac{1}{2}\left(\frac{1}{p}-\frac{1}{q}\right)\ln n-\frac{1}{2}\ln\frac{\|X\|_p}{\|X\|}-\frac{1}{2q}\ln\left(1+\frac{q}{p}\right). C q ( X n ) ≥ 2 1 ( p 1 − q 1 ) ln n − 2 1 ln ∥ X ∥ ∥ X ∥ p − 2 q 1 ln ( 1 + p q ) . (b) C p ( X n ) → ∞ C_p(X_n)\to\infty C p ( X n ) → ∞ as n → ∞ n\to\infty n → ∞ .
(c) For q < p q < p q < p nothing follows: C q ( X n ) C_q(X_n) C q ( X n ) may stay bounded or tend to infinity.
Consequently, for any compact X X X with s j > 0 s_j > 0 s j > 0 for all j j j : if sup n C q ( X n ) < ∞ \sup_n C_q(X_n) < \infty sup n C q ( X n ) < ∞ , then X ∉ S r X \notin S_r X ∈ / S r for every r ≤ q r \le q r ≤ q , and κ ( X n ) = O ( n 2 / q ) \kappa(X_n) = O(n^{2/q}) κ ( X n ) = O ( n 2/ q ) .
Proof. (a) We have ∑ j ≤ n s j q ≥ s 1 q = ∥ X ∥ q \sum_{j\le n}s_j^q \ge s_1^q = \lVert X\rVert^q ∑ j ≤ n s j q ≥ s 1 q = ∥ X ∥ q , and, as in the proof of Proposition 4.4.3, s j ≤ ∥ X ∥ p j − 1 / p s_j \le\lVert X\rVert_p j^{-1/p} s j ≤ ∥ X ∥ p j − 1/ p , so
∑ j ≤ n s j − q ≥ ∥ X ∥ p − q ∑ j ≤ n j q / p ≥ ∥ X ∥ p − q n q / p + 1 1 + q / p . \sum_{j\le n}s_j^{-q} \ge\lVert X\rVert_p^{-q}\sum_{j\le n}j^{q/p} \ge\lVert X\rVert_p^{-q}\frac{n^{q/p+1}}{1+q/p}. j ≤ n ∑ s j − q ≥ ∥ X ∥ p − q j ≤ n ∑ j q / p ≥ ∥ X ∥ p − q 1 + q / p n q / p + 1 .
Insert this into 2 q C q ( X n ) = ln ∑ j ≤ n s j q + ln ∑ j ≤ n s j − q − 2 ln n 2q\,C_q(X_n)=\ln\sum_{j\le n}s_j^q+\ln\sum_{j\le n}s_j^{-q}-2\ln n 2 q C q ( X n ) = ln ∑ j ≤ n s j q + ln ∑ j ≤ n s j − q − 2 ln n .
(b) Let ε n = sup k > n / 2 k s k p \varepsilon_n=\sup_{k>n/2}ks_k^p ε n = sup k > n /2 k s k p , which tends to 0 0 0 . Then ∑ n / 2 < j ≤ n s j − p ≥ n 2 ⋅ n 2 ε n \sum_{n/2<j\le n}s_j^{-p}\ge\frac{n}{2}\cdot\frac{n}{2\varepsilon_n} ∑ n /2 < j ≤ n s j − p ≥ 2 n ⋅ 2 ε n n , hence
e 2 p C p ( X n ) ≥ s 1 p 4 ε n → ∞ . e^{2pC_p(X_n)}\ge\frac{s_1^p}{4\varepsilon_n}\to\infty. e 2 p C p ( X n ) ≥ 4 ε n s 1 p → ∞.
(c) For s j = j − a s_j=j^{-a} s j = j − a with a > 1 / p a>1/p a > 1/ p , Theorem 4.4.5 below (applied to λ j = j a \lambda_j=j^a λ j = j a , since C q C_q C q is invariant under inversion) shows that C q ( X n ) C_q(X_n) C q ( X n ) is bounded for q < 1 / a q<1/a q < 1/ a and unbounded for q ≥ 1 / a q\ge1/a q ≥ 1/ a , and 1 / a < p 1/a<p 1/ a < p .
The last statement is the contrapositive of (a), (b), together with C q ( X n ) ≥ OLC ( X n ) − 1 q ln n C_q(X_n)\ge\operatorname{OLC}(X_n)-\frac{1}{q}\ln n C q ( X n ) ≥ OLC ( X n ) − q 1 ln n , which follows from 1 n ∑ j ≤ n s j q ≥ s 1 q n \frac{1}{n}\sum_{j\le n}s_j^q\ge\frac{s_1^q}{n} n 1 ∑ j ≤ n s j q ≥ n s 1 q and 1 n ∑ j ≤ n s j − q ≥ s n − q n \frac{1}{n}\sum_{j\le n}s_j^{-q}\ge\frac{s_n^{-q}}{n} n 1 ∑ j ≤ n s j − q ≥ n s n − q .
Thus an ℓ p \ell_p ℓ p spectrum forces every contrast with q ≥ p q\ge p q ≥ p to diverge, at a rate with slope at least 1 2 ( 1 p − 1 q ) \frac{1}{2}\left(\frac{1}{p}-\frac{1}{q}\right) 2 1 ( p 1 − q 1 ) , which increases to the slope 1 2 p \frac{1}{2p} 2 p 1 of OLC \operatorname{OLC} OLC as q → ∞ q\to\infty q → ∞ , and it imposes no condition on the contrasts with q < p q<p q < p . The slope in (a) is optimal: for s j = j − a s_j=j^{-a} s j = j − a the exact slope is 1 2 ( a − 1 q ) \frac{1}{2}\left(a-\frac{1}{q}\right) 2 1 ( a − q 1 ) by Theorem 4.4.5, and a a a can be taken arbitrarily close to 1 / p 1/p 1/ p . In terms of the two exponents
q c ( X ) = sup { q : sup n C q ( X n ) < ∞ } , p 0 ( X ) = inf { p : X ∈ S p } , q_c(X)=\sup\{q:\sup_n C_q(X_n)<\infty\},\qquad p_0(X)=\inf\{p:X\in S_p\}, q c ( X ) = sup { q : n sup C q ( X n ) < ∞ } , p 0 ( X ) = inf { p : X ∈ S p } ,
Proposition 4.4.4 says that q c ( X ) ≤ p 0 ( X ) q_c(X)\le p_0(X) q c ( X ) ≤ p 0 ( X ) . Equality holds for power-type spectra, where both are equal to 1 / a 1/a 1/ a . It can fail for irregular spectra: if s s s takes the value e − k ! e^{-k!} e − k ! on a block of length N k = ⌈ e b k ⋅ k ! ⌉ N_k=\lceil e^{b k\cdot k!}\rceil N k = ⌈ e bk ⋅ k ! ⌉ , k ≥ 1 k\ge1 k ≥ 1 , with some b > 0 b>0 b > 0 , then ∑ k N k e − p k ! = ∞ \sum_kN_ke^{-pk!}=\infty ∑ k N k e − p k ! = ∞ for every p p p , so p 0 ( X ) = ∞ p_0(X)=\infty p 0 ( X ) = ∞ , whereas for n = ∑ k ≤ K N k + 1 n=\sum_{k\le K}N_k+1 n = ∑ k ≤ K N k + 1 , the first index of the ( K + 1 ) (K+1) ( K + 1 ) -st block, one has n ≤ 2 K N K n\le2KN_K n ≤ 2 K N K and e 2 q C q ( X n ) ≥ N K e q ( K + 1 ) ! − q K ! / n 2 ≥ e ( q − b ) K ⋅ K ! / ( 8 K 2 ) → ∞ e^{2qC_q(X_n)}\ge N_Ke^{q(K+1)!-qK!}/n^2\ge e^{(q-b)K\cdot K!}/(8K^2)\to\infty e 2 q C q ( X n ) ≥ N K e q ( K + 1 )! − q K ! / n 2 ≥ e ( q − b ) K ⋅ K ! / ( 8 K 2 ) → ∞ for q > b q>b q > b , so q c ( X ) ≤ b q_c(X)\le b q c ( X ) ≤ b . Long plateaus followed by large drops make the contrast unbounded without making the spectrum summable.
Power-type spectra: a phase transition at the abscissa
For operators obeying a Weyl law the Schatten contrasts of the sections have a phase transition in p p p , located at the abscissa 1 / γ 1/\gamma 1/ γ of the zeta function. Below it they converge to the contrast of an unbounded operator in the sense of § 7.3 ; above it they grow logarithmically.
Theorem 4.4.5. Suppose that λ n = c n γ ( 1 + o ( 1 ) ) \lambda_n=cn^\gamma(1+o(1)) λ n = c n γ ( 1 + o ( 1 )) as n → ∞ n\to\infty n → ∞ , with c , γ > 0 c,\gamma>0 c , γ > 0 . Then, as n → ∞ n\to\infty n → ∞ :
(a) OLC ( L n ) = γ 2 ln n + 1 2 ln c λ 1 + o ( 1 ) \operatorname{OLC}(L_n)=\frac{\gamma}{2}\ln n+\frac{1}{2}\ln\frac{c}{\lambda_1}+o(1) OLC ( L n ) = 2 γ ln n + 2 1 ln λ 1 c + o ( 1 ) .
(b) For 0 < p < 1 / γ 0<p<1/\gamma 0 < p < 1/ γ ,
C p ( L n ) ⟶ − 1 2 p ln ( 1 − γ 2 p 2 ) = C p ( ν γ ) , C_p(L_n)\longrightarrow-\frac{1}{2p}\ln(1-\gamma^2p^2)=C_p(\nu_\gamma), C p ( L n ) ⟶ − 2 p 1 ln ( 1 − γ 2 p 2 ) = C p ( ν γ ) , where ν γ \nu_\gamma ν γ is the law of γ ln U \gamma\ln U γ ln U for U U U uniformly distributed on [ 0 , 1 ] [0,1] [ 0 , 1 ] . Moreover the logarithmic spectral measures of L n / λ n L_n/\lambda_n L n / λ n converge weakly to ν γ \nu_\gamma ν γ , which is the logarithmic spectral measure of the multiplication operator by x γ x^\gamma x γ on L 2 [ 0 , 1 ] L^2[0,1] L 2 [ 0 , 1 ] , and p c ( ν γ ) = 1 / γ p_c(\nu_\gamma)=1/\gamma p c ( ν γ ) = 1/ γ .
(c) For p = 1 / γ p=1/\gamma p = 1/ γ , C p ( L n ) = γ 2 ( ln ln n − ln 2 ) + o ( 1 ) C_p(L_n)=\frac{\gamma}{2}(\ln\ln n-\ln2)+o(1) C p ( L n ) = 2 γ ( ln ln n − ln 2 ) + o ( 1 ) .
(d) For p > 1 / γ p>1/\gamma p > 1/ γ ,
C p ( L n ) = ( γ 2 − 1 2 p ) ln n + 1 2 p ln c p ζ L ( p ) 1 + γ p + o ( 1 ) , C_p(L_n)=\left(\frac{\gamma}{2}-\frac{1}{2p}\right)\ln n+\frac{1}{2p}\ln\frac{c^p\zeta_L(p)}{1+\gamma p}+o(1), C p ( L n ) = ( 2 γ − 2 p 1 ) ln n + 2 p 1 ln 1 + γ p c p ζ L ( p ) + o ( 1 ) ,
and consequently
OLC ( L n ) − C p ( L n ) = 1 2 p ln n − 1 2 p ln λ 1 p ζ L ( p ) 1 + γ p + o ( 1 ) . \operatorname{OLC}(L_n)-C_p(L_n)=\frac{1}{2p}\ln n-\frac{1}{2p}\ln\frac{\lambda_1^p\zeta_L(p)}{1+\gamma p}+o(1). OLC ( L n ) − C p ( L n ) = 2 p 1 ln n − 2 p 1 ln 1 + γ p λ 1 p ζ L ( p ) + o ( 1 ) .
Proof. We use the following elementary fact: if a j > 0 a_j>0 a j > 0 and a j = j β ( 1 + o ( 1 ) ) a_j=j^\beta(1+o(1)) a j = j β ( 1 + o ( 1 )) , then ∑ j ≤ n a j \sum_{j\le n}a_j ∑ j ≤ n a j is asymptotic to n β + 1 / ( β + 1 ) n^{\beta+1}/(\beta+1) n β + 1 / ( β + 1 ) if β > − 1 \beta>-1 β > − 1 and to ln n \ln n ln n if β = − 1 \beta=-1 β = − 1 , and converges if β < − 1 \beta<-1 β < − 1 . {(citet}(a)} is (7) . For the other parts, 2 p C p ( L n ) = ln ( 1 n ∑ j ≤ n λ j p ) + ln ( 1 n ∑ j ≤ n λ j − p ) 2pC_p(L_n)=\ln\left(\frac{1}{n}\sum_{j\le n}\lambda_j^p\right)+\ln\left(\frac{1}{n}\sum_{j\le n}\lambda_j^{-p}\right) 2 p C p ( L n ) = ln ( n 1 ∑ j ≤ n λ j p ) + ln ( n 1 ∑ j ≤ n λ j − p ) . Since λ j p = c p j γ p ( 1 + o ( 1 ) ) \lambda_j^p=c^pj^{\gamma p}(1+o(1)) λ j p = c p j γ p ( 1 + o ( 1 )) ,
1 n ∑ j ≤ n λ j p = c p n γ p 1 + γ p ( 1 + o ( 1 ) ) , \frac{1}{n}\sum_{j\le n}\lambda_j^p=\frac{c^pn^{\gamma p}}{1+\gamma p}(1+o(1)), n 1 j ≤ n ∑ λ j p = 1 + γ p c p n γ p ( 1 + o ( 1 )) ,
while 1 n ∑ j ≤ n λ j − p \frac{1}{n}\sum_{j\le n}\lambda_j^{-p} n 1 ∑ j ≤ n λ j − p is asymptotic to c − p n − γ p / ( 1 − γ p ) c^{-p}n^{-\gamma p}/(1-\gamma p) c − p n − γ p / ( 1 − γ p ) if γ p < 1 \gamma p<1 γ p < 1 , to c − p ln n / n c^{-p}\ln n/n c − p ln n / n if γ p = 1 \gamma p=1 γ p = 1 , and to ζ L ( p ) / n \zeta_L(p)/n ζ L ( p ) / n if γ p > 1 \gamma p>1 γ p > 1 . Multiplying gives ( ( 1 + γ p ) ( 1 − γ p ) ) − 1 1 2 ln n ((1+\gamma p)(1-\gamma p))^{-1}\frac{1}{2}\ln n (( 1 + γ p ) ( 1 − γ p ) ) − 1 2 1 ln n and c p ζ L ( p ) n γ p − 1 / ( 1 + γ p ) c^p\zeta_L(p)n^{\gamma p-1}/(1+\gamma p) c p ζ L ( p ) n γ p − 1 / ( 1 + γ p ) respectively, which is {(citet}(b)}–{(citet}(d)}; the last formula in {(citet}(d)} follows by subtracting from {(citet}(a)}. In {(citet}(b)}, ∫ e ± p x d ν γ = ∫ 0 1 u ± γ p d u = ( 1 ± γ p ) − 1 \int e^{\pm px}\,d\nu_\gamma=\int_0^1u^{\pm\gamma p}\,du=(1\pm\gamma p)^{-1} ∫ e ± p x d ν γ = ∫ 0 1 u ± γ p d u = ( 1 ± γ p ) − 1 for γ p < 1 \gamma p<1 γ p < 1 , and the integral with the minus sign diverges for γ p ≥ 1 \gamma p\ge1 γ p ≥ 1 . Finally, for u ∈ ( 0 , 1 ] u\in(0,1] u ∈ ( 0 , 1 ] the proportion of indices j ≤ n j\le n j ≤ n with λ j / λ n ≤ u γ \lambda_j/\lambda_n\le u^\gamma λ j / λ n ≤ u γ tends to u u u , which is the weak convergence.
Thus, for a power-type spectrum, the sections L n L_n L n , rescaled by their norm, converge in distribution to a bounded operator with unbounded inverse in a finite von Neumann algebra, whose critical exponent is the abscissa of the spectral zeta function of L L L . For p p p below the abscissa the Schatten contrast of the sections stays bounded although κ ( L n ) → ∞ \kappa(L_n)\to\infty κ ( L n ) → ∞ ; for p p p above it, it grows like a multiple of ln n \ln n ln n with a slope γ 2 − 1 2 p \frac{\gamma}{2}-\frac{1}{2p} 2 γ − 2 p 1 that increases to the slope γ 2 \frac{\gamma}{2} 2 γ of OLC.
Lacunary spectra
Theorem 4.4.6. Suppose that λ j + 1 / λ j ≥ ϑ \lambda_{j+1}/\lambda_j\ge\vartheta λ j + 1 / λ j ≥ ϑ for all j j j , with some ϑ > 1 \vartheta>1 ϑ > 1 . Then for every p ∈ ( 0 , ∞ ) p\in(0,\infty) p ∈ ( 0 , ∞ ) and every n n n ,
OLC ( L n ) − ln n p ≤ C p ( L n ) ≤ OLC ( L n ) − ln n p + 1 p ln 1 1 − ϑ − p . \operatorname{OLC}(L_n)-\frac{\ln n}{p}\le C_p(L_n)\le\operatorname{OLC}(L_n)-\frac{\ln n}{p}+\frac{1}{p}\ln\frac{1}{1-\vartheta^{-p}}. OLC ( L n ) − p ln n ≤ C p ( L n ) ≤ OLC ( L n ) − p ln n + p 1 ln 1 − ϑ − p 1 . In particular C p ( L n ) → ∞ C_p(L_n)\to\infty C p ( L n ) → ∞ for every p > 0 p>0 p > 0 .
Proof. We have 2 p C p ( L n ) = p ln λ n λ 1 − 2 ln n + ln ∑ j ≤ n ( λ j λ n ) p + ln ∑ j ≤ n ( λ 1 λ j ) p 2pC_p(L_n)=p\ln\frac{\lambda_n}{\lambda_1}-2\ln n+\ln\sum_{j\le n}\left(\frac{\lambda_j}{\lambda_n}\right)^p+\ln\sum_{j\le n}\left(\frac{\lambda_1}{\lambda_j}\right)^p 2 p C p ( L n ) = p ln λ 1 λ n − 2 ln n + ln ∑ j ≤ n ( λ n λ j ) p + ln ∑ j ≤ n ( λ j λ 1 ) p , and each of the two sums lies between 1 1 1 and ∑ k ≥ 0 ϑ − k p = ( 1 − ϑ − p ) − 1 \sum_{k\ge0}\vartheta^{-kp}=(1-\vartheta^{-p})^{-1} ∑ k ≥ 0 ϑ − k p = ( 1 − ϑ − p ) − 1 . Finally, OLC ( L n ) ≥ n − 1 2 ln ϑ \operatorname{OLC}(L_n)\ge\frac{n-1}{2}\ln\vartheta OLC ( L n ) ≥ 2 n − 1 ln ϑ .
Theorems 4.4.5 and 4.4.6 separate the two growth classes by the coefficient of ln n \ln n ln n in the gap between the contrasts:
OLC ( L n ) − C p ( L n ) = { 1 2 p ln n + O ( 1 ) , λ n ∼ c n γ , p > 1 / γ , 1 p ln n + O ( 1 ) , λ j + 1 / λ j ≥ ϑ > 1. (8) \operatorname{OLC}(L_n)-C_p(L_n)=
\begin{cases}
\frac{1}{2p}\ln n+O(1), & \lambda_n\sim cn^\gamma,\ p>1/\gamma,\\
\frac{1}{p}\ln n+O(1), & \lambda_{j+1}/\lambda_j\ge\vartheta>1.
\end{cases} \tag*{(8)} OLC ( L n ) − C p ( L n ) = { 2 p 1 ln n + O ( 1 ) , p 1 ln n + O ( 1 ) , λ n ∼ c n γ , p > 1/ γ , λ j + 1 / λ j ≥ ϑ > 1. (8)
Heuristically: if the largest and the smallest eigenvalue of a positive invertible B ∈ M n ( C ) B\in M_n(\mathbb{C}) B ∈ M n ( C ) have multiplicities m 1 m_1 m 1 and m r m_r m r , then 1 n Tr B p = m 1 n ∥ B ∥ p ( 1 + o ( 1 ) ) \frac{1}{n}\operatorname{Tr}B^p=\frac{m_1}{n}\|B\|^p(1+o(1)) n 1 Tr B p = n m 1 ∥ B ∥ p ( 1 + o ( 1 )) and 1 n Tr B − p = m r n ∥ B − 1 ∥ p ( 1 + o ( 1 ) ) \frac{1}{n}\operatorname{Tr}B^{-p}=\frac{m_r}{n}\|B^{-1}\|^p(1+o(1)) n 1 Tr B − p = n m r ∥ B − 1 ∥ p ( 1 + o ( 1 )) as p → ∞ p\to\infty p → ∞ , so C p ( B ) = OLC ( B ) − 1 2 p ln n 2 m 1 m r + o ( 1 p ) C_p(B)=\operatorname{OLC}(B)-\frac{1}{2p}\ln\frac{n^2}{m_1m_r}+o\left(\frac{1}{p}\right) C p ( B ) = OLC ( B ) − 2 p 1 ln m 1 m r n 2 + o ( p 1 ) . For lacunary spectra both edges of the spectrum of L n L_n L n are isolated and the correction is the full 1 2 p ln n 2 m 1 m r = ln n p \frac{1}{2p}\ln\frac{n^2}{m_1m_r}=\frac{\ln n}{p} 2 p 1 ln m 1 m r n 2 = p l n n ; for power growth the upper edge is not isolated, the top of the spectrum behaves like a continuum, and only the lower edge contributes.
Example 4.4.7. (i) The Dirichlet Laplacian on [ 0 , π ] [0,\pi] [ 0 , π ] . Here λ j = j 2 \lambda_j=j^2 λ j = j 2 , so γ = 2 \gamma=2 γ = 2 , c = 1 c=1 c = 1 and ζ L ( p ) = ζ ( 2 p ) \zeta_L(p)=\zeta(2p) ζ L ( p ) = ζ ( 2 p ) . Then OLC ( L n ) = ln n \operatorname{OLC}(L_n)=\ln n OLC ( L n ) = ln n exactly; C p ( L n ) → − 1 2 p ln ( 1 − 4 p 2 ) C_p(L_n)\to-\frac{1}{2p}\ln(1-4p^2) C p ( L n ) → − 2 p 1 ln ( 1 − 4 p 2 ) for p < 1 2 p<\frac{1}{2} p < 2 1 ; and, for instance, C 1 ( L n ) = 1 2 ln n + 1 2 ln π 18 + o ( 1 ) C_1(L_n)=\frac{1}{2}\ln n+\frac{1}{2}\ln\frac{\pi}{18}+o(1) C 1 ( L n ) = 2 1 ln n + 2 1 ln 18 π + o ( 1 ) .
(ii) Elliptic operators. For the Dirichlet Laplacian on a bounded domain in R d \mathbb{R}^d R d , and more generally for a positive self-adjoint elliptic operator of order m m m on a compact d d d -dimensional manifold without boundary, Weyl’s law ([59 ] and [57 ] for domains, [60 ] for manifolds) gives λ n ∼ c n γ \lambda_n\sim cn^\gamma λ n ∼ c n γ with γ = m / d \gamma=m/d γ = m / d . Hence OLC ( L n ) = m 2 d ln n + O ( 1 ) \operatorname{OLC}(L_n)=\frac{m}{2d}\ln n+O(1) OLC ( L n ) = 2 d m ln n + O ( 1 ) , i.e. κ ( L n ) = O ( n m / d ) \kappa(L_n)=O(n^{m/d}) κ ( L n ) = O ( n m / d ) , and the Schatten contrasts of the sections remain bounded exactly for p < d / m p<d/m p < d / m ; for the Laplacian the critical exponent is half the dimension.
(iii) The harmonic oscillator. λ j = 2 j − 1 \lambda_j=2j-1 λ j = 2 j − 1 , γ = 1 \gamma=1 γ = 1 , c = 2 c=2 c = 2 , ζ L ( p ) = ( 1 − 2 − p ) ζ ( p ) \zeta_L(p)=(1-2^{-p})\zeta(p) ζ L ( p ) = ( 1 − 2 − p ) ζ ( p ) : OLC ( L n ) = 1 2 ln ( 2 n − 1 ) \operatorname{OLC}(L_n)=\frac{1}{2}\ln(2n-1) OLC ( L n ) = 2 1 ln ( 2 n − 1 ) , the critical exponent is 1 1 1 , and C 2 ( L n ) = 1 4 ln n + 1 4 ln π 2 6 + o ( 1 ) C_2(L_n)=\frac{1}{4}\ln n+\frac{1}{4}\ln\frac{\pi^2}{6}+o(1) C 2 ( L n ) = 4 1 ln n + 4 1 ln 6 π 2 + o ( 1 ) .
(iv) Geometric and factorial spectra. For λ j = ϑ j \lambda_j=\vartheta^j λ j = ϑ j , OLC ( L n ) = n − 1 2 ln ϑ \operatorname{OLC}(L_n)=\frac{n-1}{2}\ln\vartheta OLC ( L n ) = 2 n − 1 ln ϑ and κ ( L n ) = ϑ n − 1 \kappa(L_n)=\vartheta^{n-1} κ ( L n ) = ϑ n − 1 . For λ j = j ! \lambda_j=j! λ j = j ! , κ ( L n ) = n ! \kappa(L_n)=n! κ ( L n ) = n ! and, by Stirling’s formula,
OLC ( L n ) = 1 2 ln n ! = 1 2 n ln n − 1 2 n + 1 4 ln ( 2 π n ) + o ( 1 ) . \operatorname{OLC}(L_n)=\frac{1}{2}\ln n!=\frac{1}{2}n\ln n-\frac{1}{2}n+\frac{1}{4}\ln(2\pi n)+o(1). OLC ( L n ) = 2 1 ln n ! = 2 1 n ln n − 2 1 n + 4 1 ln ( 2 π n ) + o ( 1 ) .
In both cases C p ( L n ) = OLC ( L n ) − ln n p + O ( 1 ) C_p(L_n)=\operatorname{OLC}(L_n)-\frac{\ln n}{p}+O(1) C p ( L n ) = OLC ( L n ) − p l n n + O ( 1 ) for every p p p , there is no critical exponent, and L − 1 L^{-1} L − 1 belongs to every Schatten class.
Upper and lower orders, and their contrast
Theorem 4.4.2 uses only the lower limit of 2 OLC ( L n ) / ln n 2\operatorname{OLC}(L_n)/\ln n 2 OLC ( L n ) / ln n . Put
γ ∗ ( L ) = lim inf n → ∞ 2 OLC ( L n ) ln n = lim inf n → ∞ ln λ n ln n , \gamma_*(L)=\liminf_{n\to\infty}\frac{2\operatorname{OLC}(L_n)}{\ln n}=\liminf_{n\to\infty}\frac{\ln\lambda_n}{\ln n}, γ ∗ ( L ) = n → ∞ lim inf ln n 2 OLC ( L n ) = n → ∞ lim inf ln n ln λ n ,
γ ∗ ( L ) = lim sup n → ∞ 2 OLC ( L n ) ln n = lim sup n → ∞ ln λ n ln n , \gamma^*(L)=\limsup_{n\to\infty}\frac{2\operatorname{OLC}(L_n)}{\ln n}=\limsup_{n\to\infty}\frac{\ln\lambda_n}{\ln n}, γ ∗ ( L ) = n → ∞ lim sup ln n 2 OLC ( L n ) = n → ∞ lim sup ln n ln λ n ,
both in [ 0 , ∞ ] [0,\infty] [ 0 , ∞ ] , and, when 0 < γ ∗ ≤ γ ∗ < ∞ 0<\gamma_*\le\gamma^*<\infty 0 < γ ∗ ≤ γ ∗ < ∞ , define the order contrast
Δ o r d ( L ) = γ ∗ − γ ∗ γ ∗ + γ ∗ = OMC ( diag ( γ ∗ , γ ∗ ) ) , artanh Δ o r d ( L ) = 1 2 ln γ ∗ γ ∗ . \Delta_{\mathrm{ord}}(L)=\frac{\gamma^*-\gamma_*}{\gamma^*+\gamma_*}=\operatorname{OMC}(\operatorname{diag}(\gamma_*,\gamma^*)),
\qquad\operatorname{artanh}\Delta_{\mathrm{ord}}(L)=\frac{1}{2}\ln\frac{\gamma^*}{\gamma_*}. Δ ord ( L ) = γ ∗ + γ ∗ γ ∗ − γ ∗ = OMC ( diag ( γ ∗ , γ ∗ )) , artanh Δ ord ( L ) = 2 1 ln γ ∗ γ ∗ .
We set Δ o r d ( L ) = 1 \Delta_{\mathrm{ord}}(L)=1 Δ ord ( L ) = 1 if γ ∗ = 0 < γ ∗ \gamma_*=0<\gamma^* γ ∗ = 0 < γ ∗ or γ ∗ < γ ∗ = ∞ \gamma_*<\gamma^*=\infty γ ∗ < γ ∗ = ∞ , and leave it undefined if γ ∗ = γ ∗ ∈ { 0 , ∞ } \gamma_*=\gamma^*\in\{0,\infty\} γ ∗ = γ ∗ ∈ { 0 , ∞ } . Thus Δ o r d ( L ) = 0 \Delta_{\mathrm{ord}}(L)=0 Δ ord ( L ) = 0 exactly when λ n = n γ + o ( 1 ) \lambda_n=n^{\gamma+o(1)} λ n = n γ + o ( 1 ) for some γ ∈ ( 0 , ∞ ) \gamma\in(0,\infty) γ ∈ ( 0 , ∞ ) . Let N L ( λ ) = # { n : λ n ≤ λ } N_L(\lambda)=\#\{n:\lambda_n\le\lambda\} N L ( λ ) = # { n : λ n ≤ λ } be the eigenvalue counting function.
Proposition 4.4.8. Let L L L be as above.
(a) For c , s > 0 c,s>0 c , s > 0 : γ ∗ ( c L s ) = s γ ∗ ( L ) \gamma_*(cL^s)=s\gamma_*(L) γ ∗ ( c L s ) = s γ ∗ ( L ) and γ ∗ ( c L s ) = s γ ∗ ( L ) \gamma^*(cL^s)=s\gamma^*(L) γ ∗ ( c L s ) = s γ ∗ ( L ) ; hence Δ o r d ( c L s ) = Δ o r d ( L ) \Delta_{\mathrm{ord}}(cL^s)=\Delta_{\mathrm{ord}}(L) Δ ord ( c L s ) = Δ ord ( L ) .
(b) With 1 0 = ∞ \frac{1}{0}=\infty 0 1 = ∞ and 1 ∞ = 0 \frac{1}{\infty}=0 ∞ 1 = 0 ,
1 γ ∗ ( L ) = lim sup λ → ∞ ln N L ( λ ) ln λ , 1 γ ∗ ( L ) = lim inf λ → ∞ ln N L ( λ ) ln λ ; \frac{1}{\gamma_*(L)}=\limsup_{\lambda\to\infty}\frac{\ln N_L(\lambda)}{\ln\lambda},
\qquad
\frac{1}{\gamma^*(L)}=\liminf_{\lambda\to\infty}\frac{\ln N_L(\lambda)}{\ln\lambda}; γ ∗ ( L ) 1 = λ → ∞ lim sup ln λ ln N L ( λ ) , γ ∗ ( L ) 1 = λ → ∞ lim inf ln λ ln N L ( λ ) ;
in particular Δ o r d \Delta_{\mathrm{ord}} Δ ord is also the Michelson contrast of the upper and lower orders of N L N_L N L .
(c) The abscissa of convergence of ζ L \zeta_L ζ L is 1 / γ ∗ ( L ) 1/\gamma^*(L) 1/ γ ∗ ( L ) (Theorem 4.4.2 ) and does not determine γ ∗ ( L ) \gamma^*(L) γ ∗ ( L ) : for every pair 0 ≤ a ≤ b ≤ ∞ 0\le a\le b\le\infty 0 ≤ a ≤ b ≤ ∞ there is L L L with γ ∗ ( L ) = a \gamma_*(L)=a γ ∗ ( L ) = a and γ ∗ ( L ) = b \gamma^*(L)=b γ ∗ ( L ) = b .
(d) If sup n C q ( L n ) < ∞ \sup_n C_q(L_n) < \infty sup n C q ( L n ) < ∞ for some q > 0 q > 0 q > 0 , then q ≤ min { 1 / γ ∗ , 2 / γ ∗ } q \le\min\{1/\gamma_*, 2/\gamma^*\} q ≤ min { 1/ γ ∗ , 2/ γ ∗ } . In particular the exponent q c = sup { q : sup n C q ( L n ) < ∞ } q_c = \sup\{q : \sup_n C_q(L_n) < \infty\} q c = sup { q : sup n C q ( L n ) < ∞ } (sup ∅ = 0 \sup\varnothing= 0 sup ∅ = 0 , 1 / 0 = ∞ 1/0 = \infty 1/0 = ∞ ) satisfies q c ≤ min { 1 / γ ∗ , 2 / γ ∗ } q_c \le\min\{1/\gamma_*, 2/\gamma^*\} q c ≤ min { 1/ γ ∗ , 2/ γ ∗ } , and, when Δ o r d \Delta_{\mathrm{ord}} Δ ord is defined, the second bound is the better one exactly when γ ∗ > 2 γ ∗ \gamma^* > 2\gamma_* γ ∗ > 2 γ ∗ , i.e. when Δ o r d ( L ) > 1 3 \Delta_{\mathrm{ord}}(L) > \frac{1}{3} Δ ord ( L ) > 3 1 .
Proof. (a) The eigenvalues of c L s cL^s c L s are c λ n s c\lambda_n^s c λ n s , in the same order, and ln ( c λ n s ) / ln n = s ln λ n / ln n + ln c / ln n \ln(c\lambda_n^s)/\ln n = s\ln\lambda_n/\ln n+\ln c/\ln n ln ( c λ n s ) / ln n = s ln λ n / ln n + ln c / ln n .
(b) Since N L ( λ n ) ≥ n N_L(\lambda_n) \ge n N L ( λ n ) ≥ n , we have ln N L ( λ n ) / ln λ n ≥ ln n / ln λ n \ln N_L(\lambda_n)/\ln\lambda_n \ge\ln n/\ln\lambda_n ln N L ( λ n ) / ln λ n ≥ ln n / ln λ n , so the upper limit in the first formula is at least lim sup n ln n / ln λ n = 1 / γ ∗ \limsup_n\ln n/\ln\lambda_n = 1/\gamma_* lim sup n ln n / ln λ n = 1/ γ ∗ . Conversely, if N L ( λ ) = m N_L(\lambda) = m N L ( λ ) = m then λ m ≤ λ \lambda_m \le\lambda λ m ≤ λ , so ln N L ( λ ) / ln λ ≤ ln m / ln λ m \ln N_L(\lambda)/\ln\lambda\le\ln m/\ln\lambda_m ln N L ( λ ) / ln λ ≤ ln m / ln λ m once λ m > 1 \lambda_m > 1 λ m > 1 , and m → ∞ m \to\infty m → ∞ as λ → ∞ \lambda\to\infty λ → ∞ ; hence the upper limit is at most 1 / γ ∗ 1/\gamma_* 1/ γ ∗ . For the second formula: if λ m ≤ λ < λ m + 1 \lambda_m \le\lambda< \lambda_{m+1} λ m ≤ λ < λ m + 1 then N L ( λ ) ≥ m N_L(\lambda) \ge m N L ( λ ) ≥ m and ln N L ( λ ) / ln λ ≥ ln m ln ( m + 1 ) ⋅ ln ( m + 1 ) ln λ m + 1 \ln N_L(\lambda)/\ln\lambda\ge\frac{\ln m}{\ln(m+1)} \cdot\frac{\ln(m+1)}{\ln\lambda_{m+1}} ln N L ( λ ) / ln λ ≥ l n ( m + 1 ) l n m ⋅ l n λ m + 1 l n ( m + 1 ) , so the lower limit is at least 1 / γ ∗ 1/\gamma^* 1/ γ ∗ ; and for λ < λ n \lambda< \lambda_n λ < λ n we have N L ( λ ) ≤ n − 1 N_L(\lambda) \le n-1 N L ( λ ) ≤ n − 1 , so letting λ ↑ λ n \lambda\uparrow\lambda_n λ ↑ λ n along a sequence with ln λ ≥ ln λ n − 1 n \ln\lambda\ge\ln\lambda_n - \frac{1}{n} ln λ ≥ ln λ n − n 1 shows that it is at most lim inf n ln ( n − 1 ) / ln λ n = 1 / γ ∗ \liminf_n \ln(n-1)/\ln\lambda_n = 1/\gamma^* lim inf n ln ( n − 1 ) / ln λ n = 1/ γ ∗ .
(c) Let first 0 < a ≤ b < ∞ 0 < a \le b < \infty 0 < a ≤ b < ∞ . Choose T 1 ≥ 1 T_1 \ge1 T 1 ≥ 1 and define a nondecreasing function ψ \psi ψ on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) by induction on blocks: on [ T k , T k ′ ] [T_k,T'_k] [ T k , T k ′ ] , with T k ′ = 2 T k T'_k = 2T_k T k ′ = 2 T k , let ψ ( t ) = b t \psi(t) = bt ψ ( t ) = b t ; on [ T k ′ , a − 1 T k ′ ] [T'_k,a^{-1}T'_k] [ T k ′ , a − 1 T k ′ ] let ψ ( t ) = b T k ′ \psi(t) = bT'_k ψ ( t ) = b T k ′ ; on [ a − 1 T k ′ , T k + 1 ) [a^{-1}T'_k,T_{k+1}) [ a − 1 T k ′ , T k + 1 ) , with T k + 1 = k 2 T k ′ T_{k+1} = k^2T'_k T k + 1 = k 2 T k ′ , let ψ ( t ) = a t \psi(t) = at ψ ( t ) = a t ; and ψ = 0 \psi= 0 ψ = 0 on [ 0 , T 1 ) [0,T_1) [ 0 , T 1 ) . Then a ≤ ψ ( t ) / t ≤ b a \le\psi(t)/t \le b a ≤ ψ ( t ) / t ≤ b for t ≥ T 1 t \ge T_1 t ≥ T 1 , with ψ ( t ) / t = a \psi(t)/t = a ψ ( t ) / t = a on the third kind of interval and = b = b = b on the first, so lim inf ψ ( t ) / t = a \liminf\psi(t)/t = a lim inf ψ ( t ) / t = a and lim sup ψ ( t ) / t = b \limsup\psi(t)/t = b lim sup ψ ( t ) / t = b . Put λ n = e ψ ( ln n ) + 1 \lambda_n = e^{\psi(\ln n)} + 1 λ n = e ψ ( l n n ) + 1 and let L L L be diagonal with these eigenvalues; then ln λ n / ln n \ln\lambda_n/\ln n ln λ n / ln n has lower limit a a a and upper limit b b b . If a = 0 < b a = 0 < b a = 0 < b , let ψ ( t ) = b T k ′ \psi(t) = bT'_k ψ ( t ) = b T k ′ on [ T k ′ , T k + 1 ) [T'_k,T_{k+1}) [ T k ′ , T k + 1 ) with T k + 1 = k 2 T k ′ T_{k+1} = k^2T'_k T k + 1 = k 2 T k ′ . If b = ∞ b = \infty b = ∞ , use ψ ( t ) = t 2 \psi(t) = t^2 ψ ( t ) = t 2 on the first kind of interval (with T 1 ≥ max { a , 1 } T_1 \ge\max\{a,1\} T 1 ≥ max { a , 1 } ), continue with the constant ψ ( T k ′ ) = T k ′ 2 \psi(T'_k) = T_k'^2 ψ ( T k ′ ) = T k ′2 up to t = T k ′ 2 / a t = T_k'^2/a t = T k ′2 / a and then with ψ ( t ) = a t \psi(t) = at ψ ( t ) = a t up to T k + 1 = k T k ′ 2 / a T_{k+1} = kT_k'^2/a T k + 1 = k T k ′2 / a if a > 0 a > 0 a > 0 , and with the constant T k ′ 2 T_k'^2 T k ′2 up to T k + 1 = k 2 T k ′ 2 T_{k+1} = k^2T_k'^2 T k + 1 = k 2 T k ′2 if a = 0 a = 0 a = 0 . The cases a = b ∈ { 0 , ∞ } a = b \in\{0,\infty\} a = b ∈ { 0 , ∞ } are realised by λ n = ln ( n + 1 ) \lambda_n = \ln(n+1) λ n = ln ( n + 1 ) and λ n = e n \lambda_n = e^n λ n = e n .
(d) C q ( L n ) C_q(L_n) C q ( L n ) is invariant under inversion, and L n − 1 L_n^{-1} L n − 1 is the n n n -th section of the compact operator X = L − 1 X = L^{-1} X = L − 1 in the sense of Proposition 4.4.3 . By the last statement of Proposition 4.4.4 , sup n C q ( X n ) < ∞ \sup_n C_q(X_n) < \infty sup n C q ( X n ) < ∞ implies X ∉ S r X \notin S_r X ∈ / S r for r ≤ q r \le q r ≤ q , i.e. q ≤ 1 / γ ∗ q \le1/\gamma_* q ≤ 1/ γ ∗ by Theorem 4.4.2 , and κ ( L n ) = κ ( X n ) = O ( n 2 / q ) \kappa(L_n) = \kappa(X_n) = O(n^{2/q}) κ ( L n ) = κ ( X n ) = O ( n 2/ q ) , i.e. γ ∗ ≤ 2 / q \gamma^* \le2/q γ ∗ ≤ 2/ q .
So the zeta function and the Schatten classes of L − 1 L^{-1} L − 1 see only the lower order, and the order contrast is a second, scale-free number: it is a property of the family { c L s } \{cL^s\} { c L s } rather than of L L L (upper and lower limits of scales of spaces are studied in [61 ] ), and its logarithmic form is 1 2 ln ( γ ∗ / γ ∗ ) \frac{1}{2}\ln(\gamma^*/\gamma_*) 2 1 ln ( γ ∗ / γ ∗ ) is the logarithmic contrast of the pair of orders. When the orders coincide, finer information is in the constants (Subsection 4.4.7); when they are degenerate (γ ∗ = γ ∗ = 0 \gamma_* = \gamma^* = 0 γ ∗ = γ ∗ = 0 , as for λ n = ln ( n + 1 ) \lambda_n = \ln(n+1) λ n = ln ( n + 1 ) , or = ∞ = \infty = ∞ , as for λ n = e n \lambda_n = e^n λ n = e n ), the same construction can be repeated with ln ln λ n / ln ln n \ln\ln\lambda_n/\ln\ln n ln ln λ n / ln ln n or ln λ n / n \ln\lambda_n/n ln λ n / n in place of ln λ n / ln n \ln\lambda_n/\ln n ln λ n / ln n .
Local contrasts and relativistic addition
For n ≥ 1 n \ge1 n ≥ 1 let
δ n ( L ) = OMC ( diag ( λ n , λ n + 1 ) ) = λ n + 1 − λ n λ n + 1 + λ n ∈ [ 0 , 1 ) \delta_n(L) = \operatorname{OMC}(\operatorname{diag}(\lambda_n,\lambda_{n+1})) = \frac{\lambda_{n+1}-\lambda_n}{\lambda_{n+1}+\lambda_n} \in[0,1) δ n ( L ) = OMC ( diag ( λ n , λ n + 1 )) = λ n + 1 + λ n λ n + 1 − λ n ∈ [ 0 , 1 )
be the local contrasts of L L L , and let x ⊕ y = x + y 1 + x y x \oplus y = \frac{x+y}{1+xy} x ⊕ y = 1 + x y x + y on ( − 1 , 1 ) (-1,1) ( − 1 , 1 ) be the relativistic addition of velocities.
Proposition 4.4.9. (a) OLC ( L n ) = ∑ k < n artanh δ k \operatorname{OLC}(L_n) = \sum_{k<n}\operatorname{artanh}\delta_k OLC ( L n ) = ∑ k < n artanh δ k and OMC ( L n ) = δ 1 ⊕ δ 2 ⊕ ⋯ ⊕ δ n − 1 \operatorname{OMC}(L_n) = \delta_1 \oplus\delta_2 \oplus\cdots\oplus\delta_{n-1} OMC ( L n ) = δ 1 ⊕ δ 2 ⊕ ⋯ ⊕ δ n − 1 .
(b) 2 lim inf n → ∞ n artanh δ n ≤ γ ∗ ( L ) ≤ γ ∗ ( L ) ≤ 2 lim sup n → ∞ n artanh δ n 2\liminf_{n\to\infty} n\operatorname{artanh}\delta_n \le\gamma_*(L) \le\gamma^*(L) \le2\limsup_{n\to\infty} n\operatorname{artanh}\delta_n 2 lim inf n → ∞ n artanh δ n ≤ γ ∗ ( L ) ≤ γ ∗ ( L ) ≤ 2 lim sup n → ∞ n artanh δ n ; if δ n → 0 \delta_n \to0 δ n → 0 , the artanh \operatorname{artanh} artanh may be omitted.
(c) If n δ n → γ 2 ∈ [ 0 , ∞ ] n\delta_n \to\frac{\gamma}{2} \in[0,\infty] n δ n → 2 γ ∈ [ 0 , ∞ ] , then γ ∗ = γ ∗ = γ \gamma_* = \gamma^* = \gamma γ ∗ = γ ∗ = γ . The lacunarity hypothesis λ j + 1 / λ j ≥ ϑ \lambda_{j+1}/\lambda_j \ge\vartheta λ j + 1 / λ j ≥ ϑ of Theorem 4.4.6 is the condition δ j ≥ ϑ − 1 ϑ + 1 \delta_j \ge\frac{\vartheta-1}{\vartheta+1} δ j ≥ ϑ + 1 ϑ − 1 for all j j j , and it implies γ ∗ = ∞ \gamma_* = \infty γ ∗ = ∞ .
(d) The inequalities in (b) can be strict, also when γ ∗ = γ ∗ \gamma_*=\gamma^* γ ∗ = γ ∗ .
Proof. (a) artanh δ k = 1 2 ln ( λ k + 1 / λ k ) \operatorname{artanh}\delta_k=\frac{1}{2}\ln(\lambda_{k+1}/\lambda_k) artanh δ k = 2 1 ln ( λ k + 1 / λ k ) ; summing and using (7) gives the first formula, and the second follows from the addition theorem tanh ( u + v ) = tanh u ⊕ tanh v \tanh(u+v)=\tanh u\mathbin{\oplus}\tanh v tanh ( u + v ) = tanh u ⊕ tanh v .
(b) Put a n = ln λ n − ln λ 1 = 2 ∑ k < n artanh δ k a_n=\ln\lambda_n-\ln\lambda_1=2\sum_{k<n}\operatorname{artanh}\delta_k a n = ln λ n − ln λ 1 = 2 ∑ k < n artanh δ k and b n = ln n b_n=\ln n b n = ln n . Since b n b_n b n increases strictly to ∞ \infty ∞ , the Stolz–Cesàro theorem gives
lim inf a n + 1 − a n b n + 1 − b n ≤ lim inf a n b n ≤ lim sup a n b n ≤ lim sup a n + 1 − a n b n + 1 − b n . \liminf\frac{a_{n+1}-a_n}{b_{n+1}-b_n}\leq\liminf\frac{a_n}{b_n}\leq\limsup\frac{a_n}{b_n}\leq\limsup\frac{a_{n+1}-a_n}{b_{n+1}-b_n}. lim inf b n + 1 − b n a n + 1 − a n ≤ lim inf b n a n ≤ lim sup b n a n ≤ lim sup b n + 1 − b n a n + 1 − a n .
and a n + 1 − a n b n + 1 − b n = 2 artanh δ n ln ( 1 + 1 / n ) \frac{a_{n+1}-a_n}{b_{n+1}-b_n}=\frac{2\operatorname{artanh}\delta_n}{\ln(1+1/n)} b n + 1 − b n a n + 1 − a n = l n ( 1 + 1/ n ) 2 artanh δ n , where n ln ( 1 + 1 / n ) → 1 n\ln(1+1/n)\to1 n ln ( 1 + 1/ n ) → 1 . The middle terms are γ ∗ \gamma_* γ ∗ and γ ∗ \gamma^* γ ∗ , since ln λ 1 / ln n → 0 \ln\lambda_1/\ln n\to0 ln λ 1 / ln n → 0 . If δ n → 0 \delta_n\to0 δ n → 0 , then artanh δ n / δ n → 1 \operatorname{artanh}\delta_n/\delta_n\to1 artanh δ n / δ n → 1 .
(c) The first statement follows from (b) (if n δ n → ∞ n\delta_n\to\infty n δ n → ∞ , use artanh δ n ≥ δ n \operatorname{artanh}\delta_n\geq\delta_n artanh δ n ≥ δ n ). The second is
λ j + 1 − λ j λ j + 1 + λ j ≥ ϑ − 1 ϑ + 1 ⟺ λ j + 1 / λ j ≥ ϑ ; \frac{\lambda_{j+1}-\lambda_j}{\lambda_{j+1}+\lambda_j}\geq\frac{\vartheta-1}{\vartheta+1}\Longleftrightarrow\lambda_{j+1}/\lambda_j\geq\vartheta; λ j + 1 + λ j λ j + 1 − λ j ≥ ϑ + 1 ϑ − 1 ⟺ λ j + 1 / λ j ≥ ϑ ;
then n δ n → ∞ n\delta_n\to\infty n δ n → ∞ .
(d) Let r n = 1 r_n=1 r n = 1 if ⌊ log 4 n ⌋ \lfloor\log_4 n\rfloor ⌊ log 4 n ⌋ is even and r n = 3 r_n=3 r n = 3 otherwise, and λ 1 = 1 \lambda_1=1 λ 1 = 1 , λ n + 1 = λ n ( 1 + 1 / n ) r n \lambda_{n+1}=\lambda_n(1+1/n)^{r_n} λ n + 1 = λ n ( 1 + 1/ n ) r n . Then 2 n artanh δ n = r n n ln ( 1 + 1 / n ) 2n\operatorname{artanh}\delta_n=r_n n\ln(1+1/n) 2 n artanh δ n = r n n ln ( 1 + 1/ n ) has lower limit 1 1 1 and upper limit 3 3 3 . On the other hand ∑ 4 j ≤ k < 4 j + 1 ln ( 1 + 1 / k ) = ln 4 \sum_{4^j\leq k<4^{j+1}}\ln(1+1/k)=\ln4 ∑ 4 j ≤ k < 4 j + 1 ln ( 1 + 1/ k ) = ln 4 , so for 4 J ≤ n < 4 J + 1 4^J\leq n<4^{J+1} 4 J ≤ n < 4 J + 1 , writing n = 4 J x n=4^Jx n = 4 J x , ln λ n = ln 4 ∑ j < J ρ j + ρ J ln x \ln\lambda_n=\ln4\sum_{j<J}\rho_j+\rho_J\ln x ln λ n = ln 4 ∑ j < J ρ j + ρ J ln x , where ρ j ∈ { 1 , 3 } \rho_j\in\{1,3\} ρ j ∈ { 1 , 3 } is the value of r k r_k r k on the block 4 j ≤ k < 4 j + 1 4^j\leq k<4^{j+1} 4 j ≤ k < 4 j + 1 and ∑ j < J ρ j ∈ { 2 J − 1 , 2 J } \sum_{j<J}\rho_j\in\{2J-1,2J\} ∑ j < J ρ j ∈ { 2 J − 1 , 2 J } , and ln λ n / ln n → 2 \ln\lambda_n/\ln n\to2 ln λ n / ln n → 2 . Thus γ ∗ = γ ∗ = 2 \gamma_*=\gamma^*=2 γ ∗ = γ ∗ = 2 , while the outer bounds in (b) are 1 1 1 and 3 3 3 .
Part (a) is the equality case of the inequality OMC ( A B ) ≤ OMC ( A ) ⊕ OMC ( B ) \operatorname{OMC}(AB)\leq\operatorname{OMC}(A)\mathbin{\oplus}\operatorname{OMC}(B) OMC ( A B ) ≤ OMC ( A ) ⊕ OMC ( B ) of Theorem 1.3.1 for the commuting factors of L n L_n L n obtained by splitting its spectrum at consecutive eigenvalues: the contrast of a section is the relativistic sum of its local contrasts, and its logarithmic contrast, the sum of the corresponding rapidities. Part (b) says that the logarithmic rates γ ∗ , γ ∗ \gamma_*,\gamma^* γ ∗ , γ ∗ are bounded by the lower and upper mean rapidity per unit of ln n \ln n ln n ; geometric spectra λ j = ϑ j \lambda_j=\vartheta^j λ j = ϑ j have constant local contrast ϑ − 1 ϑ + 1 \frac{\vartheta-1}{\vartheta+1} ϑ + 1 ϑ − 1 , and power-type spectra have local contrasts of order γ 2 n \frac{\gamma}{2n} 2 n γ on logarithmic average, 1 ln n ∑ k < n 2 artanh δ k → γ \frac{1}{\ln n}\sum_{k<n}2\operatorname{artanh}\delta_k\to\gamma l n n 1 ∑ k < n 2 artanh δ k → γ , though not termwise (repeated eigenvalues give δ k = 0 \delta_k=0 δ k = 0 , and (d) shows oscillation).
Oscillating Weyl constants
Theorem 4.4.5 assumes λ n ∼ c n γ \lambda_n\sim cn^\gamma λ n ∼ c n γ . Suppose instead that, for some γ ∈ ( 0 , ∞ ) \gamma\in(0,\infty) γ ∈ ( 0 , ∞ ) (then necessarily γ ∗ = γ ∗ = γ \gamma_*=\gamma^*=\gamma γ ∗ = γ ∗ = γ ),
0 < D ∗ = lim inf n → ∞ λ n n γ ≤ D ∗ = lim sup n → ∞ λ n n γ < ∞ . 0<D_*=\liminf_{n\to\infty}\frac{\lambda_n}{n^\gamma}\leq D^*=\limsup_{n\to\infty}\frac{\lambda_n}{n^\gamma}<\infty. 0 < D ∗ = n → ∞ lim inf n γ λ n ≤ D ∗ = n → ∞ lim sup n γ λ n < ∞.
and let ϖ ( L ) = 1 2 ln ( D ∗ / D ∗ ) = OLC ( diag ( D ∗ , D ∗ ) ) \varpi(L)=\frac{1}{2}\ln(D^*/D_*)=\operatorname{OLC}(\operatorname{diag}(D_*,D^*)) ϖ ( L ) = 2 1 ln ( D ∗ / D ∗ ) = OLC ( diag ( D ∗ , D ∗ )) be the logarithmic contrast of the two Weyl constants; its Michelson form is ( D ∗ − D ∗ ) / ( D ∗ + D ∗ ) (D^*-D_*)/(D^*+D_*) ( D ∗ − D ∗ ) / ( D ∗ + D ∗ ) . Clearly ϖ ( c L ) = ϖ ( L ) \varpi(cL)=\varpi(L) ϖ ( c L ) = ϖ ( L ) , and ϖ ( L ) = 0 \varpi(L)=0 ϖ ( L ) = 0 exactly when λ n ∼ c n γ \lambda_n\sim cn^\gamma λ n ∼ c n γ .
Proposition 4.4.10. Under these assumptions, with ν γ \nu_\gamma ν γ as in Theorem 4.4.5 :
(a) lim sup n ( OLC ( L n ) − γ 2 ln n ) − lim inf n ( OLC ( L n ) − γ 2 ln n ) = ϖ ( L ) \limsup_n(\operatorname{OLC}(L_n)-\frac{\gamma}{2}\ln n)-\liminf_n(\operatorname{OLC}(L_n)-\frac{\gamma}{2}\ln n)=\varpi(L) lim sup n ( OLC ( L n ) − 2 γ ln n ) − lim inf n ( OLC ( L n ) − 2 γ ln n ) = ϖ ( L ) ;
(b) for 0 < p < 1 / γ 0<p<1/\gamma 0 < p < 1/ γ : lim sup n ∣ C p ( L n ) − C p ( ν γ ) ∣ ≤ ϖ ( L ) \limsup_n|C_p(L_n)-C_p(\nu_\gamma)|\leq\varpi(L) lim sup n ∣ C p ( L n ) − C p ( ν γ ) ∣ ≤ ϖ ( L ) ; for p = 1 / γ p=1/\gamma p = 1/ γ : lim sup n ∣ C p ( L n ) − γ 2 ( ln ln n − ln 2 ) ∣ ≤ ϖ ( L ) \limsup_n|C_p(L_n)-\frac{\gamma}{2}(\ln\ln n-\ln2)|\leq\varpi(L) lim sup n ∣ C p ( L n ) − 2 γ ( ln ln n − ln 2 ) ∣ ≤ ϖ ( L ) ;
(c) for p > 1 / γ p>1/\gamma p > 1/ γ , with R p ( n ) = C p ( L n ) − ( γ 2 − 1 2 p ) ln n − 1 2 p ln ζ L ( p ) 1 + γ p R_p(n)=C_p(L_n)-(\frac{\gamma}{2}-\frac{1}{2p})\ln n-\frac{1}{2p}\ln\frac{\zeta_L(p)}{1+\gamma p} R p ( n ) = C p ( L n ) − ( 2 γ − 2 p 1 ) ln n − 2 p 1 ln 1 + γ p ζ L ( p ) ,
1 2 ln D ∗ ≤ lim inf n R p ( n ) ≤ lim sup n R p ( n ) ≤ 1 2 ln D ∗ . \frac{1}{2}\ln D_*\leq\liminf_nR_p(n)\leq\limsup_nR_p(n)\leq\frac{1}{2}\ln D^*. 2 1 ln D ∗ ≤ n lim inf R p ( n ) ≤ n lim sup R p ( n ) ≤ 2 1 ln D ∗ .
so the oscillation of R p ( n ) R_p(n) R p ( n ) is at most ϖ ( L ) \varpi(L) ϖ ( L ) .
For ϖ ( L ) = 0 \varpi(L)=0 ϖ ( L ) = 0 , (b) and (c) reduce to Theorem 4.4.5 (b)–(d).
Proof. (a) By (7) , OLC ( L n ) − γ 2 ln n = 1 2 ln ( λ n / n γ ) − 1 2 ln λ 1 \operatorname{OLC}(L_n)-\frac{\gamma}{2}\ln n=\frac{1}{2}\ln(\lambda_n/n^\gamma)-\frac{1}{2}\ln\lambda_1 OLC ( L n ) − 2 γ ln n = 2 1 ln ( λ n / n γ ) − 2 1 ln λ 1 .
For (b) and (c) fix ε ∈ ( 0 , D ∗ ) \varepsilon\in(0,D_*) ε ∈ ( 0 , D ∗ ) and j 0 j_0 j 0 such that ( D ∗ − ε ) j γ ≤ λ j ≤ ( D ∗ + ε ) j γ (D_*-\varepsilon)j^\gamma\leq\lambda_j\leq(D^*+\varepsilon)j^\gamma ( D ∗ − ε ) j γ ≤ λ j ≤ ( D ∗ + ε ) j γ for j ≥ j 0 j\geq j_0 j ≥ j 0 . For s > − 1 s>-1 s > − 1 , ∑ j ≤ n j s ∼ n s + 1 / ( s + 1 ) \sum_{j\leq n}j^s\sim n^{s+1}/(s+1) ∑ j ≤ n j s ∼ n s + 1 / ( s + 1 ) , and the terms with j < j 0 j<j_0 j < j 0 are negligible when the sum diverges. Hence, for every p > 0 p>0 p > 0 ,
( D ∗ − ε ) p ≤ lim inf n 1 + γ p n γ p ⋅ 1 n ∑ j ≤ n λ j p ≤ lim sup n 1 + γ p n γ p ⋅ 1 n ∑ j ≤ n λ j p ≤ ( D ∗ + ε ) p . (D_*-\varepsilon)^p\leq\liminf_{n}\frac{1+\gamma p}{n^{\gamma p}}\cdot\frac{1}{n}\sum_{j\leq n}\lambda_j^p\leq\limsup_{n}\frac{1+\gamma p}{n^{\gamma p}}\cdot\frac{1}{n}\sum_{j\leq n}\lambda_j^p\leq(D^*+\varepsilon)^p. ( D ∗ − ε ) p ≤ n lim inf n γ p 1 + γ p ⋅ n 1 j ≤ n ∑ λ j p ≤ n lim sup n γ p 1 + γ p ⋅ n 1 j ≤ n ∑ λ j p ≤ ( D ∗ + ε ) p .
and, for γ p < 1 \gamma p<1 γ p < 1 , the same holds for 1 − γ p n − γ p ⋅ 1 n ∑ j ≤ n λ j − p \frac{1-\gamma p}{n^{-\gamma p}}\cdot\frac{1}{n}\sum_{j\leq n}\lambda_j^{-p} n − γ p 1 − γ p ⋅ n 1 ∑ j ≤ n λ j − p with the bounds ( D ∗ + ε ) − p (D^*+\varepsilon)^{-p} ( D ∗ + ε ) − p and ( D ∗ − ε ) − p (D_*-\varepsilon)^{-p} ( D ∗ − ε ) − p .
(b) Multiplying, and using 2 p C p ( L n ) = ln 1 n ∑ j ≤ n λ j p + ln 1 n ∑ j ≤ n λ j − p 2pC_p(L_n)=\ln\frac{1}{n}\sum_{j\leq n}\lambda_j^p+\ln\frac{1}{n}\sum_{j\leq n}\lambda_j^{-p} 2 p C p ( L n ) = ln n 1 ∑ j ≤ n λ j p + ln n 1 ∑ j ≤ n λ j − p and 2 p C p ( ν γ ) = − ln ( 1 − γ 2 p 2 ) 2pC_p(\nu_\gamma)=-\ln(1-\gamma^2p^2) 2 p C p ( ν γ ) = − ln ( 1 − γ 2 p 2 ) ,
lim sup n ∣ 2 p C p ( L n ) − 2 p C p ( ν γ ) ∣ ≤ p ln D ∗ + ε D ∗ − ε . \limsup_n\left|2pC_p(L_n)-2pC_p(\nu_\gamma)\right|\leq p\ln\frac{D^*+\varepsilon}{D_*-\varepsilon}. n lim sup ∣ 2 p C p ( L n ) − 2 p C p ( ν γ ) ∣ ≤ p ln D ∗ − ε D ∗ + ε .
Divide by 2 p 2p 2 p and let ε → 0 \varepsilon\to0 ε → 0 . For p = 1 / γ p=1/\gamma p = 1/ γ the first sum is as before, and 1 n ∑ j ≤ n λ j − p = 1 n ∑ j ≤ n d j − p j − 1 \frac{1}{n}\sum_{j\leq n}\lambda_j^{-p}=\frac{1}{n}\sum_{j\leq n}d_j^{-p}j^{-1} n 1 ∑ j ≤ n λ j − p = n 1 ∑ j ≤ n d j − p j − 1 with d j = λ j / j γ d_j=\lambda_j/j^\gamma d j = λ j / j γ lies between ( D ∗ + ε ) − p ln n n (D^*+\varepsilon)^{-p}\frac{\ln n}{n} ( D ∗ + ε ) − p n l n n and ( D ∗ − ε ) − p ln n n (D_*-\varepsilon)^{-p}\frac{\ln n}{n} ( D ∗ − ε ) − p n l n n asymptotically; with 2 p C p = ln ln n − ln 2 2pC_p=\ln\ln n-\ln2 2 p C p = ln ln n − ln 2 in the case ϖ = 0 \varpi=0 ϖ = 0 (Theorem 4.4.5 (c)) the same argument gives the bound.
(c) For γ p > 1 \gamma p>1 γ p > 1 we have ζ L ( p ) < ∞ \zeta_L(p)<\infty ζ L ( p ) < ∞ , since γ ∗ = γ \gamma_*=\gamma γ ∗ = γ (Theorem 4.4.2 ), and 1 n ∑ j ≤ n λ j − p = ζ L ( p ) n ( 1 + o ( 1 ) ) \frac{1}{n}\sum_{j\leq n}\lambda_j^{-p}=\frac{\zeta_L(p)}{n}(1+o(1)) n 1 ∑ j ≤ n λ j − p = n ζ L ( p ) ( 1 + o ( 1 )) . Put d ~ n = ( ( 1 + γ p ) n − γ p − 1 ∑ j ≤ n λ j p ) 1 / p \widetilde d_n=((1+\gamma p)n^{-\gamma p-1}\sum_{j\leq n}\lambda_j^p)^{1/p} d n = (( 1 + γ p ) n − γ p − 1 ∑ j ≤ n λ j p ) 1/ p , a weighted mean of the d j = λ j / j γ d_j=\lambda_j/j^\gamma d j = λ j / j γ (it is not d n d_n d n itself, and lags behind it). Hence 2 p C p ( L n ) = ( γ p − 1 ) ln n + ln ζ L ( p ) 1 + γ p + p ln d ~ n + o ( 1 ) 2pC_p(L_n)=(\gamma p-1)\ln n+\ln\frac{\zeta_L(p)}{1+\gamma p}+p\ln\widetilde d_n+o(1) 2 p C p ( L n ) = ( γ p − 1 ) ln n + ln 1 + γ p ζ L ( p ) + p ln d n + o ( 1 ) , and by the bounds at the beginning of the proof lim inf d ~ n ≥ D ∗ − ε \liminf\widetilde d_n\geq D_*-\varepsilon lim inf d n ≥ D ∗ − ε and lim sup d ~ n ≤ D ∗ + ε \limsup\widetilde d_n\leq D^*+\varepsilon lim sup d n ≤ D ∗ + ε ; divide by 2 p 2p 2 p . For ϖ = 0 \varpi=0 ϖ = 0 these are the statements of Theorem 4.4.5 , with c = D ∗ = D ∗ c=D_*=D^* c = D ∗ = D ∗ . □
Example 4.4.11 (The Sierpinski gasket). For the standard Laplacian on the Sierpinski gasket (with Dirichlet conditions, or the Neumann Laplacian plus the identity; neither D ∗ D_* D ∗ nor D ∗ D^* D ∗ changes, since by Dirichlet–Neumann bracketing the two counting functions differ by a bounded amount [62 ] , and the shift by 1 1 1 does not change λ n / n γ \lambda_n/n^\gamma λ n / n γ asymptotically), the counting function satisfies c 1 λ d s / 2 ≤ N ( λ ) ≤ c 2 λ d s / 2 c_1\lambda^{d_s/2}\leq N(\lambda)\leq c_2\lambda^{d_s/2} c 1 λ d s /2 ≤ N ( λ ) ≤ c 2 λ d s /2 for large λ \lambda λ [62 ] , with spectral dimension d s = 2 ln 3 / ln 5 d_s=2\ln3/\ln5 d s = 2 ln 3/ ln 5 , but N ( λ ) λ − d s / 2 N(\lambda)\lambda^{-d_s/2} N ( λ ) λ − d s /2 has no limit as λ → ∞ \lambda\to\infty λ → ∞ [63 ] . With γ = 2 / d s \gamma=2/d_s γ = 2/ d s , the two-sided bound gives ( n / c 2 ) γ ≤ λ n ≤ ( ( n − 1 ) / c 1 ) γ (n/c_2)^\gamma\leq\lambda_n\leq((n-1)/c_1)^\gamma ( n / c 2 ) γ ≤ λ n ≤ (( n − 1 ) / c 1 ) γ for large n n n , so 0 < D ∗ ≤ D ∗ < ∞ 0<D_*\leq D^*<\infty 0 < D ∗ ≤ D ∗ < ∞ ; and D ∗ = D ∗ = c D_*=D^*=c D ∗ = D ∗ = c would give N ( λ ) ∼ ( λ / c ) 1 / γ N(\lambda)\sim(\lambda/c)^{1/\gamma} N ( λ ) ∼ ( λ / c ) 1/ γ . Hence γ ∗ = γ ∗ = γ = ln 5 / ln 3 \gamma_*=\gamma^*=\gamma=\ln5/\ln3 γ ∗ = γ ∗ = γ = ln 5/ ln 3 , Δ o r d = 0 \Delta_{\mathrm{ord}}=0 Δ ord = 0 , but ϖ > 0 \varpi>0 ϖ > 0 : Theorem 4.4.5 does not apply, and Proposition 4.4.10 does. The value of ϖ \varpi ϖ is not computed here. An elementary example with ϖ \varpi ϖ computed is the sequence of Proposition 4.4.9 (d): there ln λ n − 2 ln n ∈ [ − ln 4 , 0 ] \ln\lambda_n-2\ln n\in[-\ln4,0] ln λ n − 2 ln n ∈ [ − ln 4 , 0 ] , with both ends approached, so γ = 2 \gamma=2 γ = 2 , D ∗ = 1 4 D_*=\frac{1}{4} D ∗ = 4 1 , D ∗ = 1 D^*=1 D ∗ = 1 and ϖ = ln 2 \varpi=\ln2 ϖ = ln 2 .
Thus the spectral information of a section sequence is graded: the lower rate γ ∗ \gamma_* γ ∗ (seen by the zeta function), the order contrast Δ o r d \Delta_{\mathrm{ord}} Δ ord , and, when Δ o r d = 0 \Delta_{\mathrm{ord}}=0 Δ ord = 0 , the contrast ϖ \varpi ϖ of the Weyl constants; the last two are Michelson (or logarithmic) contrasts of an upper and a lower limit.
Essential contrast
For a bounded operator A A A on an infinite-dimensional separable Hilbert space H \mathcal H H , OLC ( A ) \operatorname{OLC}(A) OLC ( A ) sees the whole spectrum, while the Schatten contrasts with respect to the trace of B ( H ) B(\mathcal H) B ( H ) are infinite for invertible A A A (see the beginning of § 7.7 ) and trivially otherwise. The upper and lower limits of the spectrum are seen by the Calkin algebra Q ( H ) = B ( H ) / K ( H ) \mathcal Q(\mathcal H)=B(\mathcal H)/\mathcal K(\mathcal H) Q ( H ) = B ( H ) / K ( H ) . Let π : B ( H ) → Q ( H ) \pi:B(\mathcal H)\to\mathcal Q(\mathcal H) π : B ( H ) → Q ( H ) be the quotient map, and recall that the essential spectrum σ e s s ( A ) \sigma_{\mathrm{ess}}(A) σ ess ( A ) is the spectrum of π ( A ) \pi(A) π ( A ) . For A A A with π ( A ) \pi(A) π ( A ) invertible put
OLC e s s ( A ) = OLC ( π ( A ) ) = 1 2 ln ( ∥ π ( A ) ∥ ∥ π ( A ) − 1 ∥ ) , (9) \operatorname{OLC}_{\mathrm{ess}}(A)=\operatorname{OLC}(\pi(A))=\frac{1}{2}\ln\left(\|\pi(A)\|\|\pi(A)^{-1}\|\right), \tag*{(9)} OLC ess ( A ) = OLC ( π ( A )) = 2 1 ln ( ∥ π ( A ) ∥∥ π ( A ) − 1 ∥ ) , (9)
OMC e s s ( A ) = tanh OLC e s s ( A ) , \operatorname{OMC}_{\mathrm{ess}}(A)=\tanh\operatorname{OLC}_{\mathrm{ess}}(A), OMC ess ( A ) = tanh OLC ess ( A ) ,
the contrasts of [1 , 2 ] in the unital C ∗ C^{*} C ∗ -algebra Q ( H ) \mathcal{Q}(\mathcal{H}) Q ( H ) .
Proposition 4.5.1. (a) If A ≥ 0 A \ge0 A ≥ 0 and π ( A ) \pi(A) π ( A ) is invertible, then, with a ∗ = max σ e s s ( A ) a^{*}=\max\sigma_{\mathrm{ess}}(A) a ∗ = max σ ess ( A ) and a ∗ = min σ e s s ( A ) > 0 a_{*}=\min\sigma_{\mathrm{ess}}(A)>0 a ∗ = min σ ess ( A ) > 0 ,
OMC e s s ( A ) = a ∗ − a ∗ a ∗ + a ∗ . \operatorname{OMC}_{\mathrm{ess}}(A)=\frac{a^{*}-a_{*}}{a^{*}+a_{*}}. OMC ess ( A ) = a ∗ + a ∗ a ∗ − a ∗ . For a diagonal operator A = diag ( α n ) A=\operatorname{diag}(\alpha_n) A = diag ( α n ) with α n > 0 \alpha_n>0 α n > 0 , sup α n < ∞ \sup\alpha_n<\infty sup α n < ∞ and lim inf α n > 0 \liminf\alpha_n>0 lim inf α n > 0 , one has a ∗ = lim sup α n a^{*}=\limsup\alpha_n a ∗ = lim sup α n and a ∗ = lim inf α n a_{*}=\liminf\alpha_n a ∗ = lim inf α n .
(b) For invertible A A A , OLC e s s ( A ) ≤ OLC ( A ) \operatorname{OLC}_{\mathrm{ess}}(A)\le\operatorname{OLC}(A) OLC ess ( A ) ≤ OLC ( A ) . For A , B A,B A , B with π ( A ) , π ( B ) \pi(A),\pi(B) π ( A ) , π ( B ) invertible, compact K K K , λ ≠ 0 \lambda\ne0 λ = 0 and unitaries U , V U,V U , V :
OLC e s s ( A + K ) = OLC e s s ( A ) , OLC e s s ( λ U A V ) = OLC e s s ( A ) , \operatorname{OLC}_{\mathrm{ess}}(A+K)=\operatorname{OLC}_{\mathrm{ess}}(A),\qquad
\operatorname{OLC}_{\mathrm{ess}}(\lambda UAV)=\operatorname{OLC}_{\mathrm{ess}}(A), OLC ess ( A + K ) = OLC ess ( A ) , OLC ess ( λ U A V ) = OLC ess ( A ) , and
OLC e s s ( A B ) ≤ OLC e s s ( A ) + OLC e s s ( B ) . \operatorname{OLC}_{\mathrm{ess}}(AB)\le\operatorname{OLC}_{\mathrm{ess}}(A)+\operatorname{OLC}_{\mathrm{ess}}(B). OLC ess ( A B ) ≤ OLC ess ( A ) + OLC ess ( B ) . (c) If A ≥ 0 A\ge0 A ≥ 0 and π ( A ) \pi(A) π ( A ) is invertible, then
OLC e s s ( A ) = inf { OLC ( A + K ) : K = K ∗ compact , A + K positive and invertible } , \operatorname{OLC}_{\mathrm{ess}}(A)=\inf\{\operatorname{OLC}(A+K):K=K^{*}\text{ compact},\ A+K\text{ positive and invertible}\}, OLC ess ( A ) = inf { OLC ( A + K ) : K = K ∗ compact , A + K positive and invertible } , and the infimum may be taken over self-adjoint operators K K K of finite rank.
Proof. (a) π ( A ) \pi(A) π ( A ) is a positive invertible element of a C ∗ C^{*} C ∗ -algebra with spectrum σ e s s ( A ) \sigma_{\mathrm{ess}}(A) σ ess ( A ) , so ∥ π ( A ) ∥ = a ∗ \lVert\pi(A)\rVert=a^{*} ∥ π ( A )∥ = a ∗ and ∥ π ( A ) − 1 ∥ = 1 / a ∗ \lVert\pi(A)^{-1}\rVert=1/a_{*} ∥ π ( A ) − 1 ∥ = 1/ a ∗ . For a diagonal operator the essential spectrum is the set of cluster values of the sequence ( α n ) (\alpha_n) ( α n ) , i.e. of the t t t with # { n : ∣ α n − t ∣ < ε } = ∞ \#\{n:|\alpha_n-t|<\varepsilon\}=\infty # { n : ∣ α n − t ∣ < ε } = ∞ for all ε > 0 \varepsilon>0 ε > 0 , whose largest and smallest elements are lim sup α n \limsup\alpha_n lim sup α n and lim inf α n \liminf\alpha_n lim inf α n .
(b) ∥ π ( A ) ∥ ≤ ∥ A ∥ \lVert\pi(A)\rVert\le\lVert A\rVert ∥ π ( A )∥ ≤ ∥ A ∥ and ∥ π ( A ) − 1 ∥ = ∥ π ( A − 1 ) ∥ ≤ ∥ A − 1 ∥ \lVert\pi(A)^{-1}\rVert=\lVert\pi(A^{-1})\rVert\le\lVert A^{-1}\rVert ∥ π ( A ) − 1 ∥ = ∥ π ( A − 1 )∥ ≤ ∥ A − 1 ∥ . The second and third statements hold because π ( A + K ) = π ( A ) \pi(A+K)=\pi(A) π ( A + K ) = π ( A ) and π ( U ) , π ( V ) \pi(U),\pi(V) π ( U ) , π ( V ) are unitary, and the last because the condition number is submultiplicative in any unital C ∗ C^{*} C ∗ -algebra.
(c) By (b), OLC e s s ( A ) = OLC e s s ( A + K ) ≤ OLC ( A + K ) \operatorname{OLC}_{\mathrm{ess}}(A)=\operatorname{OLC}_{\mathrm{ess}}(A+K)\le\operatorname{OLC}(A+K) OLC ess ( A ) = OLC ess ( A + K ) ≤ OLC ( A + K ) for every admissible K K K . Conversely let 0 < ε < a ∗ 0<\varepsilon<a_{*} 0 < ε < a ∗ and P = 1 R ∖ ( a ∗ − ε , a ∗ + ε ) ( A ) P=\mathbf{1}_{\mathbb{R}\setminus(a_{*}-\varepsilon,a^{*}+\varepsilon)}(A) P = 1 R ∖ ( a ∗ − ε , a ∗ + ε ) ( A ) . The part of σ ( A ) \sigma(A) σ ( A ) outside ( a ∗ − ε , a ∗ + ε ) (a_{*}-\varepsilon,a^{*}+\varepsilon) ( a ∗ − ε , a ∗ + ε ) is a compact set of isolated eigenvalues of finite multiplicity, hence finite, so P P P has finite rank. With c = 1 2 ( a ∗ + a ∗ ) c=\frac{1}{2}(a_{*}+a^{*}) c = 2 1 ( a ∗ + a ∗ ) and K = ( c − A ) P K=(c-A)P K = ( c − A ) P , the operator A + K = A ( 1 − P ) + c P A+K=A(1-P)+cP A + K = A ( 1 − P ) + c P is positive, its spectrum lies in [ a ∗ − ε , a ∗ + ε ] [a_{*}-\varepsilon,a^{*}+\varepsilon] [ a ∗ − ε , a ∗ + ε ] , and
OLC ( A + K ) ≤ 1 2 ln a ∗ + ε a ∗ − ε ⟶ OLC e s s ( A ) . \operatorname{OLC}(A+K)\le\frac{1}{2}\ln\frac{a^{*}+\varepsilon}{a_{*}-\varepsilon}\longrightarrow\operatorname{OLC}_{\mathrm{ess}}(A). OLC ( A + K ) ≤ 2 1 ln a ∗ − ε a ∗ + ε ⟶ OLC ess ( A ) .
In a factor of type I I ∞ \mathrm{II}_{\infty} II ∞ , statements (b) and (c) hold for the norm-closed ideal J \mathcal{J} J generated by the projections of finite trace [64 ] in place of K ( H ) \mathcal{K}(\mathcal{H}) K ( H ) , with “finite rank” replaced by “with support projection of finite trace”, and the proof of (c) modified as follows. Take a continuous g g g with 0 ≤ g ≤ 1 0\le g\le1 0 ≤ g ≤ 1 , g = 0 g=0 g = 0 on [ a ∗ , a ∗ ] [a_{*},a^{*}] [ a ∗ , a ∗ ] and g = 1 g=1 g = 1 outside ( a ∗ − ε , a ∗ + ε ) (a_{*}-\varepsilon,a^{*}+\varepsilon) ( a ∗ − ε , a ∗ + ε ) ; then π ( g ( A ) ) = g ( π ( A ) ) = 0 \pi(g(A))=g(\pi(A))=0 π ( g ( A )) = g ( π ( A )) = 0 , so g ( A ) ∈ J g(A)\in\mathcal{J} g ( A ) ∈ J , and the spectral projection P = 1 S ( A ) P=\mathbf{1}_{S}(A) P = 1 S ( A ) of the set S S S where g = 1 g=1 g = 1 satisfies P = P g ( A ) P=Pg(A) P = P g ( A ) ; hence P ∈ J P\in\mathcal{J} P ∈ J is a projection in a norm-closed ideal of a semifinite algebra, so it has finite trace. For unbounded operators with compact resolvent the essential spectrum is empty and the essential contrast is not defined; its role is taken by the order contrast and by the contrast of the Weyl constants of § 7.7 , which are again contrasts of an upper and a lower limit.
Concluding remarks
The infinite contrast of an unbounded operator is graded by several numbers: the threshold p c p_c p c and the type of the transition there, for operators affiliated with a finite algebra; for finite sections, the lower rate γ ∗ \gamma_{*} γ ∗ and two contrasts of upper and lower limits, the scale-free order contrast and the contrast ϖ \varpi ϖ of the Weyl constants; and the essential contrast, for bounded operators read modulo compact ones. We mention three open questions.
Weyl remainders. Theorem 4.4.5 and Proposition 4.4.10 use only the leading term of the counting function. How do the remainder term in Weyl’s law, and more generally regularly varying counting functions, enter the next terms of C p ( L n ) C_p(L_n) C p ( L n ) ?
(2) Sharpness. Is the bound ϖ \varpi ϖ in Proposition 4.4.10(b) attained? For log-periodic Weyl constants, as on p.c.f. self-similar fractals, does C p ( L n ) C_p(L_n) C p ( L n ) itself oscillate log-periodically below the abscissa, and with which amplitude?
(3) Non-spectral sections. What is the growth of the contrasts for the sections used in numerical analysis (finite elements), for which Proposition 4.4.1 gives only a lower bound?
Contrasts of random products
Introduction
The operator logarithmic contrast OLC ( A ) = 1 2 ln κ ( A ) \operatorname{OLC}(A)=\frac{1}{2}\ln\kappa(A) OLC ( A ) = 2 1 ln κ ( A ) , κ ( A ) = ∥ A ∥ ∥ A − 1 ∥ \kappa(A)=\lVert A\rVert\lVert A^{-1}\rVert κ ( A ) = ∥ A ∥ ∥ A − 1 ∥ , and the operator Michelson contrast OMC ( A ) = tanh OLC ( A ) \operatorname{OMC}(A)=\tanh\operatorname{OLC}(A) OMC ( A ) = tanh OLC ( A ) were introduced and studied in [1 , 2 ] : OLC is subadditive under products, homogeneous under powers of positive elements, and vanishes exactly on multiples of unitaries. The Schatten contrasts C p C_p C p replace the operator norm by normalised Schatten norms; the elementary facts about them that we use are collected in § 8.2 or proved where they are used. In this chapter we ask how these quantities behave along products of independent random factors, and along the opposite process, convergence to the tracial state.
Growth: three kinds of independence. Let A N = X N ⋯ X 1 A_N=X_N\cdots X_1 A N = X N ⋯ X 1 with identically distributed invertible factors.
Tensor independence. If the factors sit in different tensor factors, every contrast is exactly additive: C p ( A N ) = N C p ( X ) C_p(A_N)=NC_p(X) C p ( A N ) = N C p ( X ) for all p ∈ ( 0 , ∞ ] p\in(0,\infty] p ∈ ( 0 , ∞ ] (Proposition 5.3.1). The slopes depend on p p p .
Matrix independence. For i.i.d. factors in GL d ( C ) \operatorname{GL}_d(\mathbb{C}) GL d ( C ) , d d d fixed, all contrasts have the same slope 1 2 ( γ 1 − γ d ) \frac{1}{2}(\gamma_1-\gamma_d) 2 1 ( γ 1 − γ d ) , where γ 1 ≥ γ d \gamma_1\geq\gamma_d γ 1 ≥ γ d are the extreme Lyapunov exponents, and OLC ( A N ) \operatorname{OLC}(A_N) OLC ( A N ) fluctuates on the scale N \sqrt{N} N (Theorem 5.4.1 ). This is the theorem of Furstenberg and Kesten and the central limit theorem of Goldsheid–Guivarc’h and Benoist–Quint, read for contrasts; the existence of the limit follows directly from the subadditivity of OLC by Kingman’s theorem, and its value from Furstenberg–Kesten.
Free independence. For ∗ \ast ∗ -free factors in a finite von Neumann algebra, C 2 C_2 C 2 is exactly additive, all slopes collapse to C 2 ( X ) C_2(X) C 2 ( X ) , and
C p ( A N ) = N C 2 ( X ) + ( 1 2 − 1 p ) ln N + O ( 1 ) ( 2 ≤ p < ∞ ) , C_p(A_N)=NC_2(X)+\left(\frac{1}{2}-\frac{1}{p}\right)\ln N+O(1)\qquad(2\leq p<\infty), C p ( A N ) = N C 2 ( X ) + ( 2 1 − p 1 ) ln N + O ( 1 ) ( 2 ≤ p < ∞ ) ,
with an explicit constant in place of O ( 1 ) O(1) O ( 1 ) for even integers p p p , and
OLC ( A N ) = N C 2 ( X ) + 1 2 ln N + 1 2 + 1 4 ln ( v v ′ ) + o ( 1 ) , \operatorname{OLC}(A_N)=NC_2(X)+\frac{1}{2}\ln N+\frac{1}{2}+\frac{1}{4}\ln(vv')+o(1), OLC ( A N ) = N C 2 ( X ) + 2 1 ln N + 2 1 + 4 1 ln ( v v ′ ) + o ( 1 ) ,
where v , v ′ v,v' v , v ′ are normalised variances of X ∗ X X^\ast X X ∗ X and ( X ∗ X ) − 1 (X^\ast X)^{-1} ( X ∗ X ) − 1 (Theorem 5.5.2 ).
The second-order term of OLC ( A N ) \operatorname{OLC}(A_N) OLC ( A N ) is therefore 0 0 0 , of order N \sqrt{N} N and random, or 1 2 ln N \frac{1}{2}\ln N 2 1 ln N and deterministic, in the three cases. For p < 2 p<2 p < 2 we prove lower bounds with the conjectured logarithmic term, and for all p ∉ 2 N p\notin2\mathbb{N} p ∈ / 2 N we conjecture the full expansion with an explicit constant, supported by high-precision numerical evaluation of the exact free formulas (Conjecture 5.5.4 ). For Haar-rotated factors, the matrix slope is bounded by the free slope C 2 ( X ) C_2(X) C 2 ( X ) and converges to it as d → ∞ d \to\infty d → ∞ (Proposition 5.6.1 ).
Decay: convergence to the tracial state. A positive element has contrast 0 exactly when it is a multiple of 1, i.e. when the corresponding state is the trace. Since 2 OLC(ρ \rho ρ ) is the Hilbert projective distance from ρ \rho ρ to 1, the Birkhoff–Hopf theorem gives geometric decay of OLC under unital positive maps of finite projective diameter, with the contraction ratio tanh ( Δ / 4 ) \tanh(\Delta/4) tanh ( Δ/4 ) (Proposition 5.7.1 ). For convolution powers on the circle, the Michelson contrast of the density decays like 2 ∣ a k 0 ∣ ∣ μ ^ ( k 0 ) ∣ N 2|a_{k_0}| |\hat{\mu}(k_0)|^N 2∣ a k 0 ∣∣ μ ^ ( k 0 ) ∣ N (Proposition 5.7.2 ); this is the classical fact that phase noise multiplies the visibility of interference fringes by the modulus of the characteristic function of the noise, and it is in this sense that the Michelson contrast was introduced in optics. For random unital channels the decay rate is a Lyapunov exponent (Proposition 5.7.3 ).
What is new. Proposition 5.6.1 (comparison of slopes), Proposition 5.5.1, Theorem 5.5.2 (b)–(d) and Conjecture 5.5.4 appear to be new; Theorem 5.5.2 (a) is a direct corollary of results of Kargin and of Deng–Zhong on free multiplicative convolution powers [65 , 66 ] , and the integer moment asymptotics used in the proof are a short Lagrange inversion computation, related to limit theorems in [67 , 68 ] . Theorem 5.4.1 and the results of § 8.7 are reformulations of classical theorems, included because, in the language of contrasts, they complete the comparison and give the growth and decay of contrasts a common form. Table 5.1 summarises the status of the results.
Table 5.1. Status of the results.
Preliminaries
Throughout, M \mathcal{M} M is a von Neumann algebra with a faithful normal tracial state τ \tau τ , and for X ∈ M X \in\mathcal{M} X ∈ M and 0 < p < ∞ 0 < p < \infty 0 < p < ∞ we write ∥ X ∥ p = τ ( ∣ X ∣ p ) 1 / p \|X\|_p = \tau(|X|^p)^{1/p} ∥ X ∥ p = τ ( ∣ X ∣ p ) 1/ p (a quasi-norm for p < 1 p < 1 p < 1 ) and ∥ X ∥ ∞ = ∥ X ∥ \|X\|_\infty= \|X\| ∥ X ∥ ∞ = ∥ X ∥ . For invertible A ∈ M A \in\mathcal{M} A ∈ M put
C p ( A ) = 1 2 ( ln ∥ A ∥ p + ln ∥ A − 1 ∥ p ) = 1 2 p ln ( τ ( ∣ A ∣ p ) τ ( ∣ A − 1 ∣ p ) ) , C ∞ ( A ) = OLC ( A ) . C_p(A) = \tfrac{1}{2}\bigl(\ln\|A\|_p + \ln\|A^{-1}\|_p\bigr) = \frac{1}{2p}\ln\bigl(\tau(|A|^p)\tau(|A^{-1}|^p)\bigr), \qquad C_\infty(A) = \operatorname{OLC}(A). C p ( A ) = 2 1 ( ln ∥ A ∥ p + ln ∥ A − 1 ∥ p ) = 2 p 1 ln ( τ ( ∣ A ∣ p ) τ ( ∣ A − 1 ∣ p ) ) , C ∞ ( A ) = OLC ( A ) .
Let ν A \nu_A ν A be the distribution of ln ∣ A ∣ \ln|A| ln ∣ A ∣ with respect to τ \tau τ . We use the following elementary facts, with C p ( ν ) = 1 2 p ln ( ∫ e p x d ν ∫ e − p x d ν ) C_p(\nu)=\frac{1}{2p}\ln\left(\int e^{px}\,d\nu\int e^{-px}\,d\nu\right) C p ( ν ) = 2 p 1 ln ( ∫ e p x d ν ∫ e − p x d ν ) :
(i) C p ( A ) = C p ( ν A ) C_p(A)=C_p(\nu_A) C p ( A ) = C p ( ν A ) : in the polar decomposition A = U ∣ A ∣ A=U|A| A = U ∣ A ∣ the factor U U U is unitary and ∣ A − 1 ∣ = U ∣ A ∣ − 1 U ∗ |A^{-1}|=U|A|^{-1}U^* ∣ A − 1 ∣ = U ∣ A ∣ − 1 U ∗ , so τ ( ∣ A − 1 ∣ p ) = τ ( ∣ A ∣ − p ) = ∫ e − p x d ν A \tau(|A^{-1}|^p)=\tau(|A|^{-p})=\int e^{-px}\,d\nu_A τ ( ∣ A − 1 ∣ p ) = τ ( ∣ A ∣ − p ) = ∫ e − p x d ν A ; in particular C p ( A − 1 ) = C p ( A ) C_p(A^{-1})=C_p(A) C p ( A − 1 ) = C p ( A ) ;
(ii) C p ( A ) ≥ 0 C_p(A)\geq0 C p ( A ) ≥ 0 , with equality iff ∣ A ∣ |A| ∣ A ∣ is a multiple of 1 1 1 , by the Cauchy–Schwarz inequality
1 = ( ∫ e p x / 2 e − p x / 2 d ν A ) 2 ≤ ∫ e p x d ν A ∫ e − p x d ν A ; 1=\left(\int e^{px/2}e^{-px/2}\,d\nu_A\right)^2\leq\int e^{px}\,d\nu_A\int e^{-px}\,d\nu_A; 1 = ( ∫ e p x /2 e − p x /2 d ν A ) 2 ≤ ∫ e p x d ν A ∫ e − p x d ν A ;
(iii) p ↦ C p ( A ) p\mapsto C_p(A) p ↦ C p ( A ) is nondecreasing with limit OLC ( A ) \operatorname{OLC}(A) OLC ( A ) , since on a probability space the L p L^p L p -norms of e x e^x e x and e − x e^{-x} e − x increase to the essential suprema;
(iv) if ν A ′ = ν A ∗ ν B \nu_{A'}=\nu_A\ast\nu_B ν A ′ = ν A ∗ ν B then C p ( A ′ ) = C p ( A ) + C p ( B ) C_p(A')=C_p(A)+C_p(B) C p ( A ′ ) = C p ( A ) + C p ( B ) , since Laplace transforms multiply under convolution.
Lemma 5.2.1. Let X 1 , … , X N ∈ M X_1,\ldots,X_N\in\mathcal{M} X 1 , … , X N ∈ M be invertible, p 1 , … , p N ∈ ( 0 , ∞ ] p_1,\ldots,p_N\in(0,\infty] p 1 , … , p N ∈ ( 0 , ∞ ] and 1 r = ∑ i 1 p i \frac{1}{r}=\sum_i\frac{1}{p_i} r 1 = ∑ i p i 1 . Then
C r ( X N ⋯ X 1 ) ≤ ∑ i C p i ( X i ) . C_r(X_N\cdots X_1)\leq\sum_i C_{p_i}(X_i). C r ( X N ⋯ X 1 ) ≤ i ∑ C p i ( X i ) .
In particular C p / N ( X N ⋯ X 1 ) ≤ ∑ i C p ( X i ) C_{p/N}(X_N\cdots X_1)\leq\sum_i C_p(X_i) C p / N ( X N ⋯ X 1 ) ≤ ∑ i C p ( X i ) and OLC ( X N ⋯ X 1 ) ≤ ∑ i OLC ( X i ) \operatorname{OLC}(X_N\cdots X_1)\leq\sum_i\operatorname{OLC}(X_i) OLC ( X N ⋯ X 1 ) ≤ ∑ i OLC ( X i ) .
Proof. By the Hölder inequality ∥ X Y ∥ r ≤ ∥ X ∥ p ∥ Y ∥ q \|XY\|_r\leq\|X\|_p\|Y\|_q ∥ X Y ∥ r ≤ ∥ X ∥ p ∥ Y ∥ q , 1 r = 1 p + 1 q \frac{1}{r}=\frac{1}{p}+\frac{1}{q} r 1 = p 1 + q 1 , valid for all 0 < p , q ≤ ∞ 0<p,q\leq\infty 0 < p , q ≤ ∞ [31 , 32 ] , applied repeatedly to the product and to its inverse X 1 − 1 ⋯ X N − 1 X_1^{-1}\cdots X_N^{-1} X 1 − 1 ⋯ X N − 1 .
Free multiplicative convolution. For compactly supported probability measures μ , μ ′ \mu,\mu' μ , μ ′ on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) , μ ⊠ μ ′ \mu\boxtimes\mu' μ ⊠ μ ′ is the distribution of a 1 / 2 b a 1 / 2 a^{1/2}ba^{1/2} a 1/2 b a 1/2 for free positive a , b a,b a , b with distributions μ , μ ′ \mu,\mu' μ , μ ′ [82 , 49 ] . With the moment series ψ μ ( z ) = ∑ k ≥ 1 m k ( μ ) z k \psi_\mu(z)=\sum_{k\geq1}m_k(\mu)z^k ψ μ ( z ) = ∑ k ≥ 1 m k ( μ ) z k , its compositional inverse χ μ \chi_\mu χ μ and the S S S -transform S μ ( w ) = 1 + w χ μ ( w ) S_\mu(w)=\frac{1+w}{\chi_\mu(w)} S μ ( w ) = χ μ ( w ) 1 + w , one has S μ ⊠ μ ′ = S μ S μ ′ S_{\mu\boxtimes\mu'}=S_\mu S_{\mu'} S μ ⊠ μ ′ = S μ S μ ′ . If m 1 ( μ ) = 1 m_1(\mu)=1 m 1 ( μ ) = 1 and μ \mu μ has variance V V V , then S μ ( w ) = 1 − V w + O ( w 2 ) S_\mu(w)=1-Vw+O(w^2) S μ ( w ) = 1 − V w + O ( w 2 ) .
Tensor independence and a general bound
Proposition 5.3.1. For i = 1 , … , N i=1,\ldots,N i = 1 , … , N let X i X_i X i be an invertible element of a von Neumann algebra M i \mathcal{M}_i M i with a faithful normal tracial state τ i \tau_i τ i , and let A N = X 1 ⊗ ⋯ ⊗ X N A_N=X_1\otimes\cdots\otimes X_N A N = X 1 ⊗ ⋯ ⊗ X N in the tensor product ⨂ i ( M i , τ i ) \bigotimes_i(\mathcal{M}_i,\tau_i) ⨂ i ( M i , τ i ) . Then C p ( A N ) = ∑ i C p ( X i ) C_p(A_N)=\sum_i C_p(X_i) C p ( A N ) = ∑ i C p ( X i ) for every p ∈ ( 0 , ∞ ] p\in(0,\infty] p ∈ ( 0 , ∞ ] . In particular, for identically distributed X i X_i X i , C p ( A N ) = N C p ( X ) C_p(A_N)=N C_p(X) C p ( A N ) = N C p ( X ) , and the slope C p ( X ) C_p(X) C p ( X ) depends on p p p unless ∣ X ∣ |X| ∣ X ∣ is a multiple of 1 1 1 .
Proof. ∣ A N ∣ = ⨂ ∣ X i ∣ |A_N|=\bigotimes|X_i| ∣ A N ∣ = ⨂ ∣ X i ∣ , so ν A N = ν X 1 ∗ ⋯ ∗ ν X N \nu_{A_N}=\nu_{X_1}\ast\cdots\ast\nu_{X_N} ν A N = ν X 1 ∗ ⋯ ∗ ν X N ; for p < ∞ p<\infty p < ∞ use the additivity of C p C_p C p under convolution, and for p = ∞ p=\infty p = ∞ that the extreme points of the supports add (equivalently, the case p = ∞ p=\infty p = ∞ follows from the finite-p p p case and fact (iii), C p → OLC C_p\to\operatorname{OLC} C p → OLC ; for positive definite matrices it is also Chapter 1). If ∣ X ∣ |X| ∣ X ∣ is not a multiple of 1 1 1 , then p ↦ C p ( X ) p\mapsto C_p(X) p ↦ C p ( X ) is strictly increasing, since the L p L^p L p -norms of a nonconstant function on a probability space are strictly increasing in p p p .
The tensor product is the commuting model: the factors X i ⊗ 1 ⊗ ⋯ X_i\otimes1\otimes\cdots X i ⊗ 1 ⊗ ⋯ commute, and the product of commuting independent positive elements behaves in the same way. Lemma 5.2.1 gives, for arbitrary products in M \mathcal{M} M , the bound C p / N ( A N ) ≤ ∑ i C p ( X i ) C_{p/N}(A_N)\leq\sum_i C_p(X_i) C p / N ( A N ) ≤ ∑ i C p ( X i ) ; in the tensor case C p / N ( A N ) = N C p / N ( X ) ≤ N C p ( X ) C_{p/N}(A_N)=N C_{p/N}(X)\leq N C_p(X) C p / N ( A N ) = N C p / N ( X ) ≤ N C p ( X ) , so the bound loses only the change of exponent.
Matrix independence
Let X 1 , X 2 , … X_1,X_2,\ldots X 1 , X 2 , … be i.i.d. random elements of G L d ( C ) \mathrm{GL}_d(\mathbb{C}) GL d ( C ) with E ln + ∥ X 1 ∥ < ∞ \mathbb{E}\ln^{+}\lVert X_1\rVert<\infty E ln + ∥ X 1 ∥ < ∞ and E ln + ∥ X 1 − 1 ∥ < ∞ \mathbb{E}\ln^{+}\lVert X_1^{-1}\rVert<\infty E ln + ∥ X 1 − 1 ∥ < ∞ , A N = X N ⋯ X 1 A_N=X_N\cdots X_1 A N = X N ⋯ X 1 , and let γ 1 ≥ γ d \gamma_1\geq\gamma_d γ 1 ≥ γ d be the largest and the smallest Lyapunov exponent: 1 N ln ∥ A N ∥ → γ 1 \frac{1}{N}\ln\lVert A_N\rVert\to\gamma_1 N 1 ln ∥ A N ∥ → γ 1 and 1 N ln ∥ A N − 1 ∥ → − γ d \frac{1}{N}\ln\lVert A_N^{-1}\rVert\to-\gamma_d N 1 ln ∥ A N − 1 ∥ → − γ d almost surely [69 ] . Contrasts in M d ( C ) M_d(\mathbb{C}) M d ( C ) are taken with the normalised trace; for a symmetric gauge function Φ \Phi Φ on R d \mathbb{R}^d R d normalised by Φ ( e 1 ) = 1 \Phi(e_1)=1 Φ ( e 1 ) = 1 , with s ( A ) s(A) s ( A ) the vector of singular values and κ Φ ( A ) = Φ ( s ( A ) ) Φ ( s ( A − 1 ) ) \kappa_\Phi(A)=\Phi(s(A))\Phi(s(A^{-1})) κ Φ ( A ) = Φ ( s ( A )) Φ ( s ( A − 1 )) , we use the normalised contrast C Φ ( A ) = 1 2 ln ( κ Φ ( A ) / κ Φ ( I ) ) C_\Phi(A)=\frac{1}{2}\ln(\kappa_\Phi(A)/\kappa_\Phi(I)) C Φ ( A ) = 2 1 ln ( κ Φ ( A ) / κ Φ ( I )) , where κ Φ ( I ) = Φ ( 1 ) 2 \kappa_\Phi(I)=\Phi(1)^2 κ Φ ( I ) = Φ ( 1 ) 2 .
Theorem 5.4.1. (a) Almost surely , 1 N OLC ( A N ) → 1 2 ( γ 1 − γ d ) \frac{1}{N}\operatorname{OLC}(A_N)\to\frac{1}{2}(\gamma_1-\gamma_d) N 1 OLC ( A N ) → 2 1 ( γ 1 − γ d ) , and the same limit holds for C p ( A N ) C_p(A_N) C p ( A N ) , 0 < p < ∞ 0<p<\infty 0 < p < ∞ , and for C Φ ( A N ) C_\Phi(A_N) C Φ ( A N ) for every symmetric gauge function Φ \Phi Φ .
(b) More precisely , 0 ≤ OLC ( A N ) − C p ( A N ) ≤ ln d p 0\leq\operatorname{OLC}(A_N)-C_p(A_N)\leq\frac{\ln d}{p} 0 ≤ OLC ( A N ) − C p ( A N ) ≤ p l n d and 0 ≤ OLC ( A N ) − C Φ ( A N ) ≤ ln Φ ( 1 ) 0\leq\operatorname{OLC}(A_N)-C_\Phi(A_N)\leq\ln\Phi(1) 0 ≤ OLC ( A N ) − C Φ ( A N ) ≤ ln Φ ( 1 ) for every N N N .
(c) If, in addition , ln ∥ X 1 ∥ \ln\lVert X_1\rVert ln ∥ X 1 ∥ and ln ∥ X 1 − 1 ∥ \ln\lVert X_1^{-1}\rVert ln ∥ X 1 − 1 ∥ have finite second moments and the support of the law of X 1 X_1 X 1 generates a semigroup that is Zariski dense in G L d ( C ) \mathrm{GL}_d(\mathbb{C}) GL d ( C ) regarded as a real algebraic group (or in G L d ( R ) \mathrm{GL}_d(\mathbb{R}) GL d ( R ) for real factors), then ( OLC ( A N ) − N 2 ( γ 1 − γ d ) ) / N \left(\operatorname{OLC}(A_N)-\frac{N}{2}(\gamma_1-\gamma_d)\right)/\sqrt{N} ( OLC ( A N ) − 2 N ( γ 1 − γ d ) ) / N converges in distribution to a centred Gaussian law, nondegenerate for d ≥ 2 d\geq2 d ≥ 2 .
Proof. (a) OLC is subadditive (Lemma 5.2.1 with all p i = ∞ p_i=\infty p i = ∞ ; see also [2 ] ), so N ↦ OLC ( X N ⋯ X 1 ) N\mapsto\operatorname{OLC}(X_N\cdots X_1) N ↦ OLC ( X N ⋯ X 1 ) is a subadditive process over the shift, and Kingman’s subadditive ergodic theorem [70 ] gives an almost sure limit of OLC ( A N ) / N \operatorname{OLC}(A_N)/N OLC ( A N ) / N ; it equals 1 2 ( γ 1 − γ d ) \frac{1}{2}(\gamma_1-\gamma_d) 2 1 ( γ 1 − γ d ) by the theorem of Furstenberg and Kesten applied to ln ∥ A N ∥ \ln\lVert A_N\rVert ln ∥ A N ∥ and ln ∥ A N − 1 ∥ \ln\lVert A_N^{-1}\rVert ln ∥ A N − 1 ∥ . The other contrasts follow from (b). (b) If s 1 ≥ ⋯ ≥ s d s_1\geq\cdots\geq s_d s 1 ≥ ⋯ ≥ s d are the singular values of A A A , then τ ( ∣ A ∣ p ) ≥ s 1 p / d \tau(|A|^p)\geq s_1^p/d τ ( ∣ A ∣ p ) ≥ s 1 p / d and τ ( ∣ A ∣ − p ) ≥ s d − p / d \tau(|A|^{-p})\geq s_d^{-p}/d τ ( ∣ A ∣ − p ) ≥ s d − p / d , so C p ( A ) ≥ OLC ( A ) − ln d p C_p(A)\geq\operatorname{OLC}(A)-\frac{\ln d}{p} C p ( A ) ≥ OLC ( A ) − p l n d ; and s 1 ≤ Φ ( s ) ≤ Φ ( 1 ) s 1 s_1\leq\Phi(s)\leq\Phi(1)s_1 s 1 ≤ Φ ( s ) ≤ Φ ( 1 ) s 1 for the normalised Φ \Phi Φ , and likewise for A − 1 A^{-1} A − 1 , so κ ( A ) ≤ κ Φ ( A ) ≤ Φ ( 1 ) 2 κ ( A ) \kappa(A)\leq\kappa_\Phi(A)\leq\Phi(1)^2\kappa(A) κ ( A ) ≤ κ Φ ( A ) ≤ Φ ( 1 ) 2 κ ( A ) , which gives both bounds for C Φ C_\Phi C Φ . Finally C p ≤ OLC C_p\leq\operatorname{OLC} C p ≤ OLC because p ↦ C p p\mapsto C_p p ↦ C p increases to OLC. (c) 2 OLC ( A N ) = ln s 1 ( A N ) − ln s d ( A N ) 2\operatorname{OLC}(A_N)=\ln s_1(A_N)-\ln s_d(A_N) 2 OLC ( A N ) = ln s 1 ( A N ) − ln s d ( A N ) is a linear functional of the Cartan projection of A N A_N A N . The central limit theorem for the Cartan projection is due to Goldsheid and Guivarc’h [71 ] under exponential moments and to Benoist and Quint [72 Theorems 4.11, 4.16] (see also [73 ] ) under a second moment; for Zariski dense laws the Gaussian is nondegenerate on the Cartan subspace of the semisimple part, on which ln s 1 − ln s d \ln s_1-\ln s_d ln s 1 − ln s d does not vanish for d ≥ 2 d\geq2 d ≥ 2 . For the norm cocycle alone see Le Page [83 ] .
Free independence
In this section X 1 , X 2 , ⋯ ∈ M X_1,X_2,\cdots\in\mathcal{M} X 1 , X 2 , ⋯ ∈ M are ∗ * ∗ -free and invertible, and A N = X N ⋯ X 1 A_N=X_N\cdots X_1 A N = X N ⋯ X 1 .
Proposition 5.5.1. For ∗ * ∗ -free X 1 , … , X N X_1,\ldots,X_N X 1 , … , X N , not necessarily identically distributed,
τ ( ∣ A N ∣ 2 ) = ∏ i τ ( ∣ X i ∣ 2 ) , τ ( ∣ A N − 1 ∣ 2 ) = ∏ i τ ( ∣ X i − 1 ∣ 2 ) , C 2 ( A N ) = ∑ i C 2 ( X i ) . \tau(|A_N|^2)=\prod_i\tau(|X_i|^2),\qquad\tau(|A_N^{-1}|^2)=\prod_i\tau(|X_i^{-1}|^2),\qquad C_2(A_N)=\sum_i C_2(X_i). τ ( ∣ A N ∣ 2 ) = i ∏ τ ( ∣ X i ∣ 2 ) , τ ( ∣ A N − 1 ∣ 2 ) = i ∏ τ ( ∣ X i − 1 ∣ 2 ) , C 2 ( A N ) = i ∑ C 2 ( X i ) .
Proof. τ ( ∣ A N ∣ 2 ) = τ ( A N − 1 ∗ X N ∗ X N A N − 1 ) = τ ( X N ∗ X N ⋅ A N − 1 A N − 1 ∗ ) \tau(|A_N|^2)=\tau(A_{N-1}^*X_N^*X_NA_{N-1})=\tau(X_N^*X_N\cdot A_{N-1}A_{N-1}^*) τ ( ∣ A N ∣ 2 ) = τ ( A N − 1 ∗ X N ∗ X N A N − 1 ) = τ ( X N ∗ X N ⋅ A N − 1 A N − 1 ∗ ) , and X N ∗ X N X_N^*X_N X N ∗ X N is free from A N − 1 A N − 1 ∗ ∈ W ∗ ( X 1 , … , X N − 1 ) A_{N-1}A_{N-1}^*\in W^*(X_1,\ldots,X_{N-1}) A N − 1 A N − 1 ∗ ∈ W ∗ ( X 1 , … , X N − 1 ) , so the trace factorises; induct on N N N . For the inverse, ∣ A N − 1 ∣ 2 = ( X N A N − 1 A N − 1 ∗ X N ∗ ) − 1 |A_N^{-1}|^2=(X_NA_{N-1}A_{N-1}^*X_N^*)^{-1} ∣ A N − 1 ∣ 2 = ( X N A N − 1 A N − 1 ∗ X N ∗ ) − 1 , so τ ( ∣ A N − 1 ∣ 2 ) = τ ( ( X N ∗ X N ) − 1 ⋅ ∣ A N − 1 − 1 ∣ 2 ) \tau(|A_N^{-1}|^2)=\tau((X_N^*X_N)^{-1}\cdot|A_{N-1}^{-1}|^2) τ ( ∣ A N − 1 ∣ 2 ) = τ (( X N ∗ X N ) − 1 ⋅ ∣ A N − 1 − 1 ∣ 2 ) , with the same freeness, and τ ( ( X N ∗ X N ) − 1 ) = τ ( ( X N X N ∗ ) − 1 ) = τ ( ∣ X N − 1 ∣ 2 ) \tau((X_N^*X_N)^{-1})=\tau((X_NX_N^*)^{-1})=\tau(|X_N^{-1}|^2) τ (( X N ∗ X N ) − 1 ) = τ (( X N X N ∗ ) − 1 ) = τ ( ∣ X N − 1 ∣ 2 ) .
By Theorem 5.5.2 (d) below and the strict monotonicity of p ↦ C p ( X ) p\mapsto C_p(X) p ↦ C p ( X ) , p = 2 p=2 p = 2 is the only exponent for which such an identity holds for identically distributed factors with non-scalar X ∗ X X^*X X ∗ X .
For the rest of the section the X i X_i X i are identically distributed. Let μ \mu μ be the distribution of X ∗ X X^*X X ∗ X and μ − 1 \mu^{-1} μ − 1 that of ( X ∗ X ) − 1 (X^*X)^{-1} ( X ∗ X ) − 1 , and put
m = τ ( X ∗ X ) , m ′ = τ ( ( X ∗ X ) − 1 ) , v = Var ( μ ) m 2 , v ′ = Var ( μ − 1 ) m ′ 2 , C 2 ( X ) = 1 4 ln ( m m ′ ) . m=\tau(X^*X),\qquad m'=\tau((X^*X)^{-1}),\qquad v=\frac{\operatorname{Var}(\mu)}{m^2},\qquad v'=\frac{\operatorname{Var}(\mu^{-1})}{m'^2},\qquad C_2(X)=\frac{1}{4}\ln(mm'). m = τ ( X ∗ X ) , m ′ = τ (( X ∗ X ) − 1 ) , v = m 2 Var ( μ ) , v ′ = m ′2 Var ( μ − 1 ) , C 2 ( X ) = 4 1 ln ( m m ′ ) .
Since X X ∗ XX^* X X ∗ and X ∗ X X^*X X ∗ X have the same distribution, m ′ = τ ( ∣ X − 1 ∣ 2 ) m'=\tau(|X^{-1}|^2) m ′ = τ ( ∣ X − 1 ∣ 2 ) .
Theorem 5.5.2. Assume that X ∗ X X^*X X ∗ X is not a multiple of 1 1 1 . Then:
(a) OLC ( A N ) = N C 2 ( X ) + 1 2 ln N + 1 2 + 1 4 ln ( v v ′ ) + o ( 1 ) \operatorname{OLC}(A_N)=N C_2(X)+\frac{1}{2}\ln N+\frac{1}{2}+\frac{1}{4}\ln(vv')+o(1) OLC ( A N ) = N C 2 ( X ) + 2 1 ln N + 2 1 + 4 1 ln ( v v ′ ) + o ( 1 ) .
(b) For every p ∈ [ 2 , ∞ ) p\in[2,\infty) p ∈ [ 2 , ∞ ) , C p ( A N ) = N C 2 ( X ) + ( 1 2 − 1 p ) ln N + O ( 1 ) C_p(A_N)=N C_2(X)+\left(\frac{1}{2}-\frac{1}{p}\right)\ln N+O(1) C p ( A N ) = N C 2 ( X ) + ( 2 1 − p 1 ) ln N + O ( 1 ) ; for p = 2 k p=2k p = 2 k , k ∈ N k\in\mathbb{N} k ∈ N , the O ( 1 ) O(1) O ( 1 ) term is κ p + O ( N − 1 ) \kappa_p+O(N^{-1}) κ p + O ( N − 1 ) with
κ p = 1 2 p [ ( p 2 − 1 ) ln v v ′ p 2 4 − 2 ln Γ ( 1 + p 2 ) ] . \kappa_p=\frac{1}{2p}\left[\left(\frac{p}{2}-1\right)\ln\frac{vv'p^2}{4}-2\ln\Gamma\left(1+\frac{p}{2}\right)\right]. κ p = 2 p 1 [ ( 2 p − 1 ) ln 4 v v ′ p 2 − 2 ln Γ ( 1 + 2 p ) ] . (c) For every p ∈ ( 0 , 2 ) p\in(0,2) p ∈ ( 0 , 2 ) , N C 2 ( X ) + ( 1 2 − 1 p ) ln N − O ( 1 ) ≤ C p ( A N ) ≤ N C 2 ( X ) N C_2(X)+\left(\frac{1}{2}-\frac{1}{p}\right)\ln N-O(1)\leq C_p(A_N)\leq N C_2(X) N C 2 ( X ) + ( 2 1 − p 1 ) ln N − O ( 1 ) ≤ C p ( A N ) ≤ N C 2 ( X ) .
(d) Consequently lim N → ∞ C p ( A N ) / N = C 2 ( X ) \lim_{N\to\infty}C_p(A_N)/N=C_2(X) lim N → ∞ C p ( A N ) / N = C 2 ( X ) for every p ∈ ( 0 , ∞ ] p\in(0,\infty] p ∈ ( 0 , ∞ ] .
Proof. Step 1: distributions. The distribution of ∣ A N ∣ 2 |A_N|^2 ∣ A N ∣ 2 is μ ⊠ N \mu^{\boxtimes N} μ ⊠ N . Indeed A N ∗ A N = A N − 1 ∗ Y A N − 1 A_N^*A_N=A_{N-1}^*Y A_{N-1} A N ∗ A N = A N − 1 ∗ Y A N − 1 with Y = X N ∗ X N Y=X_N^*X_N Y = X N ∗ X N , whose moments are τ ( ( Y 1 / 2 A N − 1 A N − 1 ∗ Y 1 / 2 ) k ) \tau((Y^{1/2}A_{N-1}A_{N-1}^*Y^{1/2})^k) τ (( Y 1/2 A N − 1 A N − 1 ∗ Y 1/2 ) k ) ; the element A N − 1 A N − 1 ∗ A_{N-1}A_{N-1}^* A N − 1 A N − 1 ∗ has the same distribution as A N − 1 ∗ A N − 1 A_{N-1}^*A_{N-1} A N − 1 ∗ A N − 1 and is free from Y Y Y , and both are bounded and positive, so the distribution is μ ⊠ μ ⊠ ( N − 1 ) \mu\boxtimes\mu^{\boxtimes(N-1)} μ ⊠ μ ⊠ ( N − 1 ) . Similarly ∣ A N − 1 ∣ 2 = ( A N A N ∗ ) − 1 |A_N^{-1}|^2=(A_NA_N^*)^{-1} ∣ A N − 1 ∣ 2 = ( A N A N ∗ ) − 1 has distribution ( μ ⊠ N ) − 1 = ( μ − 1 ) ⊠ N (\mu^{\boxtimes N})^{-1}=(\mu^{-1})^{\boxtimes N} ( μ ⊠ N ) − 1 = ( μ − 1 ) ⊠ N , since ( a 1 / 2 b a 1 / 2 ) − 1 = a − 1 / 2 b − 1 a − 1 / 2 (a^{1/2}ba^{1/2})^{-1}=a^{-1/2}b^{-1}a^{-1/2} ( a 1/2 b a 1/2 ) − 1 = a − 1/2 b − 1 a − 1/2 .
Step 2: norms. Let μ 1 \mu_1 μ 1 be the image of μ \mu μ under x ↦ x / m x\mapsto x/m x ↦ x / m , a compactly supported measure on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) with mean 1 1 1 and variance v > 0 v>0 v > 0 . By Kargin [65 ] (for integer powers) and Deng–Zhong [66 ] , the largest point of the support of μ 1 ⊠ N \mu_1^{\boxtimes N} μ 1 ⊠ N is e v N ( 1 + o ( 1 ) ) e^{vN}(1+o(1)) e v N ( 1 + o ( 1 )) . As μ ⊠ N \mu^{\boxtimes N} μ ⊠ N is the image of μ 1 ⊠ N \mu_1^{\boxtimes N} μ 1 ⊠ N under x ↦ m N x x\mapsto m^Nx x ↦ m N x ,
∥ A N ∥ 2 = m N e v N ( 1 + o ( 1 ) ) , ∥ A N − 1 ∥ 2 = m ′ N e v ′ N ( 1 + o ( 1 ) ) , (10) \lVert A_N\rVert^2=m^N e^{vN}(1+o(1)),\qquad\lVert A_N^{-1}\rVert^2=m'^N e^{v'N}(1+o(1)), \tag*{(10)} ∥ A N ∥ 2 = m N e v N ( 1 + o ( 1 )) , ∥ A N − 1 ∥ 2 = m ′ N e v ′ N ( 1 + o ( 1 )) , (10)
the second by the same argument for μ − 1 \mu^{-1} μ − 1 . Hence OLC ( A N ) = 1 4 ln ( ∥ A N ∥ 2 ∥ A N − 1 ∥ 2 ) \operatorname{OLC}(A_N)=\frac{1}{4}\ln(\lVert A_N\rVert^2\lVert A_N^{-1}\rVert^2) OLC ( A N ) = 4 1 ln (∥ A N ∥ 2 ∥ A N − 1 ∥ 2 ) is as in (a).
Step 3: integer moments. Normalise m = 1 m=1 m = 1 . Then S μ ⊠ N = S μ N S_{\mu^{\boxtimes N}}=S_\mu^N S μ ⊠ N = S μ N , so χ μ ⊠ N ( w ) = w 1 + w S μ ( w ) N \chi_{\mu^{\boxtimes N}}(w)=\frac{w}{1+w}S_\mu(w)^N χ μ ⊠ N ( w ) = 1 + w w S μ ( w ) N , i.e. w = z φ ( w ) w=z\varphi(w) w = z φ ( w ) with z = χ μ ⊠ N ( w ) z=\chi_{\mu^{\boxtimes N}}(w) z = χ μ ⊠ N ( w ) and φ ( w ) = ( 1 + w ) S μ ( w ) − N \varphi(w)=(1+w)S_\mu(w)^{-N} φ ( w ) = ( 1 + w ) S μ ( w ) − N . By the Lagrange inversion formula, for k ≥ 1 k\geq1 k ≥ 1 ,
m k ( μ ⊠ N ) = [ z k ] ψ μ ⊠ N ( z ) = 1 k [ w k − 1 ] ( 1 + w ) k S μ ( w ) − N k . m_k(\mu^{\boxtimes N})=[z^k]\psi_{\mu^{\boxtimes N}}(z)=\frac{1}{k}[w^{k-1}](1+w)^kS_\mu(w)^{-Nk}. m k ( μ ⊠ N ) = [ z k ] ψ μ ⊠ N ( z ) = k 1 [ w k − 1 ] ( 1 + w ) k S μ ( w ) − N k .
Write S μ ( w ) − 1 = exp ( v w + w 2 g ( w ) ) S_\mu(w)^{-1}=\exp(vw+w^2g(w)) S μ ( w ) − 1 = exp ( v w + w 2 g ( w )) with g g g holomorphic near 0 0 0 . The coefficient of w k − 1 w^{k-1} w k − 1 in ( 1 + w ) k exp ( N k ( v w + w 2 g ( w ) ) ) (1+w)^k\exp(Nk(vw+w^2g(w))) ( 1 + w ) k exp ( N k ( v w + w 2 g ( w ))) is a polynomial in N N N of degree k − 1 k-1 k − 1 : a monomial N j w i N^jw^i N j w i arises only with i ≥ j i\geq j i ≥ j , so the term N k − 1 N^{k-1} N k − 1 comes only from ( N k v w ) k − 1 / ( k − 1 ) ! (Nkvw)^{k-1}/(k-1)! ( N k v w ) k − 1 / ( k − 1 )! . Hence
m k ( μ ⊠ N ) = k k − 1 k ! ( v N ) k − 1 ( 1 + O ( N − 1 ) ) , (11) m_k(\mu^{\boxtimes N})=\frac{k^{k-1}}{k!}(vN)^{k-1}(1+O(N^{-1})), \tag*{(11)} m k ( μ ⊠ N ) = k ! k k − 1 ( v N ) k − 1 ( 1 + O ( N − 1 )) , (11)
and for general m m m the right-hand side is multiplied by m k N m^{kN} m k N . (For k = 2 k=2 k = 2 this is the exact identity m 2 = 1 + v N m_2=1+vN m 2 = 1 + v N ; for the Marchenko–Pastur law, where S μ ( w ) = ( 1 + w ) − 1 S_\mu(w)=(1+w)^{-1} S μ ( w ) = ( 1 + w ) − 1 and v = 1 v=1 v = 1 , the exact formula above gives the Fuss–Catalan numbers.) Thus ln τ ( ∣ A N ∣ k ) = k N ln m + ( k − 1 ) ln ( v k N ) − ln Γ ( k + 1 ) + O ( N − 1 ) \ln\tau(|A_N|^k)=kN\ln m+(k-1)\ln(vkN)-\ln\Gamma(k+1)+O(N^{-1}) ln τ ( ∣ A N ∣ k ) = k N ln m + ( k − 1 ) ln ( v k N ) − ln Γ ( k + 1 ) + O ( N − 1 ) .
Step 4: interpolation. Write ℓ s = ln τ ( ∣ A N ∣ s ) \ell_s=\ln\tau(|A_N|^s) ℓ s = ln τ ( ∣ A N ∣ s ) for s > 0 s>0 s > 0 and ℓ ∞ = ln ∥ A N ∥ \ell_\infty=\ln\lVert A_N\rVert ℓ ∞ = ln ∥ A N ∥ , so that ℓ 2 = N ln m \ell_2=N\ln m ℓ 2 = N ln m and, by (10) , ℓ ∞ = N 2 ln m + 1 2 ln N + O ( 1 ) \ell_\infty=\frac{N}{2}\ln m+\frac{1}{2}\ln N+O(1) ℓ ∞ = 2 N ln m + 2 1 ln N + O ( 1 ) . The inequality τ ( ∣ A ∣ s ) ≤ ∥ A ∥ s − t τ ( ∣ A ∣ t ) \tau(|A|^s)\leq\lVert A\rVert^{s-t}\tau(|A|^t) τ ( ∣ A ∣ s ) ≤ ∥ A ∥ s − t τ ( ∣ A ∣ t ) for s ≥ t s\geq t s ≥ t gives:
for p ≥ 2 p \ge2 p ≥ 2 : ℓ p ≤ ℓ 2 + ( p − 2 ) ℓ ∞ \ell_p \le\ell_2 + (p - 2)\ell_\infty ℓ p ≤ ℓ 2 + ( p − 2 ) ℓ ∞ , and, for an integer k ≥ p / 2 k \ge p/2 k ≥ p /2 , ℓ p ≥ ℓ 2 k − ( 2 k − p ) ℓ ∞ \ell_p \ge\ell_{2k} - (2k - p)\ell_\infty ℓ p ≥ ℓ 2 k − ( 2 k − p ) ℓ ∞ ;
Inserting the asymptotics of ℓ 2 \ell_2 ℓ 2 , ℓ 2 k \ell_{2k} ℓ 2 k and ℓ ∞ \ell_\infty ℓ ∞ , the upper and lower bounds for p ≥ 2 p \ge2 p ≥ 2 both equal 1 p ℓ p = N 2 ln m + ( 1 2 − 1 p ) ln N + O ( 1 ) \frac{1}{p}\ell_p = \frac{N}{2}\ln m + \left(\frac{1}{2} - \frac{1}{p}\right)\ln N + O(1) p 1 ℓ p = 2 N ln m + ( 2 1 − p 1 ) ln N + O ( 1 ) ; for p < 2 p < 2 p < 2 the lower bound is the same and the upper bound is N 2 ln m \frac{N}{2}\ln m 2 N ln m . The same holds for A N − 1 A_N^{-1} A N − 1 with m ′ m' m ′ , μ − 1 \mu^{-1} μ − 1 . Since C p ( A N ) = 1 2 ( 1 p ℓ p + 1 p ℓ p ′ ) C_p(A_N) = \frac{1}{2}\left(\frac{1}{p}\ell_p + \frac{1}{p}\ell'_p\right) C p ( A N ) = 2 1 ( p 1 ℓ p + p 1 ℓ p ′ ) , with ℓ ′ \ell' ℓ ′ the analogous quantities for A N − 1 A_N^{-1} A N − 1 , this gives (b) and (c), and (d) follows from (a)–(c). For p = 2 k p = 2k p = 2 k no interpolation is needed: inserting the moment asymptotics for A N A_N A N and A N − 1 A_N^{-1} A N − 1 directly into C 2 k ( A N ) = 1 4 k ( ℓ 2 k + ℓ 2 k ′ ) C_{2k}(A_N) = \frac{1}{4k}(\ell_{2k} + \ell'_{2k}) C 2 k ( A N ) = 4 k 1 ( ℓ 2 k + ℓ 2 k ′ ) gives the constant κ 2 k \kappa_{2k} κ 2 k .
Remark 5.5.3. Kargin [75 ] proved 1 N ln ∥ A N ∥ → 1 2 ln τ ( X ∗ X ) \frac{1}{N}\ln\|A_N\| \to\frac{1}{2}\ln\tau(X^*X) N 1 ln ∥ A N ∥ → 2 1 ln τ ( X ∗ X ) , and [65 , 66 ] the linear growth of the support of μ ⊠ N \mu^{\boxtimes N} μ ⊠ N , so (a) is a direct consequence; the integer moment asymptotics (11) are a short Lagrange inversion computation, and related limit theorems for free multiplicative powers are in [67 ] (for Marchenko–Pastur laws see [68 ] ). Parts (b)–(d), i.e. the Schatten norm asymptotics ∥ A N ∥ p ≍ m N / 2 N 1 2 − 1 p \|A_N\|_p \asymp m^{N/2}N^{\frac{1}{2}-\frac{1}{p}} ∥ A N ∥ p ≍ m N /2 N 2 1 − p 1 for p ≥ 2 p \ge2 p ≥ 2 , the two-sided bounds for p < 2 p < 2 p < 2 , and the collapse of all contrast slopes onto C 2 C_2 C 2 , do not seem to have been stated. The mechanism is visible in (11) : ∣ A N ∣ 2 / m N |A_N|^2/m^N ∣ A N ∣ 2 / m N has mean 1 1 1 , but its mass near the top of the spectrum, at scale N N N , is of order N − 1 N^{-1} N − 1 .
Conjecture 5.5.4. Let μ \mu μ be compactly supported in ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) and not a point mass. For every s > 0 s > 0 s > 0 , with Y N = ∣ A N ∣ 2 / m N Y_N = |A_N|^2/m^N Y N = ∣ A N ∣ 2 / m N and Y N ′ = ∣ A N − 1 ∣ 2 / m ′ N Y'_N = |A_N^{-1}|^2/m'^{N} Y N ′ = ∣ A N − 1 ∣ 2 / m ′ N ,
ln τ ( Y N s ) = ( s − 1 ) ln ( v s N ) − ln Γ ( s + 1 ) + o ( 1 ) , \ln\tau(Y_N^s) = (s - 1)\ln(vsN) - \ln\Gamma(s + 1) + o(1), ln τ ( Y N s ) = ( s − 1 ) ln ( v s N ) − ln Γ ( s + 1 ) + o ( 1 ) , and the same for Y N ′ Y'_N Y N ′ with v ′ v' v ′ in place of v v v . Consequently, for every p ∈ ( 0 , ∞ ) p \in(0,\infty) p ∈ ( 0 , ∞ ) ,
C p ( A N ) = N C 2 ( X ) + ( 1 2 − 1 p ) ln N + κ p + o ( 1 ) , C_p(A_N) = N C_2(X) + \left(\frac{1}{2} - \frac{1}{p}\right)\ln N + \kappa_p + o(1), C p ( A N ) = N C 2 ( X ) + ( 2 1 − p 1 ) ln N + κ p + o ( 1 ) , with κ p \kappa_p κ p as in Theorem 5.5.2 (b).
For integer s = k s = k s = k the first formula is (11) , and the expansion of C p C_p C p is proved for p ∈ 2 N p \in2\mathbb{N} p ∈ 2 N in Theorem 5.5.2 (b); the conjecture concerns p ∉ 2 N p \notin2\mathbb{N} p ∈ / 2 N , in particular p < 2 p < 2 p < 2 . The constants are consistent with Theorem 5.5.2 : κ 2 = 0 \kappa_2 = 0 κ 2 = 0 , and κ p → 1 2 + 1 4 ln ( v v ′ ) \kappa_p \to\frac{1}{2} + \frac{1}{4}\ln(vv') κ p → 2 1 + 4 1 ln ( v v ′ ) as p → ∞ p \to\infty p → ∞ , which is the constant in (a). For the Marchenko–Pastur law, which has 0 0 0 in its support and so lies outside the invertible setting, the moments of μ ⊠ N \mu^{\boxtimes N} μ ⊠ N at real s s s are given by the Raney (Fuss–Catalan) Mellin formula Γ ( ( N + 1 ) s + 1 ) / ( Γ ( s + 1 ) Γ ( N s + 2 ) ) \Gamma((N + 1)s + 1)/(\Gamma(s + 1)\Gamma(Ns + 2)) Γ (( N + 1 ) s + 1 ) / ( Γ ( s + 1 ) Γ ( N s + 2 )) [84 ] , and the analogue of the first formula follows from Stirling’s formula; we checked the Mellin formula numerically for N = 4 N = 4 N = 4 . A plausible route to a proof is the vague convergence of v N ( μ ⊠ N ∘ D v N ) v_N(\mu^{\boxtimes N} \circ D_{v_N}) v N ( μ ⊠ N ∘ D v N ) , D c ( x ) = c x D_c(x) = cx D c ( x ) = c x , to a σ \sigma σ -finite measure with s s s -th moment s s − 1 / Γ ( s + 1 ) s^{s-1}/\Gamma(s + 1) s s − 1 /Γ ( s + 1 ) , together with uniform integrability near 0 0 0 for s < 1 s < 1 s < 1 . Numerical evidence is given in § 8.8 .
The order of the remainder
Table 1.
For C 2 C_2 C 2 the free and the tensor case coincide: both are exactly additive. For fixed N N N , products of Haar-rotated d × d d \times d d × d factors converge to the free model as d → ∞ d \to\infty d → ∞ ; for Haar unitaries the convergence is strong by [85 ] when the deterministic factors D d D_d D d converge strongly (e.g. for discretisations of a fixed spectrum), so that norms converge as well (we do not use the orthogonal case of this); but the limits N → ∞ N \to\infty N → ∞ and d → ∞ d \to\infty d → ∞ need not commute, and for the slopes we argue directly. For free factors the largest of Kargin’s free Lyapunov exponents [86 ] is 1 2 ln τ ( X ∗ X ) = 1 2 ln m \frac{1}{2}\ln\tau(X^*X)=\frac{1}{2}\ln m 2 1 ln τ ( X ∗ X ) = 2 1 ln m , the slope of ln ∥ A N ∥ \ln\lVert A_N\rVert ln ∥ A N ∥ , and by symmetry the slope of ln ∥ A N − 1 ∥ \ln\lVert A_N^{-1}\rVert ln ∥ A N − 1 ∥ is 1 2 ln m ′ \frac{1}{2}\ln m' 2 1 ln m ′ ; together they give the free slope C 2 ( X ) C_2(X) C 2 ( X ) .
Proposition 5.6.1. Let D = diag ( δ 1 , … , δ d ) D=\operatorname{diag}(\delta_1,\ldots,\delta_d) D = diag ( δ 1 , … , δ d ) be positive definite, let U U U be Haar distributed on O ( d ) O(d) O ( d ) (or U ( d ) U(d) U ( d ) ), and let X i X_i X i be i.i.d. copies of X = U D X=UD X = U D . Then
1 2 ( γ 1 − γ d ) = 1 4 E [ ln ∑ i δ i 2 u i 2 + ln ∑ i δ i − 2 u i 2 ] ≤ C 2 ( X ) = 1 4 ln ( τ ( D 2 ) τ ( D − 2 ) ) , \frac{1}{2}(\gamma_1-\gamma_d)=\frac{1}{4}\mathbb{E}\left[\ln\sum_i\delta_i^2u_i^2+\ln\sum_i\delta_i^{-2}u_i^2\right]\le C_2(X)=\frac{1}{4}\ln\left(\tau(D^2)\tau(D^{-2})\right), 2 1 ( γ 1 − γ d ) = 4 1 E [ ln i ∑ δ i 2 u i 2 + ln i ∑ δ i − 2 u i 2 ] ≤ C 2 ( X ) = 4 1 ln ( τ ( D 2 ) τ ( D − 2 ) ) , where u u u is uniformly distributed on the unit sphere of R d \mathbb{R}^d R d (of C d \mathbb{C}^d C d , with u i 2 u_i^2 u i 2 replaced by ∣ u i ∣ 2 \lvert u_i\rvert^2 ∣ u i ∣ 2 ). If D = D d D=D_d D = D d has spectrum in a fixed interval [ a , b ] ⊂ ( 0 , ∞ ) [a,b]\subset(0,\infty) [ a , b ] ⊂ ( 0 , ∞ ) and the empirical distributions of its eigenvalues converge weakly to a measure σ \sigma σ , then both sides converge, as d → ∞ d\to\infty d → ∞ , to
1 4 ln ( ∫ x 2 d σ ∫ x − 2 d σ ) , \frac{1}{4}\ln\left(\int x^2\mathrm{d}\sigma\int x^{-2}\mathrm{d}\sigma\right), 4 1 ln ( ∫ x 2 d σ ∫ x − 2 d σ ) , the free slope of Theorem 5.5.2 .
Proof. The law of X X X is invariant under X ↦ W X X\mapsto WX X ↦ W X for orthogonal (unitary) W W W . Fix a unit vector e e e and put v 0 = e v_0=e v 0 = e , v i = X i v i − 1 / ∥ X i v i − 1 ∥ v_i=X_iv_{i-1}/\lVert X_iv_{i-1}\rVert v i = X i v i − 1 / ∥ X i v i − 1 ∥ . Then ∥ X i v i − 1 ∥ = ∥ D v i − 1 ∥ \lVert X_iv_{i-1}\rVert=\lVert Dv_{i-1}\rVert ∥ X i v i − 1 ∥ = ∥ D v i − 1 ∥ , and since U i U_i U i is Haar and independent of X 1 , … , X i − 1 X_1,\ldots,X_{i-1} X 1 , … , X i − 1 , v i v_i v i is uniformly distributed and independent of v 0 , … , v i − 1 v_0,\ldots,v_{i-1} v 0 , … , v i − 1 . Hence ln ∥ A N e ∥ = ∑ i ≤ N ln ∥ D v i − 1 ∥ \ln\lVert A_Ne\rVert=\sum_{i\le N}\ln\lVert Dv_{i-1}\rVert ln ∥ A N e ∥ = ∑ i ≤ N ln ∥ D v i − 1 ∥ , and by the law of large numbers 1 N ln ∥ A N e ∥ → E ln ∥ D u ∥ \frac{1}{N}\ln\lVert A_Ne\rVert\to\mathbb{E}\ln\lVert Du\rVert N 1 ln ∥ A N e ∥ → E ln ∥ D u ∥ almost surely, for every fixed e e e . Since ∥ A N e ∥ ≤ ∥ A N ∥ ≤ d max j ∥ A N e j ∥ \lVert A_Ne\rVert\le\lVert A_N\rVert\le\sqrt{d}\max_j\lVert A_Ne_j\rVert ∥ A N e ∥ ≤ ∥ A N ∥ ≤ d max j ∥ A N e j ∥ for an orthonormal basis ( e j ) (e_j) ( e j ) , γ 1 = E ln ∥ D u ∥ = 1 2 E ln ∑ i δ i 2 u i 2 \gamma_1=\mathbb{E}\ln\lVert Du\rVert=\frac{1}{2}\mathbb{E}\ln\sum_i\delta_i^2u_i^2 γ 1 = E ln ∥ D u ∥ = 2 1 E ln ∑ i δ i 2 u i 2 (this is the observation of Cohen and Newman [74 ] ). Next, ( A N − 1 ) ∗ = Y N ⋯ Y 1 (A_N^{-1})^*=Y_N\cdots Y_1 ( A N − 1 ) ∗ = Y N ⋯ Y 1 with Y i = X i − ∗ = U i D − 1 Y_i=X_i^{-*}=U_iD^{-1} Y i = X i −∗ = U i D − 1 , so the same argument gives − γ d = lim N 1 N ln ∥ A N − 1 ∥ = E ln ∥ D − 1 u ∥ -\gamma_d=\lim_{N}\frac{1}{N}\ln\lVert A_N^{-1}\rVert=\mathbb{E}\ln\lVert D^{-1}u\rVert − γ d = lim N N 1 ln ∥ A N − 1 ∥ = E ln ∥ D − 1 u ∥ . Since E u i 2 = 1 d \mathbb{E}u_i^2=\frac{1}{d} E u i 2 = d 1 , Jensen’s inequality gives E ln ∑ i δ i ± 2 u i 2 ≤ ln τ ( D ± 2 ) \mathbb{E}\ln\sum_i\delta_i^{\pm2}u_i^2\le\ln\tau(D^{\pm2}) E ln ∑ i δ i ± 2 u i 2 ≤ ln τ ( D ± 2 ) , and τ ( X ∗ X ) = τ ( D 2 ) \tau(X^*X)=\tau(D^2) τ ( X ∗ X ) = τ ( D 2 ) , τ ( ( X ∗ X ) − 1 ) = τ ( D − 2 ) \tau((X^*X)^{-1})=\tau(D^{-2}) τ (( X ∗ X ) − 1 ) = τ ( D − 2 ) . For the limit write u = g / ∥ g ∥ u=g/\lVert g\rVert u = g / ∥ g ∥ with g g g a standard Gaussian vector: ∑ i δ i ± 2 u i 2 \sum_i\delta_i^{\pm2}u_i^2 ∑ i δ i ± 2 u i 2 is the quotient of 1 d ∑ i δ i ± 2 g i 2 \frac{1}{d}\sum_i\delta_i^{\pm2}g_i^2 d 1 ∑ i δ i ± 2 g i 2 , which differs from τ ( D d ± 2 ) \tau(D_d^{\pm2}) τ ( D d ± 2 ) by a quantity of variance at most 2 max ( a − 4 , b 4 ) / d 2\max(a^{-4},b^4)/d 2 max ( a − 4 , b 4 ) / d , and 1 d ∑ i g i 2 → 1 \frac{1}{d}\sum_i g_i^2\to1 d 1 ∑ i g i 2 → 1 ; its logarithm is bounded by 2 max ( ∣ ln a ∣ , ∣ ln b ∣ ) 2\max(\lvert\ln a\rvert,\lvert\ln b\rvert) 2 max (∣ ln a ∣ , ∣ ln b ∣) , so by bounded convergence the expectations converge to ln ∫ x ± 2 d σ \ln\int x^{\pm2}\mathrm{d}\sigma ln ∫ x ± 2 d σ , which is also the limit of ln τ ( D d ± 2 ) \ln\tau(D_d^{\pm2}) ln τ ( D d ± 2 ) .
Thus, for isotropic factors, matrix independence never makes the contrast grow faster than free independence, and the two slopes agree in the limit of large dimension; what distinguishes them for large d d d is the remainder. (For Gaussian factors, which are not of the form U D UD U D with D D D fixed and are not invertible in the free limit, Proposition 5.6.1 does not apply; with entries of variance 1 / d 1/d 1/ d , 1 2 ( γ 1 − γ d ) ∼ 1 4 ln d \frac{1}{2}(\gamma_1-\gamma_d)\sim\frac{1}{4}\ln d 2 1 ( γ 1 − γ d ) ∼ 4 1 ln d [87 ] .)
Decay: convergence to the tracial state
A positive invertible ρ \rho ρ has OLC ( ρ ) = 0 \operatorname{OLC}(\rho)=0 OLC ( ρ ) = 0 iff ρ \rho ρ is a multiple of 1 1 1 . For a positive definite density matrix this means that ρ \rho ρ is the normalised trace, the maximally mixed state; for a probability density on a compact group, that it is the Haar measure. In this section the contrast tends to 0 0 0 .
The Hilbert metric. For positive definite x , y ∈ M d ( C ) x,y\in M_d(\mathbb{C}) x , y ∈ M d ( C ) let M ( x / y ) = inf { t > 0 : x ≤ t y } M(x/y)=\inf\{t>0:x\le ty\} M ( x / y ) = inf { t > 0 : x ≤ t y } and d H ( x , y ) = ln ( M ( x / y ) M ( y / x ) ) d_H(x,y)=\ln(M(x/y)M(y/x)) d H ( x , y ) = ln ( M ( x / y ) M ( y / x )) , the Hilbert projective metric of the cone of positive semidefinite matrices. Since M ( ρ / 1 ) = λ max ( ρ ) M(\rho/1)=\lambda_{\max}(\rho) M ( ρ /1 ) = λ m a x ( ρ ) and M ( 1 / ρ ) = λ min ( ρ ) − 1 M(1/\rho)=\lambda_{\min}(\rho)^{-1} M ( 1/ ρ ) = λ m i n ( ρ ) − 1 ,
d H ( ρ , 1 ) = ln κ ( ρ ) = 2 OLC ( ρ ) . (12) d_H(\rho,1)=\ln\kappa(\rho)=2\operatorname{OLC}(\rho). \tag*{(12)} d H ( ρ , 1 ) = ln κ ( ρ ) = 2 OLC ( ρ ) . (12)
Let Φ \Phi Φ be a unital positive linear map on M d ( C ) M_d(\mathbb{C}) M d ( C ) ; it maps positive definite matrices to positive definite ones, since x ≥ λ min ( x ) 1 x\ge\lambda_{\min}(x)1 x ≥ λ m i n ( x ) 1 implies Φ ( x ) ≥ λ min ( x ) 1 \Phi(x)\ge\lambda_{\min}(x)1 Φ ( x ) ≥ λ m i n ( x ) 1 . Let Δ ( Φ ) = sup { d H ( Φ ( x ) , Φ ( y ) ) : \Delta(\Phi)=\sup\{d_H(\Phi(x),\Phi(y)): Δ ( Φ ) = sup { d H ( Φ ( x ) , Φ ( y )) : x , y > 0 } ∈ [ 0 , ∞ ] x,y>0\} \in[0,\infty] x , y > 0 } ∈ [ 0 , ∞ ] be the projective diameter of its image; Δ ( Φ ) < ∞ \Delta(\Phi)<\infty Δ ( Φ ) < ∞ if, for instance, Φ \Phi Φ maps every nonzero positive semidefinite matrix to a positive definite one.
Proposition 5.7.1. Let Φ \Phi Φ be a unital positive linear map on M d ( C ) M_d(\mathbb{C}) M d ( C ) . Then for every positive definite ρ \rho ρ ,
OLC ( Φ ( ρ ) ) ≤ tanh ( Δ ( Φ ) 4 ) OLC ( ρ ) . \operatorname{OLC}(\Phi(\rho)) \le\tanh\left(\frac{\Delta(\Phi)}{4}\right)\operatorname{OLC}(\rho). OLC ( Φ ( ρ )) ≤ tanh ( 4 Δ ( Φ ) ) OLC ( ρ ) . In particular, if Δ ( Φ ) < ∞ \Delta(\Phi)<\infty Δ ( Φ ) < ∞ , then OLC ( Φ N ( ρ ) ) ≤ tanh ( Δ ( Φ ) / 4 ) N OLC ( ρ ) \operatorname{OLC}(\Phi^N(\rho)) \le\tanh(\Delta(\Phi)/4)^N\operatorname{OLC}(\rho) OLC ( Φ N ( ρ )) ≤ tanh ( Δ ( Φ ) /4 ) N OLC ( ρ ) decays geometrically.
Proof. By the Birkhoff–Hopf theorem [76 , 77 , 88 ] (see [78 ] for the cone of positive semidefinite matrices), d H ( Φ ( x ) , Φ ( y ) ) ≤ tanh ( Δ ( Φ ) / 4 ) d H ( x , y ) d_H(\Phi(x),\Phi(y)) \le\tanh(\Delta(\Phi)/4)d_H(x,y) d H ( Φ ( x ) , Φ ( y )) ≤ tanh ( Δ ( Φ ) /4 ) d H ( x , y ) . Take y = 1 y=1 y = 1 and use Φ ( 1 ) = 1 \Phi(1)=1 Φ ( 1 ) = 1 and (12) .
For unital trace-preserving positive maps on M d ( C ) M_d(\mathbb{C}) M d ( C ) , the Schatten contrasts C p C_p C p , p ≥ 1 p\ge1 p ≥ 1 , are also nonincreasing, so they are Lyapunov functions for unital dynamics. This is a form of Jensen’s trace inequality: let Φ \Phi Φ be unital, positive and trace-preserving, ρ = ∑ i λ i P i \rho=\sum_i\lambda_iP_i ρ = ∑ i λ i P i positive definite with spectral projections P i P_i P i , and ( e k ) (e_k) ( e k ) an orthonormal basis with Φ ( ρ ) e k = β k e k \Phi(\rho)e_k=\beta_ke_k Φ ( ρ ) e k = β k e k . Then β k = ∑ i w k i λ i \beta_k=\sum_iw_{ki}\lambda_i β k = ∑ i w k i λ i with w k i = ⟨ e k , Φ ( P i ) e k ⟩ ≥ 0 w_{ki}=\langle e_k,\Phi(P_i)e_k\rangle\ge0 w k i = ⟨ e k , Φ ( P i ) e k ⟩ ≥ 0 and ∑ i w k i = ⟨ e k , Φ ( 1 ) e k ⟩ = 1 \sum_iw_{ki}=\langle e_k,\Phi(1)e_k\rangle=1 ∑ i w k i = ⟨ e k , Φ ( 1 ) e k ⟩ = 1 , so for every convex f f f on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) , f ( β k ) ≤ ∑ i w k i f ( λ i ) = ⟨ e k , Φ ( f ( ρ ) ) e k ⟩ f(\beta_k)\le\sum_iw_{ki}f(\lambda_i)=\langle e_k,\Phi(f(\rho))e_k\rangle f ( β k ) ≤ ∑ i w k i f ( λ i ) = ⟨ e k , Φ ( f ( ρ )) e k ⟩ ; summing over k k k gives τ ( f ( Φ ( ρ ) ) ) ≤ τ ( Φ ( f ( ρ ) ) ) = τ ( f ( ρ ) ) \tau(f(\Phi(\rho)))\le\tau(\Phi(f(\rho)))=\tau(f(\rho)) τ ( f ( Φ ( ρ ))) ≤ τ ( Φ ( f ( ρ ))) = τ ( f ( ρ )) . With f ( t ) = t p f(t)=t^p f ( t ) = t p , p ≥ 1 p\ge1 p ≥ 1 , and f ( t ) = t − p f(t)=t^{-p} f ( t ) = t − p this gives τ ( Φ ( ρ ) ± p ) ≤ τ ( ρ ± p ) \tau(\Phi(\rho)^{\pm p})\le\tau(\rho^{\pm p}) τ ( Φ ( ρ ) ± p ) ≤ τ ( ρ ± p ) , hence C p ( Φ ( ρ ) ) ≤ C p ( ρ ) C_p(\Phi(\rho))\le C_p(\rho) C p ( Φ ( ρ )) ≤ C p ( ρ ) . Near the tracial state the C p C_p C p measure a different distance: C p ( ρ ) = p 2 ∥ ln ρ − τ ( ln ρ ) ∥ 2 2 ( 1 + o ( 1 ) ) C_p(\rho)=\frac{p}{2}\|\ln\rho-\tau(\ln\rho)\|_2^2(1+o(1)) C p ( ρ ) = 2 p ∥ ln ρ − τ ( ln ρ ) ∥ 2 2 ( 1 + o ( 1 )) is a squared L 2 ( τ ) L^2(\tau) L 2 ( τ ) -distance, whereas OLC ( ρ ) = 1 2 ( max − min ) \operatorname{OLC}(\rho)=\frac{1}{2}(\max-\min) OLC ( ρ ) = 2 1 ( max − min ) of the spectrum of ln ρ \ln\rho ln ρ is of first order.
The circle and the visibility of fringes. Let T = R / 2 π Z \mathbb{T}=\mathbb{R}/2\pi\mathbb{Z} T = R /2 π Z with normalised Lebesgue measure, μ \mu μ a probability measure on T \mathbb{T} T with Fourier coefficients μ ^ ( k ) = ∫ e − i k x d μ ( x ) \widehat{\mu}(k)=\int e^{-ikx}\,d\mu(x) μ ( k ) = ∫ e − ik x d μ ( x ) , and f 0 = 1 + ∑ k ≠ 0 a k e i k x f_0=1+\sum_{k\ne0}a_ke^{ikx} f 0 = 1 + ∑ k = 0 a k e ik x a continuous strictly positive probability density with ∑ k ∣ a k ∣ < ∞ \sum_k|a_k|<\infty ∑ k ∣ a k ∣ < ∞ . The density of μ ∗ N ∗ f 0 d x \mu^{*N}*f_0\,dx μ ∗ N ∗ f 0 d x is f N = 1 + ∑ k ≠ 0 μ ^ ( k ) N a k e i k x f_N=1+\sum_{k\ne0}\widehat{\mu}(k)^Na_ke^{ikx} f N = 1 + ∑ k = 0 μ ( k ) N a k e ik x , a positive invertible element of the finite von Neumann algebra L ∞ ( T ) L^\infty(\mathbb{T}) L ∞ ( T ) , and convolution with μ \mu μ is a unital trace-preserving positive map on it.
Proposition 5.7.2. Let K = { k ≠ 0 : a k ≠ 0 } K=\{k\ne0:a_k\ne0\} K = { k = 0 : a k = 0 } , r = sup k ∈ K ∣ μ ^ ( k ) ∣ r=\sup_{k\in K}|\widehat{\mu}(k)| r = sup k ∈ K ∣ μ ( k ) ∣ , and suppose that ∣ μ ^ ( k ) ∣ = r |\widehat{\mu}(k)|=r ∣ μ ( k ) ∣ = r exactly for k = ± k 0 ∈ K k=\pm k_0\in K k = ± k 0 ∈ K , that r ′ = sup { ∣ μ ^ ( k ) ∣ : k ∈ K ∖ { ± k 0 } } < r r'=\sup\{|\widehat{\mu}(k)|:k\in K\setminus\{\pm k_0\}\}<r r ′ = sup { ∣ μ ( k ) ∣ : k ∈ K ∖ { ± k 0 }} < r , and that 0 < r < 1 0<r<1 0 < r < 1 , with the convention sup ∅ = 0 \sup\varnothing=0 sup ∅ = 0 for r ′ r' r ′ (i.e. μ \mu μ is not supported on a coset of 2 π k 0 Z \frac{2\pi}{k_0}\mathbb{Z} k 0 2 π Z ; this is automatic if μ \mu μ is absolutely continuous). Then, as N → ∞ N\to\infty N → ∞ ,
OMC ( f N ) = max f N − min f N max f N + min f N = 2 ∣ a k 0 ∣ r N ( 1 + o ( 1 ) ) , OLC ( f N ) = 2 ∣ a k 0 ∣ r N ( 1 + o ( 1 ) ) . \operatorname{OMC}(f_N)=\frac{\max f_N-\min f_N}{\max f_N+\min f_N}=2|a_{k_0}|r^N(1+o(1)),\qquad\operatorname{OLC}(f_N)=2|a_{k_0}|r^N(1+o(1)). OMC ( f N ) = max f N + min f N max f N − min f N = 2∣ a k 0 ∣ r N ( 1 + o ( 1 )) , OLC ( f N ) = 2∣ a k 0 ∣ r N ( 1 + o ( 1 )) .
Proof. Since f N f_N f N is real, a − k 0 = a k 0 ‾ a_{-k_0}=\overline{a_{k_0}} a − k 0 = a k 0 and μ ^ ( − k 0 ) = μ ^ ( k 0 ) ‾ \widehat{\mu}(-k_0)=\overline{\widehat{\mu}(k_0)} μ ( − k 0 ) = μ ( k 0 ) , so f N = 1 + 2 Re ( μ ^ ( k 0 ) N a k 0 e i k 0 x ) + R N f_N=1+2\operatorname{Re}(\widehat{\mu}(k_0)^Na_{k_0}e^{ik_0x})+R_N f N = 1 + 2 Re ( μ ( k 0 ) N a k 0 e i k 0 x ) + R N with ∥ R N ∥ ∞ ≤ r ′ N ∑ k ∣ a k ∣ \|R_N\|_\infty\le r'^N\sum_k|a_k| ∥ R N ∥ ∞ ≤ r ′ N ∑ k ∣ a k ∣ . The middle term has maximum 2 ∣ a k 0 ∣ r N 2|a_{k_0}|r^N 2∣ a k 0 ∣ r N and minimum − 2 ∣ a k 0 ∣ r N -2|a_{k_0}|r^N − 2∣ a k 0 ∣ r N , attained at points of T \mathbb{T} T , so max f N − min f N = 4 ∣ a k 0 ∣ r N + O ( r ′ N ) \max f_N-\min f_N=4|a_{k_0}|r^N+O(r'^N) max f N − min f N = 4∣ a k 0 ∣ r N + O ( r ′ N ) and, since r < 1 r<1 r < 1 , max f N + min f N = 2 + O ( r N ) → 2 \max f_N+\min f_N=2+O(r^N)\to2 max f N + min f N = 2 + O ( r N ) → 2 . Finally OLC = artanh OMC \operatorname{OLC}=\operatorname{artanh}\operatorname{OMC} OLC = artanh OMC and artanh u ∼ u \operatorname{artanh}u\sim u artanh u ∼ u .
If μ \mu μ itself has a continuous positive density with absolutely summable Fourier series, one may take f 0 f_0 f 0 to be this density; then f N f_N f N is the density of μ ∗ ( N + 1 ) \mu^{*(N+1)} μ ∗ ( N + 1 ) . In optics, an interference pattern I ( x ) = I 0 ( 1 + V cos ( k 0 x + ϕ ) ) I(x)=I_0(1+V\cos(k_0x+\phi)) I ( x ) = I 0 ( 1 + V cos ( k 0 x + ϕ )) has Michelson contrast (visibility) V V V , and a random translation x ↦ x − θ x\mapsto x-\theta x ↦ x − θ with θ ∼ μ \theta\sim\mu θ ∼ μ multiplies V V V by ∣ μ ^ ( k 0 ) ∣ |\widehat{\mu}(k_0)| ∣ μ ( k 0 ) ∣ [80 Ch. 10] ; Proposition 5.7.2 says that N N N independent perturbations reduce the visibility by ∣ μ ^ ( k 0 ) ∣ N |\widehat{\mu}(k_0)|^N ∣ μ ( k 0 ) ∣ N , provided this harmonic dominates. (For Gaussian phase noise, ∣ μ ^ ( k ) ∣ = e − σ 2 k 2 / 2 |\widehat{\mu}(k)|=e^{-\sigma^2k^2/2} ∣ μ ( k ) ∣ = e − σ 2 k 2 /2 and higher harmonics decay faster, but in general they need not.) The decay of the contrast to 0 is the compact-group counterpart of the central limit theorem: on R \mathbb{R} R convolution powers spread out, on a compact group they converge to the Haar measure under the condition of Kawada and Itô [79 ] , and the Michelson contrast measures the distance.
Random unital channels. Let Φ 1 , Φ 2 , … \Phi_1,\Phi_2,\ldots Φ 1 , Φ 2 , … be i.i.d. random unital trace-preserving positive maps on M d ( C ) M_d(\mathbb{C}) M d ( C ) , d ≥ 2 d \ge2 d ≥ 2 , and ρ N = Φ N ⋯ Φ 1 ( ρ ) \rho_N=\Phi_N\cdots\Phi_1(\rho) ρ N = Φ N ⋯ Φ 1 ( ρ ) . The real space V V V of traceless self-adjoint matrices is invariant under every Φ i \Phi_i Φ i , and ρ N − τ ( ρ ) 1 = Φ N ⋯ Φ 1 ( ρ − τ ( ρ ) 1 ) \rho_N-\tau(\rho)1=\Phi_N\cdots\Phi_1(\rho-\tau(\rho)1) ρ N − τ ( ρ ) 1 = Φ N ⋯ Φ 1 ( ρ − τ ( ρ ) 1 ) . We equip V V V with the operator norm; since − ∥ x ∥ ≤ x ≤ ∥ x ∥ -\lVert x\rVert\le x \le\lVert x\rVert − ∥ x ∥ ≤ x ≤ ∥ x ∥ implies − ∥ x ∥ ≤ Φ ( x ) ≤ ∥ x ∥ -\lVert x\rVert\le\Phi(x) \le\lVert x\rVert − ∥ x ∥ ≤ Φ ( x ) ≤ ∥ x ∥ , each Φ i ∣ V \Phi_i|_V Φ i ∣ V is a contraction. Let γ = lim N 1 N ln ∥ Φ N ⋯ Φ 1 ∣ V ∥ ∈ [ − ∞ , 0 ] \gamma=\lim_{N}\frac{1}{N}\ln\lVert\Phi_N\cdots\Phi_1|_V\rVert\in[-\infty,0] γ = lim N N 1 ln ∥ Φ N ⋯ Φ 1 ∣ V ∥ ∈ [ − ∞ , 0 ] be the top Lyapunov exponent of the restrictions to V V V , which exists almost surely and is constant by [69 , 70 ] .
Proposition 5.7.3. (a) For every positive definite ρ \rho ρ , almost surely
lim sup N 1 N ln OLC ( ρ N ) ≤ γ . \limsup_{N}\frac{1}{N}\ln\operatorname{OLC}(\rho_N)\le\gamma. N lim sup N 1 ln OLC ( ρ N ) ≤ γ . If γ < 0 \gamma<0 γ < 0 , the contrast decays exponentially.
(b) Suppose, moreover, that Φ 1 ∣ V \Phi_1|_V Φ 1 ∣ V is almost surely invertible with E ln ∥ ( Φ 1 ∣ V ) − 1 ∥ < ∞ \mathbb{E}\ln\lVert(\Phi_1|_V)^{-1}\rVert<\infty E ln ∥( Φ 1 ∣ V ) − 1 ∥ < ∞ , that the group generated by the support of its law acts strongly irreducibly on V V V , and that γ < 0 \gamma<0 γ < 0 . Then for every positive definite ρ \rho ρ that is not a multiple of 1 1 1 , almost surely
lim N 1 N ln OLC ( ρ N ) = γ . \lim_{N}\frac{1}{N}\ln\operatorname{OLC}(\rho_N)=\gamma. N lim N 1 ln OLC ( ρ N ) = γ .
Proof. Normalise τ ( ρ ) = 1 \tau(\rho)=1 τ ( ρ ) = 1 and write ρ N = 1 + X N \rho_N=1+X_N ρ N = 1 + X N with X N ∈ V X_N\in V X N ∈ V . If η = ∥ X N ∥ < 1 \eta=\lVert X_N\rVert<1 η = ∥ X N ∥ < 1 , the eigenvalues of ρ N \rho_N ρ N lie in [ 1 − η , 1 + η ] [1-\eta,1+\eta] [ 1 − η , 1 + η ] and λ max − λ min ≥ η \lambda_{\max}-\lambda_{\min}\ge\eta λ m a x − λ m i n ≥ η (since X N X_N X N is traceless and nonzero, its largest and smallest eigenvalues have opposite signs and one of them has modulus η \eta η ), whence, writing λ max = 1 + a \lambda_{\max}=1+a λ m a x = 1 + a , λ min = 1 + b \lambda_{\min}=1+b λ m i n = 1 + b with b ≤ 0 b\le0 b ≤ 0 and a − b ≥ η a-b\ge\eta a − b ≥ η , we get λ max / λ min ≥ 1 + η \lambda_{\max}/\lambda_{\min}\ge1+\eta λ m a x / λ m i n ≥ 1 + η and η 4 ≤ 1 2 ln ( 1 + η ) ≤ OLC ( ρ N ) ≤ 1 2 ln 1 + η 1 − η ≤ 2 η \frac{\eta}{4}\le\frac{1}{2}\ln(1+\eta)\le\operatorname{OLC}(\rho_N)\le\frac{1}{2}\ln\frac{1+\eta}{1-\eta}\le2\eta 4 η ≤ 2 1 ln ( 1 + η ) ≤ OLC ( ρ N ) ≤ 2 1 ln 1 − η 1 + η ≤ 2 η for η ≤ 1 2 \eta\le\frac{1}{2} η ≤ 2 1 . Thus ln OLC ( ρ N ) = ln ∥ X N ∥ + O ( 1 ) \ln\operatorname{OLC}(\rho_N)=\ln\lVert X_N\rVert+O(1) ln OLC ( ρ N ) = ln ∥ X N ∥ + O ( 1 ) as soon as ∥ X N ∥ ≤ 1 2 \lVert X_N\rVert\le\frac{1}{2} ∥ X N ∥ ≤ 2 1 . (a) If γ = 0 \gamma=0 γ = 0 there is nothing to prove, since OLC ( ρ N ) ≤ OLC ( ρ ) \operatorname{OLC}(\rho_N)\le\operatorname{OLC}(\rho) OLC ( ρ N ) ≤ OLC ( ρ ) by Proposition 5.7.1; if γ < 0 \gamma<0 γ < 0 , use ∥ X N ∥ ≤ ∥ Φ N ⋯ Φ 1 ∣ V ∥ ∥ X 0 ∥ \lVert X_N\rVert\le\lVert\Phi_N\cdots\Phi_1|_V\rVert\lVert X_0\rVert ∥ X N ∥ ≤ ∥ Φ N ⋯ Φ 1 ∣ V ∥ ∥ X 0 ∥ . (b) By Furstenberg’s theorem for strongly irreducible i.i.d. products [81 Part A, Corollary III.3.4] , 1 N ln ∥ Φ N ⋯ Φ 1 ( X 0 ) ∥ → γ \frac{1}{N}\ln\lVert\Phi_N\cdots\Phi_1(X_0)\rVert\to\gamma N 1 ln ∥ Φ N ⋯ Φ 1 ( X 0 )∥ → γ almost surely for every fixed X 0 ≠ 0 X_0\ne0 X 0 = 0 .
For random mixed unitary channels Φ ( x ) = ∑ j w j U j x U j ∗ \Phi(x)=\sum_j w_jU_jxU_j^* Φ ( x ) = ∑ j w j U j x U j ∗ with independent Haar unitaries, a single map need not contract on V V V (its norm on V V V can be 1 1 1 ), but numerically γ ≈ − 0.29 \gamma\approx-0.29 γ ≈ − 0.29 in the example of § 8.8 ; the decay rate is computed by the standard QR algorithm. The three regimes of decay—geometric with ratio tanh ( Δ / 4 ) \tanh(\Delta/4) tanh ( Δ/4 ) , exact exponential 2 ∣ a k 0 ∣ r N 2|a_{k_0}|r^N 2∣ a k 0 ∣ r N , and the Lyapunov rate e γ N e^{\gamma N} e γ N —mirror the three regimes of growth in § 8.6 .
Numerical illustrations
The simulations and numerical values in this section are retained from the component manuscript. They illustrate the stated formulas; the large matrix and free-convolution simulations have not all been rerun for this consolidated version. The circle visibility ratios and the three-dimensional isotropic Lyapunov integrals were recomputed for this version and agree with the displayed rounded values. Values obtained by numerical inversion or quadrature are numerical evaluations, not exact arithmetic certificates.
Free products. Let X = U D X=UD X = U D with U U U Haar and D D D positive with spectrum uniformly distributed on [ 1 , 3 ] [1,3] [ 1 , 3 ] , so that C 2 ( X ) = 1 4 ln 13 9 = 0.0919 C_2(X)=\frac{1}{4}\ln\frac{13}{9}=0.0919 C 2 ( X ) = 4 1 ln 9 13 = 0.0919 . The free limit of 1 2 ln ( ∥ A N ∥ 2 / m N ) \frac{1}{2}\ln(\lVert A_N\rVert^2/m_N) 2 1 ln (∥ A N ∥ 2 / m N ) can be computed exactly from the S S S -transform: the right end of the support of μ X ⊞ N \mu^{\boxplus N}_X μ X ⊞ N is 1 / max w χ N ( w ) 1/\max_w\chi_N(w) 1/ max w χ N ( w ) , the maximum being taken up to the first critical point of χ N ( w ) = w 1 + w S μ 1 ( w ) N \chi_N(w)=\frac{w}{1+w}S_{\mu_1}(w)^N χ N ( w ) = 1 + w w S μ 1 ( w ) N . Simulations use orthogonal matrices of size n n n .
At N = 1024 N=1024 N = 1024 the numerical free value of OLC ( A N ) − N C 2 ( X ) \operatorname{OLC}(A_N)-NC_2(X) OLC ( A N ) − N C 2 ( X ) is 3.4537, against 3.4525 from Theorem 5.5.2 (a). In simulations with n = 500 n=500 n = 500 (eigenvalues of D D D equally spaced in [ 1 , 3 ] [1,3] [ 1 , 3 ] , endpoints included), C 2 ( A N ) / N C_2(A_N)/N C 2 ( A N ) / N stays at the value C 2 ( X ) = 0.0923 C_2(X)=0.0923 C 2 ( X ) = 0.0923 for the discretised spectrum (simulated range 0.0921–0.0923) for all N ≤ 32 N\leq32 N ≤ 32 , in accordance with Proposition 5.5.1, while C 1 ( A N ) / N C_1(A_N)/N C 1 ( A N ) / N increases and C 4 ( A N ) / N C_4(A_N)/N C 4 ( A N ) / N and OLC ( A N ) / N \operatorname{OLC}(A_N)/N OLC ( A N ) / N decrease towards it; the increments of C 1 ( A N ) − N C 2 ( X ) C_1(A_N)-NC_2(X) C 1 ( A N ) − N C 2 ( X ) and C 4 ( A N ) − N C 2 ( X ) C_4(A_N)-NC_2(X) C 4 ( A N ) − N C 2 ( X ) per doubling of N N N approach − 1 2 ln 2 -\frac{1}{2}\ln2 − 2 1 ln 2 and 1 4 ln 2 \frac{1}{4}\ln2 4 1 ln 2 . For Conjecture 5.5.4 we computed fractional moments of μ 1 ⊡ N \mu_1^{\boxdot N} μ 1 ⊡ N numerically, from the density of μ 1 ⊡ N \mu_1^{\boxdot N} μ 1 ⊡ N obtained by inverting χ N \chi_N χ N ; for c N ( s ) = ln τ ( Y N ) − ( s − 1 ) ln N c_N(s)=\ln\tau(Y_N)-(s-1)\ln N c N ( s ) = ln τ ( Y N ) − ( s − 1 ) ln N at N = 1024 N=1024 N = 1024 :
Matrix products. For d = 3 d=3 d = 3 and X = U D X=UD X = U D with U U U Haar orthogonal and D = diag ( 1 , 2 , 3 ) D=\operatorname{diag}(1,2,3) D = diag ( 1 , 2 , 3 ) , the formula of Proposition 5.6.1 gives γ 1 = 0.7111 \gamma_1=0.7111 γ 1 = 0.7111 , γ 3 = 0.4743 \gamma_3=0.4743 γ 3 = 0.4743 and 1 2 ( γ 1 − γ 3 ) = 0.1184 \frac{1}{2}(\gamma_1-\gamma_3)=0.1184 2 1 ( γ 1 − γ 3 ) = 0.1184 , against the free slope C 2 ( X ) = 0.1875 C_2(X)=0.1875 C 2 ( X ) = 0.1875 for the same D D D ; computing OLC ( A N ) \operatorname{OLC}(A_N) OLC ( A N ) through the accumulated QR factors (which agree with ln s 1 \ln s_1 ln s 1 and ln s 3 \ln s_3 ln s 3 up to O P ( 1 ) O_P(1) O P ( 1 ) ) gives 1 N OLC ( A N ) ≈ 0.118 \frac{1}{N}\operatorname{OLC}(A_N)\approx0.118 N 1 OLC ( A N ) ≈ 0.118 for N = 1600 N=1600 N = 1600 , averaged over 200 samples, and standard deviations 0.25 N 0.25\sqrt{N} 0.25 N , 0.23 N 0.23\sqrt{N} 0.23 N , 0.23 N 0.23\sqrt{N} 0.23 N for N = 100 , 400 , 1600 N=100,400,1600 N = 100 , 400 , 1600 . For D d D_d D d with eigenvalues equally spaced in [ 1 , 3 ] [1,3] [ 1 , 3 ] , the slope 1 2 ( γ 1 − γ d ) \frac{1}{2}(\gamma_1-\gamma_d) 2 1 ( γ 1 − γ d ) is 0.0961, 0.0922, 0.0920 for d = 10 , 100 , 1000 d=10,100,1000 d = 10 , 100 , 1000 , approaching the free slope 1 4 ln 13 9 = 0.0919 \frac{1}{4}\ln\frac{13}{9}=0.0919 4 1 ln 9 13 = 0.0919 .
Decay. On the circle, with μ \mu μ the measure with atoms 0, 1, 2.5 of weights 0.5, 0.3, 0.2 and f 0 = 1 + 0.8 cos x + 0.3 cos ( 2 x + 1 ) f_0=1+0.8\cos x+0.3\cos(2x+1) f 0 = 1 + 0.8 cos x + 0.3 cos ( 2 x + 1 ) , one has ∣ μ ^ ( 1 ) ∣ = 0.6248 |\widehat{\mu}(1)|=0.6248 ∣ μ ( 1 ) ∣ = 0.6248 , ∣ μ ^ ( 2 ) ∣ = 0.4394 |\widehat{\mu}(2)|=0.4394 ∣ μ ( 2 ) ∣ = 0.4394 , and the ratio of OMC ( f N ) \operatorname{OMC}(f_N) OMC ( f N ) to 2 ∣ a 1 ∣ ∣ μ ^ ( 1 ) ∣ N 2|a_1||\widehat{\mu}(1)|^N 2∣ a 1 ∣∣ μ ( 1 ) ∣ N is 0.9954, 1.00003, 1 + 1.3 ⋅ 10 − 7 1+1.3\cdot10^{-7} 1 + 1.3 ⋅ 1 0 − 7 for N = 5 , 10 , 20 N=5,10,20 N = 5 , 10 , 20 (the relative correction is O ( r ′ N + ( r ′ / r ) 2 N ) O(r^{\prime N}+(r'/r)^{2N}) O ( r ′ N + ( r ′ / r ) 2 N ) ; the term ( r ′ / r ) 2 N (r'/r)^{2N} ( r ′ / r ) 2 N comes from the second harmonic, which takes equal values at the antipodal extrema of the first; here r 2 < r ′ r^2<r' r 2 < r ′ , so the second term dominates for large N N N , although at N = 5 N=5 N = 5 the first is still the larger). For mixed unitary channels on M 3 ( C ) M_3(\mathbb{C}) M 3 ( C ) with two independent Haar unitaries and weights 0.7, 0.3, a long QR run gives γ = − 0.29 \gamma=-0.29 γ = − 0.29 (a single run of 400 steps gives − 0.28 -0.28 − 0.28 ). Since ∥ ρ N − 1 ∥ \lVert\rho_N-1\rVert ∥ ρ N − 1 ∥ falls below machine precision after about 100 steps, we propagate the traceless part ρ N − 1 \rho_N-1 ρ N − 1 , re-projecting onto the traceless subspace and renormalising at each step; for ρ = 3 diag ( 0.6 , 0.3 , 0.1 ) \rho=3\operatorname{diag}(0.6,0.3,0.1) ρ = 3 diag ( 0.6 , 0.3 , 0.1 ) , so that τ ( ρ ) = 1 \tau(\rho)=1 τ ( ρ ) = 1 , this gives 1 N ln ∥ ρ N − 1 ∥ = − 0.284 \frac{1}{N}\ln\lVert\rho_N-1\rVert=-0.284 N 1 ln ∥ ρ N − 1 ∥ = − 0.284 at N = 400 N=400 N = 400 , which equals 1 N ln OLC ( ρ N ) + O ( 1 N ) \frac{1}{N}\ln\operatorname{OLC}(\rho_N)+O\left(\frac{1}{N}\right) N 1 ln OLC ( ρ N ) + O ( N 1 ) by the proof of Proposition 5.7.3 .
Open problems
Prove Conjecture 5.5.4 , i.e. the asymptotics of fractional moments of μ 1 ⊡ N \mu_1^{\boxdot N} μ 1 ⊡ N , at least the exponent 1 2 − 1 p \frac{1}{2}-\frac{1}{p} 2 1 − p 1 for p < 2 p<2 p < 2 .
Non-identically distributed free factors: by Proposition 5.5.1, C 2 C_2 C 2 is additive; is C p ( A N ) − ∑ i C 2 ( X i ) = O ( ln N ) C_p(A_N)-\sum_i C_2(X_i)=O(\ln N) C p ( A N ) − ∑ i C 2 ( X i ) = O ( ln N ) under uniform bounds?
Products in a factor of type I I 1 \mathrm{II}_1 II 1 that are neither free nor tensor independent (for instance, stationary ergodic sequences of unitarily conjugated elements): do the slopes
lim C p ( A N ) / N \lim C_p(A_N)/N lim C p ( A N ) / N exist, and when do they coincide? Lemma 5.2.1 gives only C p / N ( A N ) ≤ ∑ i C p ( X i ) C_{p/N}(A_N) \le\sum_i C_p(X_i) C p / N ( A N ) ≤ ∑ i C p ( X i ) .
Fluctuations: for Haar-rotated n × n n \times n n × n factors, with n n n and N N N growing jointly, which regime interpolates between the Gaussian N \sqrt{N} N fluctuations of fixed d d d and the deterministic 1 2 ln N \frac{1}{2}\ln N 2 1 ln N of the free limit?
Symmetric cones and Lorentz boosts
Introduction
For a positive invertible operator X X X the operator Michelson contrast and the logarithmic contrast are
OMC ( X ) = λ max ( X ) − λ min ( X ) λ max ( X ) + λ min ( X ) , OLC ( X ) = artanh OMC ( X ) = 1 2 ln λ max ( X ) λ min ( X ) . \operatorname{OMC}(X)=\frac{\lambda_{\max}(X)-\lambda_{\min}(X)}{\lambda_{\max}(X)+\lambda_{\min}(X)}, \qquad\operatorname{OLC}(X)=\operatorname{artanh}\operatorname{OMC}(X)=\frac{1}{2}\ln\frac{\lambda_{\max}(X)}{\lambda_{\min}(X)}. OMC ( X ) = λ m a x ( X ) + λ m i n ( X ) λ m a x ( X ) − λ m i n ( X ) , OLC ( X ) = artanh OMC ( X ) = 2 1 ln λ m i n ( X ) λ m a x ( X ) .
[1 , 2 ] , extended to invertible operators through ∣ X ∣ \lvert X\rvert ∣ X ∣ . In [2 ] it was shown that OLC ( X Y ) ≤ OLC ( X ) + OLC ( Y ) \operatorname{OLC}(XY)\leq\operatorname{OLC}(X)+\operatorname{OLC}(Y) OLC ( X Y ) ≤ OLC ( X ) + OLC ( Y ) , that OLC is the projectivised Thompson distance from [ I ] [I] [ I ] , and that for positive n × n n\times n n × n matrices equality forces commutation exactly for n ≤ 3 n\leq3 n ≤ 3 . This chapter develops four related questions raised by these results.
What is the threshold n = 4 n=4 n = 4 a threshold of? We replace M n ( C ) M_n(\mathbb{C}) M n ( C ) by an arbitrary Euclidean Jordan algebra and the product X Y XY X Y by the quadratic representation P ( x 1 / 2 ) y P(x^{1/2})y P ( x 1/2 ) y , which is X 1 / 2 Y X 1 / 2 X^{1/2}YX^{1/2} X 1/2 Y X 1/2 for matrices. Subadditivity persists (Theorem 6.3.3 ); equality is characterised by common primitive idempotents under the extreme spectral idempotents (Theorem 6.3.4 ); and equality forces operator commutativity if and only if the rank is at most 3 or the algebra is associative (Theorem 6.4.1 ). So the threshold is one of the rank, not of the dimension (for pure tensor products, Proposition 1.3.4 already shows that it is not one of the total dimension): it holds on Lorentz cones of every dimension and on the 27-dimensional exceptional cone, and it extends to JBW-algebras (Theorem 6.10.2).
What is the defect of the Schatten contrasts a defect of? The Schatten contrasts are subadditive only up to a defect (Theorem 6.6.1 (a),(c)). On a symmetric cone the defect of every spectral contrast C Φ C_\Phi C Φ is ln ( Φ ( 1 ) / Φ ( ϵ 1 ) ) \ln(\Phi(1)/\Phi(\epsilon1)) ln ( Φ ( 1 ) /Φ ( ϵ 1 )) for Φ \Phi Φ on R rk V \mathbb{R}^{\operatorname{rk}V} R rk V , so for C p C_p C p it is ln rk V p \frac{\ln\operatorname{rk}V}{p} p l n rk V (and ln n p \frac{\ln n}{p} p l n n for positive n × n n\times n n × n matrices), and OLC is the only defect-free one (Theorem 6.6.1 ). Both basic constants, the non-commutativity threshold and the defect, are thus governed by the rank.
Is there kinematics behind the contrast? On the Lorentz cone OMC is the speed of a four-velocity and OLC its rapidity, the Thompson geometry of the unit hyperboloid is hyperbolic geometry, and subadditivity of OLC is the triangle inequality of H n H^n H n with equality exactly for collinear boosts in the same direction (Theorem 6.7.2 ); for 2 × 2 2\times2 2 × 2 matrices this is the spinor and vector representation of SL ( 2 , C ) \operatorname{SL}(2,\mathbb{C}) SL ( 2 , C ) (Remark 6.7.3 ), and it sharpens a remark of [2 ] . In every non-trivial finite-dimensional representation of SL ( 2 , C ) \operatorname{SL}(2,\mathbb{C}) SL ( 2 , C ) , with an admissible inner product, equality for two pure boosts forces equal axes (Theorem 6.9.2 ); the non-commuting 4 × 4 4\times4 4 × 4 equality case of [2 ] therefore has no interpretation as a pair of Lorentz boosts, not even for Dirac spinors, and is realised instead by orthogonal blocks, i.e. direct sums (the example after Corollary 6.9.3). The rotation k k k of the polar decomposition P ( x 1 / 2 ) P ( y 1 / 2 ) = P ( w 1 / 2 ) k P(x^{1/2})P(y^{1/2})=P(w^{1/2})k P ( x 1/2 ) P ( y 1/2 ) = P ( w 1/2 ) k (Lemma 6.5.1 ) is the Wigner rotation, and its angle is the area of the triangle whose sides are the three contrasts in the subadditivity inequality (Remark 6.7.4 ).
What does the contrast see of the Jordan product? We ask for which given contrasts u u u , v v v of x x x , y y y the contrast of x ∘ y x \circ y x ∘ y (for matrices, of X Y + Y X XY+YX X Y + Y X ) is subadditive. Strang’s bounds [89 ] hold in every JB-algebra by the Shirshov–Cohn theorem (Lemma 6.8.1 ), and on a non-associative Euclidean Jordan algebra the region is the ellipse u 2 + u v + v 2 ≤ 1 u^2+uv+v^2 \le1 u 2 + uv + v 2 ≤ 1 exactly for spin factors and the triangle u + v ≤ 1 u+v \le1 u + v ≤ 1 otherwise, including the real, complex, quaternionic and exceptional cases (Theorem 6.8.2 ); in particular, for positive n × n n \times n n × n matrices the threshold changes between n = 2 n=2 n = 2 and n = 3 n=3 n = 3 (Corollary 6.8.3 ). The multiplicative commutator of two spinor boosts is conjugate to a rotation for all angles θ ∈ ( 0 , π ) \theta\in(0,\pi) θ ∈ ( 0 , π ) exactly when u 2 + v 2 < 1 u^2+v^2<1 u 2 + v 2 < 1 , the positivity threshold of the Jordan product (Remark 6.8.4).
Finally, traces on V V V are characterised by a contrast inequality involving the symmetries s s s of V V V (Theorem 6.11.1 ).
What was known and what is new
Theorem 6.3.3 is folklore: it is the triangle inequality for the Thompson/Hilbert metric on a symmetric cone. Theorem 6.3.4 is close to, but not the same as, the description of unique Thompson geodesics and of midpoint sets on symmetric cones by Lemmens and Roelands [90 , 91 ] ; we have not found it stated. We have not found the rank threshold (Theorem 6.4.1 ) in the literature; it is a short consequence of Theorem 6.3.4 . The identification of rapidity with hyperbolic distance, and of metrics on boosts and on qubit density matrices with the rapidity metric, is classical [92 , 93 , 94 ] ; Theorem 6.7.2 records it in terms of the contrast. Theorem 6.9.2 is a direct consequence of the KAK decomposition (Lemma 6.9.1) and of the 2 × 2 2 \times2 2 × 2 case; we have not found it stated, but it is elementary. The spectral identity in Lemma 6.5.1 is stated in [95 Lemma 3.1] ; it follows at once from the polar decomposition of G ( Ω ) G(\Omega) G ( Ω ) [96 Thm. III.5.1] , [97 §2] . In Theorem 6.6.1 , (a) is proved via the min–max principle, and (b)–(d) are elementary statements about eigenvalue vectors in R rk V \mathbb{R}^{\operatorname{rk} V} R rk V ; its key inequality and the norm inequality underlying the Schatten case of (a) are in [95 ] . The Wigner angle as the area of the velocity triangle is classical [92 ] ; Remark 6.7.4 only identifies the triangle. Theorem 6.10.2 combines the proof of Theorem 6.4.1 with the structure theory of JBW-algebras [98 ] . In Section 6.8, Lemma 6.8.1 is Strang’s theorem [89 , 99 ] , transferred by the Shirshov–Cohn theorem, the positivity criterion is Gustafson’s [100 ] , and the additive commutator bound and the trace formula of Remark 6.8.4 are classical [89 , 101 ] . Theorem 6.11.1 is a Jordan analogue of characterisations of traces on C ∗ C^* C ∗ -algebras by inequalities [27 , 46 ] . All proofs are short and use only the Peirce calculus and standard structure theory; the finite-dimensional statements were also checked numerically on Herm n ( R ) \operatorname{Herm}_n(\mathbb{R}) Herm n ( R ) , Herm n ( C ) \operatorname{Herm}_n(\mathbb{C}) Herm n ( C ) and spin factors. Apart from standard facts and the classical results cited in the proofs, the chapter relies only on [1 , 2 ] .
Setting
Let V V V be a Euclidean Jordan algebra with unit e e e , rank r r r , trace tr \operatorname{tr} tr and inner product ⟨ x , y ⟩ = tr ( x ∘ y ) \langle x,y\rangle=\operatorname{tr}(x \circ y) ⟨ x , y ⟩ = tr ( x ∘ y ) , and let Ω \Omega Ω be the interior of the cone of squares [96 ] . Every x ∈ V x \in V x ∈ V has a spectral decomposition x = ∑ i = 1 r λ i c i x=\sum_{i=1}^{r}\lambda_i c_i x = ∑ i = 1 r λ i c i with respect to a Jordan frame ( c i ) (c_i) ( c i ) ; we write λ max ( x ) \lambda_{\max}(x) λ m a x ( x ) , λ min ( x ) \lambda_{\min}(x) λ m i n ( x ) for the extreme eigenvalues and E max ( x ) E_{\max}(x) E m a x ( x ) , E min ( x ) E_{\min}(x) E m i n ( x ) for the corresponding spectral idempotents (sums of the c i c_i c i with λ i = λ max \lambda_i=\lambda_{\max} λ i = λ m a x , resp. λ min \lambda_{\min} λ m i n ). The quadratic representation is P ( x ) = 2 L ( x ) 2 − L ( x 2 ) P(x)=2L(x)^2-L(x^2) P ( x ) = 2 L ( x ) 2 − L ( x 2 ) . We use the following standard facts (references are given item by item): (F1) for invertible a a a , P ( a ) P(a) P ( a ) is invertible, self-adjoint, P ( a ) Ω = Ω P(a)\Omega=\Omega P ( a ) Ω = Ω , P ( a ) e = a 2 P(a)e=a^2 P ( a ) e = a 2 , P ( a ) − 1 = P ( a − 1 ) P(a)^{-1}=P(a^{-1}) P ( a ) − 1 = P ( a − 1 ) , and ( P ( a ) y ) − 1 = P ( a − 1 ) y − 1 (P(a)y)^{-1}=P(a^{-1})y^{-1} ( P ( a ) y ) − 1 = P ( a − 1 ) y − 1 [96 Ch. II–III] ;
(F2) x ∈ Ω x\in\Omega x ∈ Ω iff all λ i > 0 \lambda_i>0 λ i > 0 [96 Ch. III] , and λ max ( z ) = max { ⟨ z , w ⟩ : w ≥ 0 , tr w = 1 } \lambda_{\max}(z)=\max\{\langle z,w\rangle:w\geq0,\operatorname{tr}w=1\} λ m a x ( z ) = max {⟨ z , w ⟩ : w ≥ 0 , tr w = 1 } , the maximum being attained at a primitive idempotent. (Proof: ⟨ λ max ( z ) e − z , w ⟩ ≥ 0 \langle\lambda_{\max}(z)e-z,w\rangle\geq0 ⟨ λ m a x ( z ) e − z , w ⟩ ≥ 0 by self-duality of the closed cone [96 Thm. III.2.1] ; equality holds at the top frame elements of z z z ; and if w = ∑ i μ i d i w=\sum_i\mu_i d_i w = ∑ i μ i d i is a spectral decomposition, then μ i ≥ 0 \mu_i\geq0 μ i ≥ 0 and ∑ i μ i = tr w = 1 \sum_i\mu_i=\operatorname{tr}w=1 ∑ i μ i = tr w = 1 , so ⟨ z , w ⟩ ≤ max i ⟨ z , d i ⟩ \langle z,w\rangle\leq\max_i\langle z,d_i\rangle ⟨ z , w ⟩ ≤ max i ⟨ z , d i ⟩ .)
(F3) for u , w ≥ 0 u,w\geq0 u , w ≥ 0 , ⟨ u , w ⟩ = 0 \langle u,w\rangle=0 ⟨ u , w ⟩ = 0 implies u ∘ w = 0 u\circ w=0 u ∘ w = 0 [102 Prop. 6] ;
(F4) Peirce decomposition V = V ( c , 1 ) ⊕ V ( c , 1 2 ) ⊕ V ( c , 0 ) V=V(c,1)\oplus V(c,\frac{1}{2})\oplus V(c,0) V = V ( c , 1 ) ⊕ V ( c , 2 1 ) ⊕ V ( c , 0 ) for an idempotent c c c , with V ( c , 1 ) ∘ V ( c , 0 ) = 0 V(c,1)\circ V(c,0)=0 V ( c , 1 ) ∘ V ( c , 0 ) = 0 ; for orthogonal idempotents f , g f,g f , g , V ( f , 1 ) ∘ V ( g , 1 ) = 0 V(f,1)\circ V(g,1)=0 V ( f , 1 ) ∘ V ( g , 1 ) = 0 ;
(F5) two elements x , y x,y x , y operator commute, i.e. [ L ( x ) , L ( y ) ] = 0 [L(x),L(y)]=0 [ L ( x ) , L ( y )] = 0 , if and only if they have a common Jordan frame [96 Lemma X.2.2] ;
(F6) if V V V is simple of rank ≥ 2 \geq2 ≥ 2 and c i ⊥ c j c_i\perp c_j c i ⊥ c j are primitive, then V i j = V ( c i , 1 2 ) ∩ V ( c j , 1 2 ) ≠ 0 V_{ij}=V(c_i,\frac{1}{2})\cap V(c_j,\frac{1}{2})\ne0 V ij = V ( c i , 2 1 ) ∩ V ( c j , 2 1 ) = 0 , and there is u ∈ V i j u\in V_{ij} u ∈ V ij with u 2 = c i + c j u^2=c_i+c_j u 2 = c i + c j ; for such u u u , P ( u ) c i = c j P(u)c_i=c_j P ( u ) c i = c j . (For u ∈ V i j u\in V_{ij} u ∈ V ij one has u 2 = 1 2 ∣ u ∣ 2 ( c i + c j ) u^2=\frac{1}{2}|u|^2(c_i+c_j) u 2 = 2 1 ∣ u ∣ 2 ( c i + c j ) [96 Prop. IV.1.4] , and then P ( u ) c i = 2 u ∘ ( u ∘ c i ) − u 2 ∘ c i = u 2 − c i = c j P(u)c_i=2u\circ(u\circ c_i)-u^2\circ c_i=u^2-c_i=c_j P ( u ) c i = 2 u ∘ ( u ∘ c i ) − u 2 ∘ c i = u 2 − c i = c j .)
Two consequences of (F4)–(F5) are used below. (F7) An element y y y operator commutes with an idempotent c c c iff y ∈ V ( c , 1 ) ⊕ V ( c , 0 ) y\in V(c,1)\oplus V(c,0) y ∈ V ( c , 1 ) ⊕ V ( c , 0 ) . Indeed, if y = y 1 + y 0 y=y_1+y_0 y = y 1 + y 0 with y ν ∈ V ( c , ν ) y_\nu\in V(c,\nu) y ν ∈ V ( c , ν ) , the union of Jordan frames of y 1 y_1 y 1 in V ( c , 1 ) V(c,1) V ( c , 1 ) and of y 0 y_0 y 0 in V ( c , 0 ) V(c,0) V ( c , 0 ) is a common frame; conversely, in a common frame c c c is a sum of frame elements, and each frame element lies in V ( c , 1 ) V(c,1) V ( c , 1 ) or in V ( c , 0 ) V(c,0) V ( c , 0 ) . (F8) If x x x and y y y operator commute, then y y y operator commutes with every spectral idempotent of x x x , since each is a sum of elements of the common frame. For a primitive idempotent c c c and an idempotent E E E we write c ≤ E c\leq E c ≤ E if c ∈ V ( E , 1 ) c\in V(E,1) c ∈ V ( E , 1 ) , i.e. E ∘ c = c E\circ c=c E ∘ c = c .
Definition 6.2.1. For x ∈ Ω x\in\Omega x ∈ Ω put κ ( x ) = λ max ( x ) / λ min ( x ) \kappa(x)=\lambda_{\max}(x)/\lambda_{\min}(x) κ ( x ) = λ m a x ( x ) / λ m i n ( x ) , OLC ( x ) = 1 2 ln κ ( x ) \operatorname{OLC}(x)=\frac{1}{2}\ln\kappa(x) OLC ( x ) = 2 1 ln κ ( x ) and OMC ( x ) = tanh OLC ( x ) = κ − 1 κ + 1 \operatorname{OMC}(x)=\tanh\operatorname{OLC}(x)=\frac{\kappa-1}{\kappa+1} OMC ( x ) = tanh OLC ( x ) = κ + 1 κ − 1 .
As for matrices [2 ] : OLC ( λ x ) = OLC ( x ) \operatorname{OLC}(\lambda x)=\operatorname{OLC}(x) OLC ( λ x ) = OLC ( x ) , OLC ( x − 1 ) = OLC ( x ) \operatorname{OLC}(x^{-1})=\operatorname{OLC}(x) OLC ( x − 1 ) = OLC ( x ) , OLC ( x t ) = ∣ t ∣ OLC ( x ) \operatorname{OLC}(x^t)=|t|\operatorname{OLC}(x) OLC ( x t ) = ∣ t ∣ OLC ( x ) , OLC ( x ) = 0 \operatorname{OLC}(x)=0 OLC ( x ) = 0 iff x ∈ R + e x\in\mathbb{R}_+e x ∈ R + e , and OLC ( x ) = δ ( [ e ] , [ x ] ) \operatorname{OLC}(x)=\delta([e],[x]) OLC ( x ) = δ ([ e ] , [ x ]) for the projectivised Thompson metric δ ( [ x ] , [ y ] ) = inf λ > 0 d T ( x , λ y ) = inf λ > 0 ∥ log P ( x − 1 / 2 ) ( λ y ) ∥ ∞ \delta([x],[y])=\inf_{\lambda>0}d_T(x,\lambda y)=\inf_{\lambda>0}\lVert\log P(x^{-1/2})(\lambda y)\rVert_\infty δ ([ x ] , [ y ]) = inf λ > 0 d T ( x , λ y ) = inf λ > 0 ∥ log P ( x − 1/2 ) ( λ y ) ∥ ∞ , which is one half of Hilbert’s projective metric d H d_H d H on Ω \Omega Ω (the proof of Proposition 1.2.1 applies verbatim). For V = Herm n ( C ) V=\operatorname{Herm}_n(\mathbb{C}) V = Herm n ( C ) , P ( x 1 / 2 ) y = X 1 / 2 Y X 1 / 2 P(x^{1/2})y=X^{1/2}YX^{1/2} P ( x 1/2 ) y = X 1/2 Y X 1/2 .
Subadditivity and the equality case
Lemma 6.3.1. Let z ∈ V z\in V z ∈ V , E = E max ( z ) E=E_{\max}(z) E = E m a x ( z ) , and w ≥ 0 w\geq0 w ≥ 0 with tr w = 1 \operatorname{tr}w=1 tr w = 1 . Then ⟨ z , w ⟩ = λ max ( z ) \langle z,w\rangle=\lambda_{\max}(z) ⟨ z , w ⟩ = λ m a x ( z ) iff w ∈ V ( E , 1 ) w\in V(E,1) w ∈ V ( E , 1 ) . In particular, for a primitive idempotent c c c : ⟨ z , c ⟩ = λ max ( z ) \langle z,c\rangle=\lambda_{\max}(z) ⟨ z , c ⟩ = λ m a x ( z ) iff c ≤ E c\leq E c ≤ E , and then z ∘ c = λ max ( z ) c z\circ c=\lambda_{\max}(z)c z ∘ c = λ m a x ( z ) c .
Proof. u = λ max ( z ) e − z ≥ 0 u=\lambda_{\max}(z)e-z\geq0 u = λ m a x ( z ) e − z ≥ 0 and ⟨ u , w ⟩ = λ max tr w − ⟨ z , w ⟩ \langle u,w\rangle=\lambda_{\max}\operatorname{tr}w-\langle z,w\rangle ⟨ u , w ⟩ = λ m a x tr w − ⟨ z , w ⟩ . If this vanishes, u ∘ w = 0 u\circ w=0 u ∘ w = 0 by (F3). Write u = ∑ i ν i c i u=\sum_i\nu_i c_i u = ∑ i ν i c i with ( c i ) (c_i) ( c i ) a frame of z z z and ν i = λ max − λ i ≥ 0 \nu_i=\lambda_{\max}-\lambda_i\geq0 ν i = λ m a x − λ i ≥ 0 . In the Peirce decomposition V = ⨁ i ≤ j V i j V=\bigoplus_{i\leq j}V_{ij} V = ⨁ i ≤ j V ij of this frame, L ( u ) L(u) L ( u ) acts on V i j V_{ij} V ij as multiplication by 1 2 ( ν i + ν j ) \frac{1}{2}(\nu_i+\nu_j) 2 1 ( ν i + ν j ) [96 Ch. IV] , so ker L ( u ) = ⨁ { V i j : ν i = ν j = 0 } = V ( E , 1 ) \ker L(u)=\bigoplus\{V_{ij}:\nu_i=\nu_j=0\}=V(E,1) ker L ( u ) = ⨁ { V ij : ν i = ν j = 0 } = V ( E , 1 ) , and w ∈ V ( E , 1 ) w\in V(E,1) w ∈ V ( E , 1 ) . Conversely, if w ∈ V ( E , 1 ) w\in V(E,1) w ∈ V ( E , 1 ) then u ∘ w = 0 u\circ w=0 u ∘ w = 0 , so ⟨ u , w ⟩ = tr ( u ∘ w ) = 0 \langle u,w\rangle=\operatorname{tr}(u\circ w)=0 ⟨ u , w ⟩ = tr ( u ∘ w ) = 0 . If c ≤ E c\leq E c ≤ E , then z = λ max E + z ′ z=\lambda_{\max}E+z' z = λ m a x E + z ′ with z ′ ∈ V ( E , 0 ) z'\in V(E,0) z ′ ∈ V ( E , 0 ) , so z ∘ c = λ max E ∘ c + z ′ ∘ c = λ max c z\circ c=\lambda_{\max}E\circ c+z'\circ c=\lambda_{\max}c z ∘ c = λ m a x E ∘ c + z ′ ∘ c = λ m a x c by (F4).
Lemma 6.3.2. Let x ∈ Ω x\in\Omega x ∈ Ω and c ≤ E max ( x ) c\leq E_{\max}(x) c ≤ E m a x ( x ) primitive. Then P ( x 1 / 2 ) c = λ max ( x ) c P(x^{1/2})c=\lambda_{\max}(x)c P ( x 1/2 ) c = λ m a x ( x ) c .
Proof. Write x 1 / 2 = α E + r x^{1/2}=\alpha E+r x 1/2 = α E + r with α = λ max ( x ) 1 / 2 \alpha=\lambda_{\max}(x)^{1/2} α = λ m a x ( x ) 1/2 , E = E max ( x ) E=E_{\max}(x) E = E m a x ( x ) , r ∈ V ( E , 0 ) r\in V(E,0) r ∈ V ( E , 0 ) . Since c ∈ V ( E , 1 ) c\in V(E,1) c ∈ V ( E , 1 ) , (F4) gives L ( r ) c = 0 L(r)c=0 L ( r ) c = 0 , L ( r 2 ) c = 0 L(r^2)c=0 L ( r 2 ) c = 0 , L ( E ) c = c L(E)c=c L ( E ) c = c ; hence P ( r ) c = 0 P(r)c=0 P ( r ) c = 0 and the mixed term L ( E ) L ( r ) c + L ( r ) L ( E ) c − L ( E ∘ r ) c = 0 L(E)L(r)c+L(r)L(E)c-L(E\circ r)c=0 L ( E ) L ( r ) c + L ( r ) L ( E ) c − L ( E ∘ r ) c = 0 , so P ( x 1 / 2 ) c = α 2 P ( E ) c = α 2 c P(x^{1/2})c=\alpha^2P(E)c=\alpha^2c P ( x 1/2 ) c = α 2 P ( E ) c = α 2 c .
Theorem 6.3.3. For x , y ∈ Ω x,y \in\Omega x , y ∈ Ω ,
λ max ( P ( x 1 / 2 ) y ) ≤ λ max ( x ) λ max ( y ) , λ min ( P ( x 1 / 2 ) y ) ≥ λ min ( x ) λ min ( y ) , \lambda_{\max}(P(x^{1/2})y) \le\lambda_{\max}(x)\lambda_{\max}(y), \qquad\lambda_{\min}(P(x^{1/2})y) \ge\lambda_{\min}(x)\lambda_{\min}(y), λ m a x ( P ( x 1/2 ) y ) ≤ λ m a x ( x ) λ m a x ( y ) , λ m i n ( P ( x 1/2 ) y ) ≥ λ m i n ( x ) λ m i n ( y ) ,
hence OLC ( P ( x 1 / 2 ) y ) ≤ OLC ( x ) + OLC ( y ) \operatorname{OLC}(P(x^{1/2})y) \le\operatorname{OLC}(x) + \operatorname{OLC}(y) OLC ( P ( x 1/2 ) y ) ≤ OLC ( x ) + OLC ( y ) and OMC ( P ( x 1 / 2 ) y ) ≤ OMC ( x ) + OMC ( y ) 1 + OMC ( x ) OMC ( y ) \operatorname{OMC}(P(x^{1/2})y) \le\frac{\operatorname{OMC}(x)+\operatorname{OMC}(y)}{1+\operatorname{OMC}(x)\operatorname{OMC}(y)} OMC ( P ( x 1/2 ) y ) ≤ 1 + OMC ( x ) OMC ( y ) OMC ( x ) + OMC ( y ) .
Proof. λ min ( y ) e ≤ y ≤ λ max ( y ) e \lambda_{\min}(y)e \le y \le\lambda_{\max}(y)e λ m i n ( y ) e ≤ y ≤ λ m a x ( y ) e and P ( x 1 / 2 ) P(x^{1/2}) P ( x 1/2 ) preserves the order (F1), so λ min ( y ) x ≤ P ( x 1 / 2 ) y ≤ λ max ( y ) x \lambda_{\min}(y)x \le P(x^{1/2})y \le\lambda_{\max}(y)x λ m i n ( y ) x ≤ P ( x 1/2 ) y ≤ λ m a x ( y ) x , and λ min ( x ) e ≤ x ≤ λ max ( x ) e \lambda_{\min}(x)e \le x \le\lambda_{\max}(x)e λ m i n ( x ) e ≤ x ≤ λ m a x ( x ) e . Equivalently: OLC is the distance from [ e ] [e] [ e ] in δ T \delta_T δ T , P ( x − 1 / 2 ) P(x^{-1/2}) P ( x − 1/2 ) is a δ T \delta_T δ T -isometry mapping [ x ] [x] [ x ] to [ e ] [e] [ e ] and [ P ( x 1 / 2 ) y ] [P(x^{1/2})y] [ P ( x 1/2 ) y ] to [ y ] [y] [ y ] , and the inequality is the triangle inequality.
Theorem 6.3.4 (Equality). For x , y ∈ Ω x,y \in\Omega x , y ∈ Ω , OLC ( P ( x 1 / 2 ) y ) = OLC ( x ) + OLC ( y ) \operatorname{OLC}(P(x^{1/2})y) = \operatorname{OLC}(x) + \operatorname{OLC}(y) OLC ( P ( x 1/2 ) y ) = OLC ( x ) + OLC ( y ) if and only if there are primitive idempotents c , c ′ c,c' c , c ′ with
c ≤ E max ( x ) , c ≤ E max ( y ) , c ′ ≤ E min ( x ) , c ′ ≤ E min ( y ) c \le E_{\max}(x), \qquad c \le E_{\max}(y), \qquad c' \le E_{\min}(x), \qquad c' \le E_{\min}(y) c ≤ E m a x ( x ) , c ≤ E m a x ( y ) , c ′ ≤ E m i n ( x ) , c ′ ≤ E m i n ( y ) .
Proof. By Theorem 6.3.3 , equality for OLC is equivalent to equality in both eigenvalue bounds. Top. By (F2) and self-adjointness, λ max ( P ( x 1 / 2 ) y ) = max c ⟨ y , P ( x 1 / 2 ) c ⟩ \lambda_{\max}(P(x^{1/2})y) = \max_c \langle y,P(x^{1/2})c\rangle λ m a x ( P ( x 1/2 ) y ) = max c ⟨ y , P ( x 1/2 ) c ⟩ over primitive c c c . For each state w w w , w ′ = P ( x 1 / 2 ) w ≥ 0 w' = P(x^{1/2})w \ge0 w ′ = P ( x 1/2 ) w ≥ 0 has tr w ′ = ⟨ x , w ⟩ ≤ λ max ( x ) \operatorname{tr} w' = \langle x,w\rangle\le\lambda_{\max}(x) tr w ′ = ⟨ x , w ⟩ ≤ λ m a x ( x ) and ⟨ y , w ′ ⟩ ≤ λ max ( y ) tr w \langle y,w'\rangle\le\lambda_{\max}(y)\operatorname{tr} w ⟨ y , w ′ ⟩ ≤ λ m a x ( y ) tr w . So equality forces a primitive c c c with ⟨ x , c ⟩ = λ max ( x ) \langle x,c\rangle= \lambda_{\max}(x) ⟨ x , c ⟩ = λ m a x ( x ) , i.e. c ≤ E max ( x ) c \le E_{\max}(x) c ≤ E m a x ( x ) (Lemma 6.3.1 ), and then w ′ = λ max ( x ) c w' = \lambda_{\max}(x)c w ′ = λ m a x ( x ) c (Lemma 6.3.2 ) and ⟨ y , c ⟩ = λ max ( y ) \langle y,c\rangle= \lambda_{\max}(y) ⟨ y , c ⟩ = λ m a x ( y ) , i.e. c ≤ E max ( y ) c \le E_{\max}(y) c ≤ E m a x ( y ) . Conversely, for such c c c , λ max ( P ( x 1 / 2 ) y ) ≥ ⟨ y , P ( x 1 / 2 ) c ⟩ = λ max ( x ) λ max ( y ) \lambda_{\max}(P(x^{1/2})y) \ge\langle y,P(x^{1/2})c\rangle= \lambda_{\max}(x)\lambda_{\max}(y) λ m a x ( P ( x 1/2 ) y ) ≥ ⟨ y , P ( x 1/2 ) c ⟩ = λ m a x ( x ) λ m a x ( y ) . Bottom. By (F1), λ min ( P ( x 1 / 2 ) y ) − 1 = λ max ( P ( x − 1 / 2 ) y − 1 ) \lambda_{\min}(P(x^{1/2})y)^{-1} = \lambda_{\max}(P(x^{-1/2})y^{-1}) λ m i n ( P ( x 1/2 ) y ) − 1 = λ m a x ( P ( x − 1/2 ) y − 1 ) and E max ( x − 1 ) = E min ( x ) E_{\max}(x^{-1}) = E_{\min}(x) E m a x ( x − 1 ) = E m i n ( x ) ; apply the top case to x − 1 , y − 1 x^{-1},y^{-1} x − 1 , y − 1 .
Remark 6.3.5. For V = Herm n ( C ) V = \operatorname{Herm}_n(\mathbb{C}) V = Herm n ( C ) the same condition (common eigenvectors in E max ( X ) ∩ E max ( Y ) E_{\max}(X) \cap E_{\max}(Y) E m a x ( X ) ∩ E m a x ( Y ) and in E min ( X ) ∩ E min ( Y ) E_{\min}(X) \cap E_{\min}(Y) E m i n ( X ) ∩ E m i n ( Y ) ) is also necessary and sufficient for κ ( X Y ) = κ ( X ) κ ( Y ) \kappa(XY) = \kappa(X)\kappa(Y) κ ( X Y ) = κ ( X ) κ ( Y ) : necessity is Lemma 1.3.2 , and sufficiency follows by evaluating X Y XY X Y and ( X Y ) − 1 (XY)^{-1} ( X Y ) − 1 on the common eigenvectors. So Lemma 1.3.2 is in fact an equivalence, and the products X Y XY X Y and X 1 / 2 Y X 1 / 2 X^{1/2}YX^{1/2} X 1/2 Y X 1/2 have the same equality case.
The rank threshold
Theorem 6.4.1. For a Euclidean Jordan algebra V V V the following are equivalent:
(a) for all x , y ∈ Ω x,y \in\Omega x , y ∈ Ω , OLC ( P ( x 1 / 2 ) y ) = OLC ( x ) + OLC ( y ) \operatorname{OLC}(P(x^{1/2})y) = \operatorname{OLC}(x) + \operatorname{OLC}(y) OLC ( P ( x 1/2 ) y ) = OLC ( x ) + OLC ( y ) implies that x x x and y y y operator commute;
(b) rk V ≤ 3 \operatorname{rk} V \le3 rk V ≤ 3 , or V V V is associative (V ≅ R r V \cong\mathbb{R}^r V ≅ R r ).
Proof. If V V V is associative, all elements operator commute. Let rk V ≤ 3 \operatorname{rk} V \le3 rk V ≤ 3 , assume equality, and take c , c ′ c,c' c , c ′ as in Theorem 6.3.4 ; we may assume x ∉ R e x \notin\mathbb{R}e x ∈ / R e . By Lemma 6.3.1 , y ∘ c = λ max ( y ) c y \circ c = \lambda_{\max}(y)c y ∘ c = λ m a x ( y ) c and y ∘ c ′ = λ min ( y ) c ′ y \circ c' = \lambda_{\min}(y)c' y ∘ c ′ = λ m i n ( y ) c ′ , so (by the Peirce decomposition) y y y has no components in V ( c , 1 2 ) V(c,\frac{1}{2}) V ( c , 2 1 ) and V ( c ′ , 1 2 ) V(c',\frac{1}{2}) V ( c ′ , 2 1 ) . As rk E max ( x ) + rk E min ( x ) ≤ 3 \operatorname{rk} E_{\max}(x) + \operatorname{rk} E_{\min}(x) \le3 rk E m a x ( x ) + rk E m i n ( x ) ≤ 3 , at least one of them is primitive. If both are, c = E max ( x ) c = E_{\max}(x) c = E m a x ( x ) , c ′ = E min ( x ) c' = E_{\min}(x) c ′ = E m i n ( x ) , and g = e − c − c ′ g = e - c - c' g = e − c − c ′ is 0 0 0 or primitive; then V ( c , 0 ) = V ( c ′ + g , 1 ) = R c ′ ⊕ V c ′ g ⊕ R g V(c,0) = V(c' + g,1) = \mathbb{R}c' \oplus V_{c'g} \oplus\mathbb{R}g V ( c , 0 ) = V ( c ′ + g , 1 ) = R c ′ ⊕ V c ′ g ⊕ R g (with V c ′ g = 0 V_{c'g} = 0 V c ′ g = 0 if g = 0 g = 0 g = 0 ), the component in V c ′ g ⊂ V ( c ′ , 1 2 ) V_{c'g} \subset V(c',\frac{1}{2}) V c ′ g ⊂ V ( c ′ , 2 1 ) is excluded, and so y ∈ V ( c , 1 ) ⊕ V ( c ′ , 1 ) ⊕ V ( g , 1 ) = R c ⊕ R c ′ ⊕ R g y \in V(c,1) \oplus V(c',1) \oplus V(g,1) = \mathbb{R}c \oplus\mathbb{R}c' \oplus\mathbb{R}g y ∈ V ( c , 1 ) ⊕ V ( c ′ , 1 ) ⊕ V ( g , 1 ) = R c ⊕ R c ′ ⊕ R g , which lies in the associative span of the spectral idempotents of x x x . If, say, E max ( x ) = c E_{\max}(x) = c E m a x ( x ) = c and E min ( x ) = e − c E_{\min}(x) = e - c E m i n ( x ) = e − c has rank 2 2 2 , then x = λ max c + λ min ( e − c ) x = \lambda_{\max}c + \lambda_{\min}(e-c) x = λ m a x c + λ m i n ( e − c ) and y = λ max ( y ) c + y 0 y = \lambda_{\max}(y)c + y_0 y = λ m a x ( y ) c + y 0 with y 0 ∈ V ( c , 0 ) y_0 \in V(c,0) y 0 ∈ V ( c , 0 ) ; a Jordan frame ( d 1 , d 2 ) (d_1,d_2) ( d 1 , d 2 ) of y 0 y_0 y 0 in V ( c , 0 ) V(c,0) V ( c , 0 ) gives a common frame ( c , d 1 , d 2 ) (c,d_1,d_2) ( c , d 1 , d 2 ) of x x x and y y y . The case E min ( x ) E_{\min}(x) E m i n ( x ) primitive is symmetric. By (F5), x x x and y y y operator commute.
Now let V V V be non-associative with rk V ≥ 4 \operatorname{rk} V \ge4 rk V ≥ 4 . Some simple component W W W has rank ≥ 2 \ge2 ≥ 2 . Choose a Jordan frame of V V V containing primitive c 2 , c 3 ∈ W c_2,c_3 \in W c 2 , c 3 ∈ W and two further elements c 1 , c 4 c_1,c_4 c 1 , c 4 , and let c 5 , … , c r c_5,\ldots,c_r c 5 , … , c r be the rest. By (F6) pick u ∈ V 23 u \in V_{23} u ∈ V 23 with u 2 = c 2 + c 3 u^2=c_2+c_3 u 2 = c 2 + c 3 . Put
x = 4 c 1 + 3 c 2 + c 3 + c 4 + 2 ∑ k ≥ 5 c k , y = 4 c 1 + 2 ( c 2 + c 3 ) + u + c 4 + 2 ∑ k ≥ 5 c k . x=4c_1+3c_2+c_3+c_4+2\sum_{k\ge5}c_k,\qquad y=4c_1+2(c_2+c_3)+u+c_4+2\sum_{k\ge5}c_k. x = 4 c 1 + 3 c 2 + c 3 + c 4 + 2 k ≥ 5 ∑ c k , y = 4 c 1 + 2 ( c 2 + c 3 ) + u + c 4 + 2 k ≥ 5 ∑ c k .
Here y ′ = 2 ( c 2 + c 3 ) + u y'=2(c_2+c_3)+u y ′ = 2 ( c 2 + c 3 ) + u lies in the rank-two subalgebra V ( c 2 + c 3 , 1 ) V(c_2+c_3,1) V ( c 2 + c 3 , 1 ) , where u u u has eigenvalues ± 1 \pm1 ± 1 , so y ′ y' y ′ has eigenvalues 3 , 1 3,1 3 , 1 and y ∈ Ω y\in\Omega y ∈ Ω . All summands of x 1 / 2 x^{1/2} x 1/2 and y y y lie in Peirce 1-spaces of the orthogonal idempotents c 1 c_1 c 1 , c 2 + c 3 c_2+c_3 c 2 + c 3 , c 4 c_4 c 4 , c k c_k c k , which multiply to zero (F4); hence P ( x 1 / 2 ) y = 16 c 1 + P ( x ′ 1 / 2 ) y ′ + c 4 + 4 ∑ k ≥ 5 c k P(x^{1/2})y=16c_1+P(x'^{1/2})y'+c_4+4\sum_{k\ge5}c_k P ( x 1/2 ) y = 16 c 1 + P ( x ′ 1/2 ) y ′ + c 4 + 4 ∑ k ≥ 5 c k with x ′ = 3 c 2 + c 3 x'=3c_2+c_3 x ′ = 3 c 2 + c 3 , and by Theorem 6.3.3 the spectrum of P ( x ′ 1 / 2 ) y ′ P(x'^{1/2})y' P ( x ′ 1/2 ) y ′ lies in [ 1 , 9 ] [1,9] [ 1 , 9 ] . Thus λ max = 16 = 4 ⋅ 4 \lambda_{\max}=16=4\cdot4 λ m a x = 16 = 4 ⋅ 4 and λ min = 1 = 1 ⋅ 1 \lambda_{\min}=1=1\cdot1 λ m i n = 1 = 1 ⋅ 1 : equality holds. But c 2 c_2 c 2 is the spectral idempotent of x x x for the simple eigenvalue 3 3 3 , and y y y has the nonzero component u ∈ V ( c 2 , 2 ˉ ) u\in V(c_2,\bar{2}) u ∈ V ( c 2 , 2 ˉ ) , so y y y does not operator commute with c 2 c_2 c 2 by (F7), hence not with x x x by (F8).
Corollary 6.4.2. (a) V = Herm n ( F ) V=\operatorname{Herm}_n(\mathbb{F}) V = Herm n ( F ) , F = R , C , H \mathbb{F}=\mathbb{R},\mathbb{C},\mathbb{H} F = R , C , H : equality forces commutativity iff n ≤ 3 n\le3 n ≤ 3 . For F = C \mathbb{F}=\mathbb{C} F = C this is Theorem 1.3.3 , by the Remark after Theorem 6.3.4 .
(b) The Albert algebra Herm 3 ( O ) \operatorname{Herm}_3(\mathbb{O}) Herm 3 ( O ) (dim 27 27 27 , rank 3 3 3 ): equality always forces operator commutativity.
(c) Spin factors V n = R × R n V_n=\mathbb{R}\times\mathbb{R}^n V n = R × R n (Lorentz cones, any n n n , rank 2 2 2 ): equality forces operator commutativity; a direct sum V m ⊕ V n V_m\oplus V_n V m ⊕ V n (rank 4 4 4 ) does not when max ( m , n ) ≥ 2 \max(m,n)\ge2 max ( m , n ) ≥ 2 . (V 1 ≅ R 2 V_1\cong\mathbb{R}^2 V 1 ≅ R 2 is associative, so V 1 ⊕ V 1 ≅ R 4 V_1\oplus V_1\cong\mathbb{R}^4 V 1 ⊕ V 1 ≅ R 4 is an exception.)
So the threshold of [2 ] is a threshold in the rank, not in the dimension; this complements Proposition 1.3.4 , where it is shown not to be one of the total dimension for pure tensor products. The obstruction in the construction is a rank-two Peirce subalgebra V ( c 2 + c 3 , 1 ) V(c_2+c_3,1) V ( c 2 + c 3 , 1 ) orthogonal to idempotents c ≤ E max c\le E_{\max} c ≤ E m a x and c ′ ≤ E min c'\le E_{\min} c ′ ≤ E m i n ; fitting these three pieces needs rank ≥ 4 \ge4 ≥ 4 . (The algebra itself may be simple, e.g. Herm 4 ( R ) \operatorname{Herm}_4(\mathbb{R}) Herm 4 ( R ) .)
Polar decomposition and reciprocity
Let G ( Ω ) = { g ∈ GL ( V ) : g Ω = Ω } G(\Omega)=\{g\in\operatorname{GL}(V):g\Omega=\Omega\} G ( Ω ) = { g ∈ GL ( V ) : g Ω = Ω } and let Aut ( V ) \operatorname{Aut}(V) Aut ( V ) be the automorphism group of the Jordan algebra. Recall that P ( a ) ∈ G ( Ω ) P(a)\in G(\Omega) P ( a ) ∈ G ( Ω ) and P ( a ) ∗ = P ( a ) P(a)^*=P(a) P ( a ) ∗ = P ( a ) for a ∈ Ω a\in\Omega a ∈ Ω (F1), and that the stabiliser of e e e in G ( Ω ) G(\Omega) G ( Ω ) is Aut ( V ) \operatorname{Aut}(V) Aut ( V ) [96 Prop. I.1.9, Prop. I.4.3, Thm. III.5.1] . (Directly: an element k k k of the stabiliser is orthogonal and preserves the characteristic function φ \varphi φ of Ω \Omega Ω ; as ∇ log φ ( z ) \nabla\log\varphi(z) ∇ log φ ( z ) is a negative multiple of z − 1 z^{-1} z − 1 , this gives k ( z − 1 ) = ( k z ) − 1 k(z^{-1})=(kz)^{-1} k ( z − 1 ) = ( k z ) − 1 ; differentiating gives P ( k z ) = k P ( z ) k − 1 P(kz)=kP(z)k^{-1} P ( k z ) = k P ( z ) k − 1 , and expanding at z = e + t a z=e+ta z = e + t a gives L ( k a ) = k L ( a ) k − 1 L(ka)=kL(a)k^{-1} L ( k a ) = k L ( a ) k − 1 .) Automorphisms are orthogonal and preserve spectra.
Lemma 6.5.1 (Reciprocity). For x , y ∈ Ω x,y\in\Omega x , y ∈ Ω put g = P ( x 1 / 2 ) P ( y 1 / 2 ) g=P(x^{1/2})P(y^{1/2}) g = P ( x 1/2 ) P ( y 1/2 ) and w = g e = P ( x 1 / 2 ) y w=ge=P(x^{1/2})y w = g e = P ( x 1/2 ) y . Then
g = P ( w 1 / 2 ) k with k ∈ Aut ( V ) , g=P(w^{1/2})k\qquad\text{with }k\in\operatorname{Aut}(V), g = P ( w 1/2 ) k with k ∈ Aut ( V ) , and
P ( y 1 / 2 ) x = k − 1 w . P(y^{1/2})x=k^{-1}w. P ( y 1/2 ) x = k − 1 w . In particular P ( x 1 / 2 ) y P(x^{1/2})y P ( x 1/2 ) y and P ( y 1 / 2 ) x P(y^{1/2})x P ( y 1/2 ) x have the same eigenvalues, with multiplicities.
Proof. k : = P ( w − 1 / 2 ) g ∈ G ( Ω ) k:=P(w^{-1/2})g\in G(\Omega) k := P ( w − 1/2 ) g ∈ G ( Ω ) and k e = P ( w − 1 / 2 ) w = e ke=P(w^{-1/2})w=e k e = P ( w − 1/2 ) w = e , so k ∈ Aut ( V ) k\in\operatorname{Aut}(V) k ∈ Aut ( V ) and k ∗ = k − 1 k^*=k^{-1} k ∗ = k − 1 . Since each P ( ⋅ ) P(\cdot) P ( ⋅ ) is self-adjoint, g ∗ = P ( y 1 / 2 ) P ( x 1 / 2 ) g^*=P(y^{1/2})P(x^{1/2}) g ∗ = P ( y 1/2 ) P ( x 1/2 ) , and also g ∗ = k ∗ P ( w 1 / 2 ) = k − 1 P ( w 1 / 2 ) g^*=k^*P(w^{1/2})=k^{-1}P(w^{1/2}) g ∗ = k ∗ P ( w 1/2 ) = k − 1 P ( w 1/2 ) ; hence P ( y 1 / 2 ) x = g ∗ e = k − 1 w P(y^{1/2})x=g^*e=k^{-1}w P ( y 1/2 ) x = g ∗ e = k − 1 w .
The spectral identity is stated in [95 Lemma 3.1] , with the reference “([17], Corollary 9)”, where [17 ] is [97 ] . We could not locate such a statement in [97 ] (its numbered results are Propositions 2.1, 2.2, 2.4–2.7, Lemma 2.3 and Theorem 2.8, with a common counter); its §2 recalls the polar decomposition g = P ( x ) k g=P(x)k g = P ( x ) k of G ( Ω ) G(\Omega) G ( Ω ) , from which the identity follows as above. For V = Herm n ( C ) V=\operatorname{Herm}_{n}(\mathbb{C}) V = Herm n ( C ) it is the similarity of X 1 / 2 Y X 1 / 2 X^{1/2}YX^{1/2} X 1/2 Y X 1/2 and Y 1 / 2 X Y 1 / 2 Y^{1/2}XY^{1/2} Y 1/2 X Y 1/2 . What we use below is the factorisation itself: on the Lorentz cone k k k is the Wigner rotation (Remark 6.7.4 ).
Spectral contrasts: the defect is the rank
Let r = rk V ≥ 2 r=\operatorname{rk} V\ge2 r = rk V ≥ 2 , let λ ( x ) ∈ R r \lambda(x)\in\mathbb{R}^{r} λ ( x ) ∈ R r be the vector of eigenvalues of x x x (with multiplicities), and let Φ \Phi Φ be a symmetric gauge function on R r \mathbb{R}^{r} R r , i.e. a norm invariant under permutations and sign changes of coordinates. With 1 = ( 1 , … , 1 ) \mathbf{1}=(1,\ldots,1) 1 = ( 1 , … , 1 ) and ϵ 1 = ( 1 , 0 , … , 0 ) \epsilon_{1}=(1,0,\ldots,0) ϵ 1 = ( 1 , 0 , … , 0 ) put
κ Φ ( x ) = Φ ( λ ( x ) ) Φ ( λ ( x − 1 ) ) , C Φ ( x ) = 1 2 ln κ Φ ( x ) κ Φ ( e ) , d Φ = ln Φ ( 1 ) Φ ( ϵ 1 ) ( x ∈ Ω ) . \kappa_{\Phi}(x)=\Phi(\lambda(x))\Phi(\lambda(x^{-1})),\qquad C_{\Phi}(x)=\frac{1}{2}\ln\frac{\kappa_{\Phi}(x)}{\kappa_{\Phi}(e)},\qquad d_{\Phi}=\ln\frac{\Phi(\mathbf{1})}{\Phi(\epsilon_{1})}\qquad(x\in\Omega). κ Φ ( x ) = Φ ( λ ( x )) Φ ( λ ( x − 1 )) , C Φ ( x ) = 2 1 ln κ Φ ( e ) κ Φ ( x ) , d Φ = ln Φ ( ϵ 1 ) Φ ( 1 ) ( x ∈ Ω ) .
For Φ = ∥ ⋅ ∥ ∞ \Phi=\|\cdot\|_{\infty} Φ = ∥ ⋅ ∥ ∞ , C Φ = OLC C_{\Phi}=\operatorname{OLC} C Φ = OLC and d Φ = 0 d_{\Phi}=0 d Φ = 0 . For Φ = ∥ ⋅ ∥ p \Phi=\|\cdot\|_{p} Φ = ∥ ⋅ ∥ p , p ≥ 1 p\ge1 p ≥ 1 , one obtains the Schatten contrasts
C p ( x ) = 1 2 p ln ( τ ( x p ) τ ( x − p ) ) , τ = 1 r tr , d p = ln r p ; C_{p}(x)=\frac{1}{2p}\ln\bigl(\tau(x^{p})\tau(x^{-p})\bigr),\qquad\tau=\frac{1}{r}\operatorname{tr},\qquad d_{p}=\frac{\ln r}{p}; C p ( x ) = 2 p 1 ln ( τ ( x p ) τ ( x − p ) ) , τ = r 1 tr , d p = p ln r ;
we use the same formula for 0 < p < 1 0<p<1 0 < p < 1 . All these quantities depend only on λ ( x ) \lambda(x) λ ( x ) , and every vector in ( 0 , ∞ ) r (0,\infty)^{r} ( 0 , ∞ ) r is λ ( x ) \lambda(x) λ ( x ) for some x ∈ Ω x\in\Omega x ∈ Ω (take x = ∑ i λ i c i x=\sum_i\lambda_i c_i x = ∑ i λ i c i for a Jordan frame). Hence a statement about a single element of Ω \Omega Ω that depends only on its spectrum is a statement about vectors in ( 0 , ∞ ) r (0,\infty)^{r} ( 0 , ∞ ) r , the same as for positive r × r r\times r r × r matrices; the product law does not depend only on spectra.
We use three elementary properties of a symmetric gauge function Φ \Phi Φ on R r \mathbb{R}^{r} R r .
(G1) If ∣ v i ∣ ≤ ∣ w i ∣ |v_i|\le|w_i| ∣ v i ∣ ≤ ∣ w i ∣ for all i i i , then Φ ( v ) ≤ Φ ( w ) \Phi(v)\le\Phi(w) Φ ( v ) ≤ Φ ( w ) . (By sign invariance we may assume v , w ≥ 0 v,w\ge0 v , w ≥ 0 , and it suffices to change one coordinate, say the first (by permutation invariance), and then to apply this coordinate by coordinate; so let v = w v=w v = w except 0 ≤ v 1 ≤ w 1 0\le v_1\le w_1 0 ≤ v 1 ≤ w 1 , and w 1 > 0 w_1>0 w 1 > 0 (otherwise v = w v=w v = w ). With w ′ = ( − w 1 , w 2 , … , w r ) w'=(-w_1,w_2,\ldots,w_r) w ′ = ( − w 1 , w 2 , … , w r ) and t = 1 2 ( 1 + v 1 / w 1 ) ∈ [ 0 , 1 ] t=\frac{1}{2}(1+v_1/w_1)\in[0,1] t = 2 1 ( 1 + v 1 / w 1 ) ∈ [ 0 , 1 ] we have v = t w + ( 1 − t ) w ′ v=tw+(1-t)w' v = tw + ( 1 − t ) w ′ , so Φ ( v ) ≤ t Φ ( w ) + ( 1 − t ) Φ ( w ′ ) = Φ ( w ) \Phi(v)\le t\Phi(w)+(1-t)\Phi(w')=\Phi(w) Φ ( v ) ≤ t Φ ( w ) + ( 1 − t ) Φ ( w ′ ) = Φ ( w ) .)
(G2) Φ ( ϵ 1 ) ∥ v ∥ ∞ ≤ Φ ( v ) ≤ Φ ( 1 ) ∥ v ∥ ∞ \Phi(\epsilon_1)\|v\|_{\infty}\le\Phi(v)\le\Phi(\mathbf{1})\|v\|_{\infty} Φ ( ϵ 1 ) ∥ v ∥ ∞ ≤ Φ ( v ) ≤ Φ ( 1 ) ∥ v ∥ ∞ . (By (G1) and permutation invariance, since ∥ v ∥ ∞ ϵ j ≤ ∣ v ∣ ≤ ∥ v ∥ ∞ 1 \|v\|_{\infty}\epsilon_j\le|v|\le\|v\|_{\infty}\mathbf{1} ∥ v ∥ ∞ ϵ j ≤ ∣ v ∣ ≤ ∥ v ∥ ∞ 1 coordinatewise, with j j j an index where ∣ v j ∣ = ∥ v ∥ ∞ |v_j|=\|v\|_{\infty} ∣ v j ∣ = ∥ v ∥ ∞ .)
(G3) If v , w ≥ 0 v,w\ge0 v , w ≥ 0 and θ ∈ ( 0 , 1 ) \theta\in(0,1) θ ∈ ( 0 , 1 ) , then Φ ( v θ w 1 − θ ) ≤ Φ ( v ) θ Φ ( w ) 1 − θ \Phi(v^{\theta}w^{1-\theta})\le\Phi(v)^{\theta}\Phi(w)^{1-\theta} Φ ( v θ w 1 − θ ) ≤ Φ ( v ) θ Φ ( w ) 1 − θ , with coordinatewise powers and product. (If v = 0 v=0 v = 0 or w = 0 w=0 w = 0 both sides vanish; otherwise, by homogeneity, let Φ ( v ) = Φ ( w ) = 1 \Phi(v)=\Phi(w)=1 Φ ( v ) = Φ ( w ) = 1 . The weighted arithmetic–geometric mean inequality gives v θ w 1 − θ ≤ θ v + ( 1 − θ ) w v^{\theta}w^{1-\theta}\le\theta v+(1-\theta)w v θ w 1 − θ ≤ θ v + ( 1 − θ ) w coordinatewise, so by (G1) and the triangle inequality Φ ( v θ w 1 − θ ) ≤ θ + ( 1 − θ ) = 1 \Phi(v^{\theta}w^{1-\theta})\le\theta+(1-\theta)=1 Φ ( v θ w 1 − θ ) ≤ θ + ( 1 − θ ) = 1 .)
We use the min–max principle for Euclidean Jordan algebras [103 ] , in the form: if u ≤ v u\le v u ≤ v in V V V , then λ j ↓ ( u ) ≤ λ j ↓ ( v ) \lambda_j^{\downarrow}(u)\le\lambda_j^{\downarrow}(v) λ j ↓ ( u ) ≤ λ j ↓ ( v ) for all j j j , where λ ↓ \lambda^{\downarrow} λ ↓ lists the eigenvalues in nonincreasing order [95 Prop. 2.3] , [104 ] . (For a direct sum, the eigenvalue lists of the components are merged, which preserves the coordinatewise order.)
Theorem 6.6.1. Let Φ \Phi Φ be a symmetric gauge function on R r \mathbb{R}^{r} R r , r = rk V ≥ 2 r=\operatorname{rk} V\ge2 r = rk V ≥ 2 .
(a) For all x , y ∈ Ω x,y\in\Omega x , y ∈ Ω , C Φ ( P ( x 1 / 2 ) y ) ≤ C Φ ( x ) + C Φ ( y ) + d Φ C_{\Phi}(P(x^{1/2})y)\le C_{\Phi}(x)+C_{\Phi}(y)+d_{\Phi} C Φ ( P ( x 1/2 ) y ) ≤ C Φ ( x ) + C Φ ( y ) + d Φ .
(b) For x ∈ Ω x\in\Omega x ∈ Ω and s ≥ 1 s\ge1 s ≥ 1 , s C Φ ( x ) ≤ C Φ ( x s ) ≤ s C Φ ( x ) + ( s − 1 ) d Φ sC_{\Phi}(x)\le C_{\Phi}(x^{s})\le sC_{\Phi}(x)+(s-1)d_{\Phi} s C Φ ( x ) ≤ C Φ ( x s ) ≤ s C Φ ( x ) + ( s − 1 ) d Φ .
(c) The constants are optimal: for a Jordan frame ( c i ) (c_i) ( c i ) and x t = t c 1 + ∑ 1 < i < r c i + t − 1 c r x_t=tc_1+\sum_{1<i<r}c_i+t^{-1}c_r x t = t c 1 + ∑ 1 < i < r c i + t − 1 c r ,
lim t → ∞ ( C Φ ( x t s ) − s C Φ ( x t ) ) = ( s − 1 ) d Φ ; \lim_{t\to\infty}(C_{\Phi}(x_t^s)-sC_{\Phi}(x_t))=(s-1)d_{\Phi}; t → ∞ lim ( C Φ ( x t s ) − s C Φ ( x t )) = ( s − 1 ) d Φ ; in particular C Φ ( P ( x t 1 / 2 ) x t ) − 2 C Φ ( x t ) → d Φ C_{\Phi}(P(x_t^{1/2})x_t)-2C_{\Phi}(x_t)\to d_{\Phi} C Φ ( P ( x t 1/2 ) x t ) − 2 C Φ ( x t ) → d Φ .
(d) C Φ ( P ( x 1 / 2 ) y ) ≤ C Φ ( x ) + C Φ ( y ) C_{\Phi}(P(x^{1/2})y) \le C_{\Phi}(x) + C_{\Phi}(y) C Φ ( P ( x 1/2 ) y ) ≤ C Φ ( x ) + C Φ ( y ) for all x , y ∈ Ω x,y \in\Omega x , y ∈ Ω if and only if d Φ = 0 d_{\Phi}=0 d Φ = 0 , if and only if Φ \Phi Φ is a multiple of ∥ ⋅ ∥ ∞ \|\cdot\|_{\infty} ∥ ⋅ ∥ ∞ , i.e. C Φ = C O L C C_{\Phi}=C_{\mathrm{OLC}} C Φ = C OLC .
In particular, for the Schatten contrasts the sharp defect is ln r p \frac{\ln r}{p} p l n r (also for 0 < p < 1 0<p<1 0 < p < 1 ): it is ln 2 p \frac{\ln2}{p} p l n 2 on every Lorentz cone, of any dimension, ln 3 p \frac{\ln3}{p} p l n 3 on the 27-dimensional exceptional cone, and ln n p \frac{\ln n}{p} p l n n on the cone of positive n × n n\times n n × n matrices.
Proof. We may assume Φ ( 1 ) = 1 \Phi(\mathbf{1})=1 Φ ( 1 ) = 1 , so that C Φ = 1 2 ln κ Φ C_{\Phi}=\frac{1}{2}\ln\kappa_{\Phi} C Φ = 2 1 ln κ Φ and d Φ = − ln Φ ( ϵ 1 ) d_{\Phi}=-\ln\Phi(\epsilon_{1}) d Φ = − ln Φ ( ϵ 1 ) ; recall that Φ \Phi Φ is monotone (G1) and Φ ( ϵ 1 ) ∥ v ∥ ∞ ≤ Φ ( v ) ≤ Φ ( 1 ) ∥ v ∥ ∞ \Phi(\epsilon_{1})\|v\|_{\infty}\le\Phi(v)\le\Phi(\mathbf{1})\|v\|_{\infty} Φ ( ϵ 1 ) ∥ v ∥ ∞ ≤ Φ ( v ) ≤ Φ ( 1 ) ∥ v ∥ ∞ (G2).
(a) Let w = P ( x 1 / 2 ) y w=P(x^{1/2})y w = P ( x 1/2 ) y . Since y ≤ λ max ( y ) e y\le\lambda_{\max}(y)e y ≤ λ m a x ( y ) e and P ( x 1 / 2 ) P(x^{1/2}) P ( x 1/2 ) preserves the order, w ≤ λ max ( y ) x w\le\lambda_{\max}(y)x w ≤ λ m a x ( y ) x , so by the min–max principle λ ↓ ( w ) ≤ λ max ( y ) λ ↓ ( x ) \lambda^{\downarrow}(w)\le\lambda_{\max}(y)\lambda^{\downarrow}(x) λ ↓ ( w ) ≤ λ m a x ( y ) λ ↓ ( x ) coordinatewise, and
Φ ( λ ( w ) ) ≤ λ max ( y ) Φ ( λ ( x ) ) ≤ Φ ( λ ( x ) ) Φ ( λ ( y ) ) Φ ( ϵ 1 ) . \Phi(\lambda(w))\le\lambda_{\max}(y)\Phi(\lambda(x))\le\frac{\Phi(\lambda(x))\Phi(\lambda(y))}{\Phi(\epsilon_{1})}. Φ ( λ ( w )) ≤ λ m a x ( y ) Φ ( λ ( x )) ≤ Φ ( ϵ 1 ) Φ ( λ ( x )) Φ ( λ ( y )) .
By (F1), w − 1 = P ( x − 1 / 2 ) y − 1 w^{-1}=P(x^{-1/2})y^{-1} w − 1 = P ( x − 1/2 ) y − 1 , so the same holds for the inverses; hence
κ Φ ( w ) ≤ Φ ( ϵ 1 ) − 2 κ Φ ( x ) κ Φ ( y ) . \kappa_{\Phi}(w)\le\Phi(\epsilon_{1})^{-2}\kappa_{\Phi}(x)\kappa_{\Phi}(y). κ Φ ( w ) ≤ Φ ( ϵ 1 ) − 2 κ Φ ( x ) κ Φ ( y ) .
For 0 < p < 1 0<p<1 0 < p < 1 the same argument applies to v ↦ ( 1 r ∑ i v i p ) 1 / p v\mapsto\left(\frac{1}{r}\sum_i v_i^p\right)^{1/p} v ↦ ( r 1 ∑ i v i p ) 1/ p , which is monotone and satisfies r − 1 / p ∥ v ∥ ∞ ≤ ( 1 r ∑ i v i p ) 1 / p r^{-1/p}\|v\|_{\infty}\le\left(\frac{1}{r}\sum_i v_i^p\right)^{1/p} r − 1/ p ∥ v ∥ ∞ ≤ ( r 1 ∑ i v i p ) 1/ p .
(b) and (c) depend only on λ = λ ( x ) \lambda=\lambda(x) λ = λ ( x ) , since λ ( x ± s ) \lambda(x^{\pm s}) λ ( x ± s ) is λ ± s \lambda^{\pm s} λ ± s up to order. (b) For s = 1 s=1 s = 1 there is nothing to prove; let s > 1 s>1 s > 1 . By (G3) with v = λ s v=\lambda^s v = λ s , w = 1 w=\mathbf{1} w = 1 , θ = 1 / s \theta=1/s θ = 1/ s , Φ ( λ ) ≤ Φ ( λ s ) 1 / s Φ ( 1 ) 1 − 1 / s = Φ ( λ s ) 1 / s \Phi(\lambda)\le\Phi(\lambda^s)^{1/s}\Phi(\mathbf{1})^{1-1/s}=\Phi(\lambda^s)^{1/s} Φ ( λ ) ≤ Φ ( λ s ) 1/ s Φ ( 1 ) 1 − 1/ s = Φ ( λ s ) 1/ s , and likewise for λ − 1 \lambda^{-1} λ − 1 ; multiplying, κ Φ ( x ) ≤ κ Φ ( x s ) \kappa_{\Phi}(x)\le\kappa_{\Phi}(x^s) κ Φ ( x ) ≤ κ Φ ( x s ) . By (G1) and (G2), Φ ( λ s ) ≤ ∥ λ ∥ ∞ s − 1 Φ ( λ ) ≤ Φ ( ϵ 1 ) 1 − s Φ ( λ ) s \Phi(\lambda^s)\le\|\lambda\|_{\infty}^{s-1}\Phi(\lambda)\le\Phi(\epsilon_{1})^{1-s}\Phi(\lambda)^s Φ ( λ s ) ≤ ∥ λ ∥ ∞ s − 1 Φ ( λ ) ≤ Φ ( ϵ 1 ) 1 − s Φ ( λ ) s , and likewise for λ − s \lambda^{-s} λ − s , so κ Φ ( x s ) ≤ Φ ( ϵ 1 ) − 2 ( s − 1 ) κ Φ ( x ) s \kappa_{\Phi}(x^s)\le\Phi(\epsilon_{1})^{-2(s-1)}\kappa_{\Phi}(x)^s κ Φ ( x s ) ≤ Φ ( ϵ 1 ) − 2 ( s − 1 ) κ Φ ( x ) s . Taking half logarithms gives (b). (c) Here λ ( x t ± s ) = ( t ± s , 1 , … , 1 , t ∓ s ) \lambda(x_t^{\pm s})=(t^{\pm s},1,\ldots,1,t^{\mp s}) λ ( x t ± s ) = ( t ± s , 1 , … , 1 , t ∓ s ) , so t − s λ ( x t s ) → ϵ 1 t^{-s}\lambda(x_t^s)\to\epsilon_1 t − s λ ( x t s ) → ϵ 1 and t − s λ ( x t − s ) → ϵ r t^{-s}\lambda(x_t^{-s})\to\epsilon_r t − s λ ( x t − s ) → ϵ r as t → ∞ t\to\infty t → ∞ ; by continuity, homogeneity and permutation invariance of Φ \Phi Φ , t − 2 s κ Φ ( x t s ) → Φ ( ϵ 1 ) 2 t^{-2s}\kappa_{\Phi}(x_t^s)\to\Phi(\epsilon_1)^2 t − 2 s κ Φ ( x t s ) → Φ ( ϵ 1 ) 2 for every s ≥ 1 s\ge1 s ≥ 1 . Hence
C Φ ( x t s ) − s C Φ ( x t ) = 1 2 ln t − 2 s κ Φ ( x t s ) ( t − 2 κ Φ ( x t ) ) s ⟶ 1 2 ln Φ ( ϵ 1 ) 2 − 2 s = ( s − 1 ) d Φ ; C_{\Phi}(x_t^s)-sC_{\Phi}(x_t)=\frac{1}{2}\ln\frac{t^{-2s}\kappa_{\Phi}(x_t^s)}{(t^{-2}\kappa_{\Phi}(x_t))^s}\longrightarrow\frac{1}{2}\ln\Phi(\epsilon_1)^{2-2s}=(s-1)d_{\Phi}; C Φ ( x t s ) − s C Φ ( x t ) = 2 1 ln ( t − 2 κ Φ ( x t ) ) s t − 2 s κ Φ ( x t s ) ⟶ 2 1 ln Φ ( ϵ 1 ) 2 − 2 s = ( s − 1 ) d Φ ;
note that P ( x t 1 / 2 ) x t = x t 2 P(x_t^{1/2})x_t=x_t^2 P ( x t 1/2 ) x t = x t 2 . For 0 < p < 1 0<p<1 0 < p < 1 the computation in (c) is the same, since it uses only continuity, homogeneity and permutation invariance.
(d) If d Φ > 0 d_{\Phi}>0 d Φ > 0 , (c) gives a violation; if d Φ = 0 d_{\Phi}=0 d Φ = 0 , (a) is subadditivity. And Φ ( ϵ 1 ) = Φ ( 1 ) \Phi(\epsilon_1)=\Phi(\mathbf{1}) Φ ( ϵ 1 ) = Φ ( 1 ) forces Φ ( v ) = Φ ( ϵ 1 ) ∥ v ∥ ∞ \Phi(v)=\Phi(\epsilon_1)\|v\|_{\infty} Φ ( v ) = Φ ( ϵ 1 ) ∥ v ∥ ∞ for all v v v , by the two-sided bound (G2).
The key inequality λ ↓ ( P ( x 1 / 2 ) y ) ≤ λ max ( y ) λ ↓ ( x ) \lambda^{\downarrow}(P(x^{1/2})y)\le\lambda_{\max}(y)\lambda^{\downarrow}(x) λ ↓ ( P ( x 1/2 ) y ) ≤ λ m a x ( y ) λ ↓ ( x ) and the Schatten case of (a) are in Tao, Jeong and Gowda: [95 Lemma 3.2] states it with the roles of x , y x,y x , y exchanged, which is equivalent by Lemma 6.5.1 , and [95 Cor. 5.2(ii)] with r = p r=p r = p , s = ∞ s=\infty s = ∞ gives the norm bound ∥ P ( x 1 / 2 ) y ∥ p ≤ ∥ x ∥ p ∥ y ∥ ∞ \|P(x^{1/2})y\|_p\le\|x\|_p\|y\|_{\infty} ∥ P ( x 1/2 ) y ∥ p ≤ ∥ x ∥ p ∥ y ∥ ∞ . Their log-majorization λ ↓ ( P ( x 1 / 2 ) y ) ≺ log λ ↓ ( x ) ∗ λ ↓ ( y ) \lambda^{\downarrow}(P(x^{1/2})y)\prec_{\log}\lambda^{\downarrow}(x)*\lambda^{\downarrow}(y) λ ↓ ( P ( x 1/2 ) y ) ≺ l o g λ ↓ ( x ) ∗ λ ↓ ( y ) [95 Thm. 3.4] gives the same bound in (a), and no better one, since the extremal pair in (c) has λ ( x t 2 ) = λ ( x t ) ∗ λ ( x t ) \lambda(x_t^2)=\lambda(x_t)*\lambda(x_t) λ ( x t 2 ) = λ ( x t ) ∗ λ ( x t ) . What is added here is the identification of the sharp defect with the rank, and the rigidity. So the two basic constants of the theory are governed by the rank: the non-commutativity threshold of Theorem 6.4.1 (rk V ≤ 3 \operatorname{rk}V\le3 rk V ≤ 3 ) and the subadditivity defect ln r p \frac{\ln r}{p} p l n r of the Schatten contrasts. On a Lorentz cone, whose eigenvalues are ρ e ± t \rho e^{\pm t} ρ e ± t ,
C p ( x ) = 1 p ln cosh ( p OLC ( x ) ) , (13) C_p(x)=\frac{1}{p}\ln\cosh(p\,\operatorname{OLC}(x)), \tag*{(13)} C p ( x ) = p 1 ln cosh ( p OLC ( x )) , (13)
a function of the rapidity alone.
The Lorentz cone: contrast as velocity
Let V n = R × R n V_n = \mathbb{R} \times\mathbb{R}^n V n = R × R n , n ≥ 1 n \ge1 n ≥ 1 , with ( x 0 , x ˉ ) ∘ ( y 0 , y ˉ ) = ( x 0 y 0 + x ˉ ⋅ y ˉ , x 0 y ˉ + y 0 x ˉ ) (x_0,\bar{x}) \circ(y_0,\bar{y}) = (x_0y_0 + \bar{x} \cdot\bar{y},\ x_0\bar{y} + y_0\bar{x}) ( x 0 , x ˉ ) ∘ ( y 0 , y ˉ ) = ( x 0 y 0 + x ˉ ⋅ y ˉ , x 0 y ˉ + y 0 x ˉ ) , e = ( 1 , 0 ) e = (1,0) e = ( 1 , 0 ) , det x = x 0 2 − ∣ x ˉ ∣ 2 \det x = x_0^2 - |\bar{x}|^2 det x = x 0 2 − ∣ x ˉ ∣ 2 , and Minkowski form ⟨ x , y ⟩ M = x 0 y 0 − x ˉ ⋅ y ˉ \langle x,y\rangle_{\mathrm{M}} = x_0y_0 - \bar{x}\cdot\bar{y} ⟨ x , y ⟩ M = x 0 y 0 − x ˉ ⋅ y ˉ , so that det x = ⟨ x , x ⟩ M \det x = \langle x,x\rangle_{\mathrm{M}} det x = ⟨ x , x ⟩ M . The cone Ω \Omega Ω is the open forward light cone, the eigenvalues of x x x are x 0 ± ∣ x ˉ ∣ x_0 \pm|\bar{x}| x 0 ± ∣ x ˉ ∣ , and x − 1 = ( x 0 , − x ˉ ) / det x x^{-1} = (x_0,-\bar{x})/\det x x − 1 = ( x 0 , − x ˉ ) / det x [96 Ch. II] . Let H n = { x ∈ Ω : det x = 1 } H^n = \{x \in\Omega: \det x = 1\} H n = { x ∈ Ω : det x = 1 } be the upper sheet of the hyperboloid with the hyperbolic distance d ( x , y ) = arcosh ⟨ x , y ⟩ M d(x,y) = \operatorname{arcosh}\langle x,y\rangle_{\mathrm{M}} d ( x , y ) = arcosh ⟨ x , y ⟩ M . Every x ∈ Ω x \in\Omega x ∈ Ω is x = ρ ( cosh η , sinh η u ˉ ) x = \rho(\cosh\eta,\sinh\eta\,\bar{u}) x = ρ ( cosh η , sinh η u ˉ ) with ρ > 0 \rho> 0 ρ > 0 , η ≥ 0 \eta\ge0 η ≥ 0 , ∣ u ˉ ∣ = 1 |\bar{u}| = 1 ∣ u ˉ ∣ = 1 , and then
OMC ( x ) = ∣ x ˉ ∣ x 0 = tanh η , OLC ( x ) = η = d ( e , x det x ) : \operatorname{OMC}(x) = \frac{|\bar{x}|}{x_0} = \tanh\eta,\qquad\operatorname{OLC}(x) = \eta= d\left(e,\frac{x}{\sqrt{\det x}}\right): OMC ( x ) = x 0 ∣ x ˉ ∣ = tanh η , OLC ( x ) = η = d ( e , det x x ) :
the Michelson contrast is the speed of the four-velocity x / det x x/\sqrt{\det x} x / det x and the logarithmic contrast its rapidity.
Lemma 6.7.1. For a = ( cosh t , sinh t u ˉ ) ∈ H n a = (\cosh t,\sinh t\,\bar{u}) \in H^n a = ( cosh t , sinh t u ˉ ) ∈ H n , P ( a ) P(a) P ( a ) is the pure Lorentz boost (a symmetric positive definite element of S O + ( 1 , n ) \mathrm{SO}^{+}(1,n) SO + ( 1 , n ) ) of rapidity 2 t 2t 2 t along u ˉ \bar{u} u ˉ , with P ( a ) e = a 2 P(a)e = a^2 P ( a ) e = a 2 ; in particular P ( a ) P(a) P ( a ) preserves ⟨ ⋅ , ⋅ ⟩ M \langle\cdot,\cdot\rangle_{\mathrm{M}} ⟨ ⋅ , ⋅ ⟩ M and H n H^n H n .
Proof. P ( a ) P(a) P ( a ) is self-adjoint for the Euclidean inner product, and positive definite since P ( a ) = P ( a 1 / 2 ) 2 P(a) = P(a^{1/2})^2 P ( a ) = P ( a 1/2 ) 2 by the fundamental formula [96 Prop. II.3.3] , and det ( P ( a ) y ) = ( det a ) 2 det y = det y \det(P(a)y) = (\det a)^2\det y = \det y det ( P ( a ) y ) = ( det a ) 2 det y = det y [96 Prop. III.4.2] , so P ( a ) P(a) P ( a ) preserves the quadratic form det \det det and hence, by polarisation, ⟨ ⋅ , ⋅ ⟩ M \langle\cdot,\cdot\rangle_{\mathrm{M}} ⟨ ⋅ , ⋅ ⟩ M . It maps Ω \Omega Ω onto Ω \Omega Ω , so it lies in the orthochronous Lorentz group; its determinant is positive since it is positive definite. A symmetric positive definite element of S O + ( 1 , n ) \mathrm{SO}^{+}(1,n) SO + ( 1 , n ) is a pure boost (polar decomposition of the Lorentz group), and P ( a ) e = a 2 = ( cosh 2 t , sinh 2 t u ˉ ) P(a)e = a^2 = (\cosh2t,\sinh2t\,\bar{u}) P ( a ) e = a 2 = ( cosh 2 t , sinh 2 t u ˉ ) by (F1), which identifies the axis and the rapidity.
Theorem 6.7.2. (a) For x , y ∈ H n x,y \in H^n x , y ∈ H n , the Thompson distance d T ( x , y ) = ∥ log P ( x − 1 / 2 ) y ∥ ∞ d_{\mathrm{T}}(x,y) = \lVert\log P(x^{-1/2})y\rVert_\infty d T ( x , y ) = ∥ log P ( x − 1/2 ) y ∥ ∞ , the projectivised Thompson distance δ ( [ x ] , [ y ] ) \delta([x],[y]) δ ([ x ] , [ y ]) and the hyperbolic distance coincide:
d T ( x , y ) = δ ( [ x ] , [ y ] ) = d ( x , y ) ; d_{\mathrm{T}}(x,y) = \delta([x],[y]) = d(x,y); d T ( x , y ) = δ ([ x ] , [ y ]) = d ( x , y ) ; Hilbert’s projective metric is 2 d ( x , y ) 2d(x,y) 2 d ( x , y ) .
(b) For x , y ∈ Ω x,y \in\Omega x , y ∈ Ω with rapidities η x = OLC ( x ) \eta_x = \operatorname{OLC}(x) η x = OLC ( x ) , η y = OLC ( y ) \eta_y = \operatorname{OLC}(y) η y = OLC ( y ) and angle θ \theta θ between x ˉ \bar{x} x ˉ and y ˉ \bar{y} y ˉ (any θ \theta θ if one of them is 0 0 0 ),
cosh OLC ( P ( x 1 / 2 ) y ) = cosh η x cosh η y + sinh η x sinh η y cos θ , \cosh\operatorname{OLC}(P(x^{1/2})y) = \cosh\eta_x\cosh\eta_y + \sinh\eta_x\sinh\eta_y\cos\theta, cosh OLC ( P ( x 1/2 ) y ) = cosh η x cosh η y + sinh η x sinh η y cos θ , the hyperbolic law of cosines for the triangle with sides η x \eta_x η x , η y \eta_y η y and angle π − θ \pi-\theta π − θ between them.
(c) Consequently OLC ( P ( x 1 / 2 ) y ) ≤ OLC ( x ) + OLC ( y ) \operatorname{OLC}(P(x^{1/2})y) \le\operatorname{OLC}(x) + \operatorname{OLC}(y) OLC ( P ( x 1/2 ) y ) ≤ OLC ( x ) + OLC ( y ) , with equality if and only if x ˉ \bar{x} x ˉ and y ˉ \bar{y} y ˉ are non-negatively proportional (this includes x ˉ = 0 \bar{x}=0 x ˉ = 0 or y ˉ = 0 \bar{y}=0 y ˉ = 0 ), and OMC ( P ( x 1 / 2 ) y ) ≤ OMC ( x ) ⊕ OMC ( y ) \operatorname{OMC}(P(x^{1/2})y) \le\operatorname{OMC}(x) \oplus\operatorname{OMC}(y) OMC ( P ( x 1/2 ) y ) ≤ OMC ( x ) ⊕ OMC ( y ) , u ⊕ v = u + v 1 + u v u \oplus v = \frac{u+v}{1+uv} u ⊕ v = 1 + uv u + v , with the same equality case. By contrast, x x x and y y y operator commute if and only if x ˉ \bar{x} x ˉ and y ˉ \bar{y} y ˉ are linearly dependent.
Proof. (a) By Lemma 6.7.1, B = P ( x − 1 / 2 ) B = P(x^{-1/2}) B = P ( x − 1/2 ) is a boost with B x = e Bx = e B x = e , so z = B y ∈ H n z = By \in H^n z = B y ∈ H n and ⟨ e , z ⟩ M = ⟨ B x , B y ⟩ M = ⟨ x , y ⟩ M \langle e,z\rangle_{\mathrm{M}} = \langle Bx,By\rangle_{\mathrm{M}} = \langle x,y\rangle_{\mathrm{M}} ⟨ e , z ⟩ M = ⟨ B x , B y ⟩ M = ⟨ x , y ⟩ M . The eigenvalues of z z z are z 0 ± ∣ z ˉ ∣ = e ± s z_0 \pm|\bar{z}| = e^{\pm s} z 0 ± ∣ z ˉ ∣ = e ± s with cosh s = z 0 = ⟨ x , y ⟩ M \cosh s = z_0 = \langle x,y\rangle_{\mathrm{M}} cosh s = z 0 = ⟨ x , y ⟩ M , so ∥ log z ∥ ∞ = s = d ( x , y ) \lVert\log z\rVert_\infty= s = d(x,y) ∥ log z ∥ ∞ = s = d ( x , y ) . Since det z = 1 \det z = 1 det z = 1 , the infimum over λ > 0 \lambda> 0 λ > 0 of ∥ log ( λ z ) ∥ ∞ \lVert\log(\lambda z)\rVert_\infty ∥ log ( λ z ) ∥ ∞ is attained at λ = 1 \lambda= 1 λ = 1 , which gives δ \delta δ , and log ( λ max ( z ) / λ min ( z ) ) = 2 s \log(\lambda_{\max}(z)/\lambda_{\min}(z)) = 2s log ( λ m a x ( z ) / λ m i n ( z )) = 2 s .
(b) Both sides are invariant under x ↦ λ x x \mapsto\lambda x x ↦ λ x , y ↦ μ y y \mapsto\mu y y ↦ μ y , so let x , y ∈ H n x,y \in H^n x , y ∈ H n . Then B = P ( x 1 / 2 ) B = P(x^{1/2}) B = P ( x 1/2 ) is a boost, z = B y ∈ H n z = By \in H^n z = B y ∈ H n , and OLC ( z ) = arcosh z 0 \operatorname{OLC}(z) = \operatorname{arcosh} z_0 OLC ( z ) = arcosh z 0 by the formula for the eigenvalues. Since B B B preserves ⟨ ⋅ , ⋅ ⟩ M \langle\cdot,\cdot\rangle_{\mathrm{M}} ⟨ ⋅ , ⋅ ⟩ M and B − 1 = P ( x − 1 / 2 ) B^{-1} = P(x^{-1/2}) B − 1 = P ( x − 1/2 ) maps e e e to x − 1 x^{-1} x − 1 , z 0 = ⟨ e , B y ⟩ M = ⟨ B − 1 e , y ⟩ M = ⟨ x − 1 , y ⟩ M = x 0 y 0 + x ˉ ⋅ y ˉ z_0 = \langle e,By\rangle_{\mathrm{M}} = \langle B^{-1}e,y\rangle_{\mathrm{M}} = \langle x^{-1},y\rangle_{\mathrm{M}} = x_0y_0 + \bar{x}\cdot\bar{y} z 0 = ⟨ e , B y ⟩ M = ⟨ B − 1 e , y ⟩ M = ⟨ x − 1 , y ⟩ M = x 0 y 0 + x ˉ ⋅ y ˉ , which is the right-hand side.
(c) The right-hand side of (b) is at most cosh ( η x + η y ) \cosh(\eta_x+\eta_y) cosh ( η x + η y ) , with equality iff cos θ = 1 \cos\theta= 1 cos θ = 1 or η x η y = 0 \eta_x\eta_y = 0 η x η y = 0 , and tanh \tanh tanh is increasing, with tanh ( η x + η y ) = tanh η x ⊕ tanh η y \tanh(\eta_x+\eta_y) = \tanh\eta_x \oplus\tanh\eta_y tanh ( η x + η y ) = tanh η x ⊕ tanh η y . For the last statement, x x x and y y y have a common Jordan frame iff they lie in a common subalgebra R e ⊕ R ( 0 , u ˉ ) \mathbb{R}e \oplus\mathbb{R}(0,\bar{u}) R e ⊕ R ( 0 , u ˉ ) , i.e. iff x ˉ , y ˉ \bar{x},\bar{y} x ˉ , y ˉ are linearly dependent (F5).
So on the Lorentz cone subadditivity of OLC is the triangle inequality of hyperbolic space, P ( x 1 / 2 ) y P(x^{1/2})y P ( x 1/2 ) y is y y y seen after the boost that takes the observer e e e to x x x , and (b) is the composition law of rapidities; equality requires collinear boosts in the same direction, which is strictly more than commutativity (opposite boosts commute, and then OLC ( P ( x 1 / 2 ) y ) = ∣ η x − η y ∣ \operatorname{OLC}(P(x^{1/2})y)=\lvert\eta_x-\eta_y\rvert OLC ( P ( x 1/2 ) y ) = ∣ η x − η y ∣ ). The identification of rapidity with hyperbolic distance on boosts and on positive 2 × 2 2 \times2 2 × 2 matrices is classical [92 , 93 , 94 ] ; the equality case of the hyperbolic triangle inequality is classical too, and Theorem 6.7.2 only records all this in terms of the contrast.
The Jordan product
For matrices X ∘ Y = 1 2 ( X Y + Y X ) X \circ Y=\frac{1}{2}(XY+YX) X ∘ Y = 2 1 ( X Y + Y X ) , and estimating the spectrum of the Jordan product of two positive matrices in terms of their spectra is a research problem of Taussky-Todd. Strang [citep][89] found the sharp bounds for the extreme eigenvalues in every dimension ≥ 2 \ge2 ≥ 2 , and in Hilbert space; see also Gustafson [100 ] , Nicholson [108 ] , de Bruijn [109 ] , and Alikakos and Bates [110 ] with the erratum [99 ] , which states Strang’s lower bound in explicit form. These bounds concern λ max \lambda_{\max} λ m a x and λ min \lambda_{\min} λ m i n separately. The contrast is their ratio, and the point of this section is that whether the two bounds can be attained at the same time depends on the algebra: it is again a question of room, as in Theorem 6.4.1 .
For x ∈ Ω x \in\Omega x ∈ Ω write μ ( x ) = 1 2 ( λ max ( x ) + λ min ( x ) ) \mu(x)=\frac{1}{2}(\lambda_{\max}(x)+\lambda_{\min}(x)) μ ( x ) = 2 1 ( λ m a x ( x ) + λ m i n ( x )) , so that the spectrum of x x x lies in μ ( x ) [ 1 − u , 1 + u ] \mu(x)[1-u,1+u] μ ( x ) [ 1 − u , 1 + u ] with u = OMC ( x ) u=\operatorname{OMC}(x) u = OMC ( x ) and both endpoints occur. For 0 ≤ u , v < 1 0 \le u,v < 1 0 ≤ u , v < 1 put
ℓ ( u , v ) = { ( 1 − u ) ( 1 − v ) , u + v ≤ 1 , 1 2 ( 1 − u 2 − v 2 ) , u + v ≥ 1 , u ⊕ v = u + v 1 + u v . \ell(u,v)=
\begin{cases}
(1-u)(1-v), & u+v\le1,\\
\frac{1}{2}(1-u^2-v^2), & u+v\ge1,
\end{cases}
\qquad
u\oplus v=\frac{u+v}{1+uv}. ℓ ( u , v ) = { ( 1 − u ) ( 1 − v ) , 2 1 ( 1 − u 2 − v 2 ) , u + v ≤ 1 , u + v ≥ 1 , u ⊕ v = 1 + uv u + v .
(The two expressions agree on u + v = 1 u+v=1 u + v = 1 , and for u + v ≠ 1 u+v\ne1 u + v = 1 one has 1 2 ( 1 − u 2 − v 2 ) < ( 1 − u ) ( 1 − v ) \frac{1}{2}(1-u^2-v^2)<(1-u)(1-v) 2 1 ( 1 − u 2 − v 2 ) < ( 1 − u ) ( 1 − v ) , since the difference is 1 2 ( u + v − 1 ) 2 \frac{1}{2}(u+v-1)^2 2 1 ( u + v − 1 ) 2 .)
Lemma 6.8.1 (Strang’s bounds in Jordan form). Let A A A be a unital J B JB J B -algebra [98 ] , A + + A_{++} A ++ the set of invertible elements of its cone of squares, and define λ max , λ min , μ , OMC \lambda_{\max},\lambda_{\min},\mu,\operatorname{OMC} λ m a x , λ m i n , μ , OMC on A + + A_{++} A ++ as above (see also § 9.11 ). Let x , y ∈ A + + x,y\in A_{++} x , y ∈ A ++ , u = OMC ( x ) u=\operatorname{OMC}(x) u = OMC ( x ) , v = OMC ( y ) v=\operatorname{OMC}(y) v = OMC ( y ) . Then
λ max ( x ∘ y ) ≤ λ max ( x ) λ max ( y ) , λ min ( x ∘ y ) ≥ μ ( x ) μ ( y ) ℓ ( u , v ) . \lambda_{\max}(x\circ y)\le\lambda_{\max}(x)\lambda_{\max}(y),\qquad
\lambda_{\min}(x\circ y)\ge\mu(x)\mu(y)\ell(u,v). λ m a x ( x ∘ y ) ≤ λ m a x ( x ) λ m a x ( y ) , λ m i n ( x ∘ y ) ≥ μ ( x ) μ ( y ) ℓ ( u , v ) .
In particular x ∘ y ∈ A + + x\circ y\in A_{++} x ∘ y ∈ A ++ whenever u 2 + v 2 < 1 u^2+v^2<1 u 2 + v 2 < 1 .
:::
Proof. The upper bound is ∥ x ∘ y ∥ ≤ ∥ x ∥ ∥ y ∥ \lVert x\circ y\rVert\le\lVert x\rVert\lVert y\rVert ∥ x ∘ y ∥ ≤ ∥ x ∥ ∥ y ∥ . Let C C C be the closed subalgebra generated by e e e , x x x , y y y . By the Shirshov–Cohn theorem [98 Thm. 7.2.5] , C C C is a JC-algebra: there is a Jordan isomorphism π \pi π of C C C onto a norm-closed Jordan algebra of self-adjoint operators on a Hilbert space, and compressing to the range of the projection π ( e ) \pi(e) π ( e ) we may assume π ( e ) = 1 \pi(e)=1 π ( e ) = 1 . The spectrum of an element of C C C is computed in the associative subalgebra generated by it and e e e , so it is the same in C C C and in A A A , and π \pi π preserves it. Hence it suffices to bound λ min \lambda_{\min} λ m i n of 1 2 ( X Y + Y X ) \frac{1}{2}(XY+YX) 2 1 ( X Y + Y X ) for positive operators X , Y X,Y X , Y with spectra in μ ( x ) [ 1 − u , 1 + u ] \mu(x)[1-u,1+u] μ ( x ) [ 1 − u , 1 + u ] and μ ( y ) [ 1 − v , 1 + v ] \mu(y)[1-v,1+v] μ ( y ) [ 1 − v , 1 + v ] (if u = 0 u=0 u = 0 or v = 0 v=0 v = 0 the claim is trivial, so let u , v > 0 u,v>0 u , v > 0 ). By Strang’s theorem [89 Thm. I] , which is stated for operators on Hilbert space (and in any case reduces to finite dimension: for a unit vector w w w and the orthogonal projection Q Q Q onto F = span { w , X w , Y w } F=\operatorname{span}\{w,Xw,Yw\} F = span { w , X w , Y w } , the compressions Q X ∣ F , Q Y ∣ F QX|_F,QY|_F QX ∣ F , Q Y ∣ F have spectra in the same intervals, and ⟨ ( X Y + Y X ) w , w ⟩ = 2 Re ⟨ Y w , X w ⟩ \langle(XY+YX)w,w\rangle=2\operatorname{Re}\langle Yw,Xw\rangle ⟨( X Y + Y X ) w , w ⟩ = 2 Re ⟨ Y w , X w ⟩ does not change when X , Y X,Y X , Y are replaced by them), in the explicit form of [99 ] ,
λ min ( X Y + Y X ) ≥ 2 m m ′ if M ′ m ′ ≤ M / m + 3 M / m − 1 \lambda_{\min}(XY+YX)\ge2mm'\ \text{if}\ \frac{M'}{m'}\le\frac{M/m+3}{M/m-1} λ m i n ( X Y + Y X ) ≥ 2 m m ′ if m ′ M ′ ≤ M / m − 1 M / m + 3
and
λ min ( X Y + Y X ) ≥ 16 M m M ′ m ′ − ( M − m ) 2 ( M ′ − m ′ ) 2 4 ( M + m ) ( M ′ + m ′ ) \lambda_{\min}(XY+YX)\ge
\frac{16MmM'm'-(M-m)^2(M'-m')^2}{4(M+m)(M'+m')} λ m i n ( X Y + Y X ) ≥ 4 ( M + m ) ( M ′ + m ′ ) 16 M m M ′ m ′ − ( M − m ) 2 ( M ′ − m ′ ) 2
otherwise, where σ ( X ) ⊂ [ m , M ] \sigma(X)\subset[m,M] σ ( X ) ⊂ [ m , M ] , σ ( Y ) ⊂ [ m ′ , M ′ ] \sigma(Y)\subset[m',M'] σ ( Y ) ⊂ [ m ′ , M ′ ] . With M , m = μ ( x ) ( 1 ± u ) M,m=\mu(x)(1\pm u) M , m = μ ( x ) ( 1 ± u ) and M ′ , m ′ = μ ( y ) ( 1 ± v ) M',m'=\mu(y)(1\pm v) M ′ , m ′ = μ ( y ) ( 1 ± v ) the condition reads u + v ≤ 1 u+v\le1 u + v ≤ 1 , 2 m m ′ = 2 μ ( x ) μ ( y ) ( 1 − u ) ( 1 − v ) 2mm'=2\mu(x)\mu(y)(1-u)(1-v) 2 m m ′ = 2 μ ( x ) μ ( y ) ( 1 − u ) ( 1 − v ) , and the second bound is μ ( x ) μ ( y ) ( 1 − u 2 − v 2 ) \mu(x)\mu(y)(1-u^2-v^2) μ ( x ) μ ( y ) ( 1 − u 2 − v 2 ) . Halving gives the claim. Positivity for u 2 + v 2 < 1 u^2+v^2<1 u 2 + v 2 < 1 follows, since then both expressions for ℓ \ell ℓ are positive; this criterion is also [100 Cor. 4.4] .
Theorem 6.8.2. Let V V V be a non-associative Euclidean Jordan algebra, and for 0 ≤ u , v ≤ 1 0\le u,v\le1 0 ≤ u , v ≤ 1 with u 2 + v 2 < 1 u^2+v^2<1 u 2 + v 2 < 1 put
S V ( u , v ) = sup { OMC ( x ∘ y ) : x , y ∈ Ω , OMC ( x ) = u , OMC ( y ) = v } . S_V(u,v)=\sup\{\operatorname{OMC}(x\circ y):x,y\in\Omega,\ \operatorname{OMC}(x)=u,\ \operatorname{OMC}(y)=v\}. S V ( u , v ) = sup { OMC ( x ∘ y ) : x , y ∈ Ω , OMC ( x ) = u , OMC ( y ) = v } . (a) If V V V is a spin factor V n V_n V n , n ≥ 2 n\ge2 n ≥ 2 , then S V ( u , v ) = u ⊕ v S_V(u,v)=u\oplus v S V ( u , v ) = u ⊕ v if u 2 + u v + v 2 ≤ 1 u^2+uv+v^2\le1 u 2 + uv + v 2 ≤ 1 , and S V ( u , v ) = ( 2 − u 2 − v 2 ) − 1 / 2 S_V(u,v)=(2-u^2-v^2)^{-1/2} S V ( u , v ) = ( 2 − u 2 − v 2 ) − 1/2 otherwise .
(b) If V V V is not a spin factor, then S V ( u , v ) = u ⊕ v S_V(u,v)=u\oplus v S V ( u , v ) = u ⊕ v if u + v ≤ 1 u+v\le1 u + v ≤ 1 , and otherwise S V ( u , v ) = κ ∗ − 1 κ ∗ + 1 S_V(u,v)=\frac{\kappa_*-1}{\kappa_*+1} S V ( u , v ) = κ ∗ + 1 κ ∗ − 1 with κ ∗ = 2 ( 1 + u ) ( 1 + v ) 1 − u 2 − v 2 \kappa_*=\frac{2(1+u)(1+v)}{1-u^2-v^2} κ ∗ = 1 − u 2 − v 2 2 ( 1 + u ) ( 1 + v ) .
In both cases the supremum is attained. Consequently, for u 2 + v 2 < 1 u^2+v^2<1 u 2 + v 2 < 1 , OLC ( x ∘ y ) ≤ OLC ( x ) + OLC ( y ) \operatorname{OLC}(x\circ y)\le\operatorname{OLC}(x)+\operatorname{OLC}(y) OLC ( x ∘ y ) ≤ OLC ( x ) + OLC ( y ) for all x , y ∈ Ω ∗ w i t h ∗ x,y\in\Omega*with* x , y ∈ Ω ∗ w i t h ∗ \operatorname{OMC}(x)=u$, OMC ( y ) = v \operatorname{OMC}(y)=v OMC ( y ) = v if and only if u 2 + u v + v 2 ≤ 1 u^2+uv+v^2\le1 u 2 + uv + v 2 ≤ 1 in case (a), and if and only if u + v ≤ 1 u+v\le1 u + v ≤ 1 in case (b).
Since tanh ( OLC ( x ) + OLC ( y ) ) = u ⊕ v \tanh(\operatorname{OLC}(x)+\operatorname{OLC}(y))=u\mathbin{\oplus}v tanh ( OLC ( x ) + OLC ( y )) = u ⊕ v , the last statement follows from (a) and (b) once we check that the second expressions exceed u ⊕ v u\mathbin{\oplus}v u ⊕ v outside the respective regions; this is done in the proof. If V V V is associative, x ∘ y x\circ y x ∘ y is the product of commuting elements and OLC ( x ∘ y ) ≤ OLC ( x ) + OLC ( y ) \operatorname{OLC}(x\circ y)\leq\operatorname{OLC}(x)+\operatorname{OLC}(y) OLC ( x ∘ y ) ≤ OLC ( x ) + OLC ( y ) always (Theorem 6.3.3 ). The ellipse u 2 + u v + v 2 ≤ 1 u^2+uv+v^2\leq1 u 2 + uv + v 2 ≤ 1 strictly contains the triangle u + v ≤ 1 u+v\leq1 u + v ≤ 1 , and both lie inside the disc u 2 + v 2 < 1 u^2+v^2<1 u 2 + v 2 < 1 where x ∘ y x\circ y x ∘ y is positive. By Corollary IV.1.5 of [96 ] the simple algebras of rank two are spin factors (and conversely V n V_n V n , n ≥ 2 n\geq2 n ≥ 2 , is simple, as is well known), so case (b) consists of the simple algebras of rank ≥ 3 \geq3 ≥ 3 and of the non-simple algebras with a non-associative summand.
Proof. (a) Write x = α ( 1 , u a ˉ ) x=\alpha(1,u\bar a) x = α ( 1 , u a ˉ ) , y = β ( 1 , v b ˉ ) y=\beta(1,v\bar b) y = β ( 1 , v b ˉ ) with α , β > 0 \alpha,\beta>0 α , β > 0 and unit vectors a ˉ , b ˉ \bar a,\bar b a ˉ , b ˉ with a ˉ ⋅ b ˉ = ϑ ∈ [ − 1 , 1 ] \bar a\cdot\bar b=\vartheta\in[-1,1] a ˉ ⋅ b ˉ = ϑ ∈ [ − 1 , 1 ] ; every ϑ \vartheta ϑ occurs since n ≥ 2 n\geq2 n ≥ 2 . From the product in V n V_n V n ,
x ∘ y = α β ( 1 + u v ϑ , u a ˉ + v b ˉ ) , OMC ( x ∘ y ) = ∣ u a ˉ + v b ˉ ∣ 1 + u v ϑ = u 2 + v 2 + 2 u v ϑ 1 + u v ϑ . x\circ y=\alpha\beta\left(1+uv\vartheta,u\bar a+v\bar b\right),\qquad\operatorname{OMC}(x\circ y)=\frac{\lvert u\bar a+v\bar b\rvert}{1+uv\vartheta}=\frac{\sqrt{u^2+v^2+2uv\vartheta}}{1+uv\vartheta}. x ∘ y = α β ( 1 + uv ϑ , u a ˉ + v b ˉ ) , OMC ( x ∘ y ) = 1 + uv ϑ ∣ u a ˉ + v b ˉ ∣ = 1 + uv ϑ u 2 + v 2 + 2 uv ϑ .
If u v = 0 uv=0 uv = 0 this is max ( u , v ) = u ⊕ v \max(u,v)=u\mathbin{\oplus}v max ( u , v ) = u ⊕ v , and u 2 + u v + v 2 < 1 u^2+uv+v^2<1 u 2 + uv + v 2 < 1 . Let u v > 0 uv>0 uv > 0 and put s = 1 + u v ϑ ∈ [ 1 − u v , 1 + u v ] s=1+uv\vartheta\in[1-uv,1+uv] s = 1 + uv ϑ ∈ [ 1 − uv , 1 + uv ] and Q = 2 − u 2 − v 2 Q=2-u^2-v^2 Q = 2 − u 2 − v 2 . Then OMC ( x ∘ y ) 2 = γ ( s ) = ( 2 s − Q ) / s 2 \operatorname{OMC}(x\circ y)^2=\gamma(s)=(2s-Q)/s^2 OMC ( x ∘ y ) 2 = γ ( s ) = ( 2 s − Q ) / s 2 , and γ ′ ( s ) = 2 ( Q − s ) / s 3 \gamma'(s)=2(Q-s)/s^3 γ ′ ( s ) = 2 ( Q − s ) / s 3 , so γ \gamma γ increases on ( 0 , Q ] (0,Q] ( 0 , Q ] and decreases on [ Q , ∞ ) [Q,\infty) [ Q , ∞ ) . Since u 2 + v 2 < 1 u^2+v^2<1 u 2 + v 2 < 1 , Q − ( 1 − u v ) = 1 − u 2 − v 2 + u v > 0 Q-(1-uv)=1-u^2-v^2+uv>0 Q − ( 1 − uv ) = 1 − u 2 − v 2 + uv > 0 , so sup γ = γ ( min { Q , 1 + u v } ) \sup\gamma=\gamma(\min\{Q,1+uv\}) sup γ = γ ( min { Q , 1 + uv }) . If Q ≥ 1 + u v Q\geq1+uv Q ≥ 1 + uv , i.e. u 2 + u v + v 2 ≤ 1 u^2+uv+v^2\leq1 u 2 + uv + v 2 ≤ 1 , this is γ ( 1 + u v ) = ( u ⊕ v ) 2 \gamma(1+uv)=(u\mathbin{\oplus}v)^2 γ ( 1 + uv ) = ( u ⊕ v ) 2 . Otherwise it is γ ( Q ) = 1 / Q \gamma(Q)=1/Q γ ( Q ) = 1/ Q , attained at ϑ = ( Q − 1 ) / ( u v ) ∈ ( − 1 , 1 ) \vartheta=(Q-1)/(uv)\in(-1,1) ϑ = ( Q − 1 ) / ( uv ) ∈ ( − 1 , 1 ) , and 1 / Q = γ ( Q ) > γ ( 1 + u v ) = ( u ⊕ v ) 2 1/Q=\gamma(Q)>\gamma(1+uv)=(u\mathbin{\oplus}v)^2 1/ Q = γ ( Q ) > γ ( 1 + uv ) = ( u ⊕ v ) 2 .
(b) By Lemma 6.8.1, κ ( x ∘ y ) ≤ λ max ( x ) λ max ( y ) μ ( x ) μ ( y ) ℓ ( u , v ) = ( 1 + u ) ( 1 + v ) ℓ ( u , v ) \kappa(x\circ y)\leq\frac{\lambda_{\max}(x)\lambda_{\max}(y)}{\mu(x)\mu(y)}\ell(u,v)=\frac{(1+u)(1+v)}{\ell(u,v)} κ ( x ∘ y ) ≤ μ ( x ) μ ( y ) λ m a x ( x ) λ m a x ( y ) ℓ ( u , v ) = ℓ ( u , v ) ( 1 + u ) ( 1 + v ) . For u + v ≤ 1 u+v\leq1 u + v ≤ 1 this is κ ( x ) κ ( y ) \kappa(x)\kappa(y) κ ( x ) κ ( y ) , i.e. OMC ( x ∘ y ) ≤ u ⊕ v \operatorname{OMC}(x\circ y)\leq u\mathbin{\oplus}v OMC ( x ∘ y ) ≤ u ⊕ v , attained by operator commuting x , y x,y x , y with primitive d ≤ E max ( x ) , E max ( y ) d\leq E_{\max}(x),E_{\max}(y) d ≤ E m a x ( x ) , E m a x ( y ) and d ′ ≤ E min ( x ) , E min ( y ) d'\leq E_{\min}(x),E_{\min}(y) d ′ ≤ E m i n ( x ) , E m i n ( y ) : for operator commuting elements x ∘ y = P ( x 1 / 2 ) y x\circ y=P(x^{1/2})y x ∘ y = P ( x 1/2 ) y , and Theorem 6.3.4 applies. Let u + v > 1 u+v>1 u + v > 1 . The upper bound is κ ∗ \kappa_* κ ∗ as stated, and it exceeds κ ( x ) κ ( y ) \kappa(x)\kappa(y) κ ( x ) κ ( y ) because 2 ( 1 − u ) ( 1 − v ) > 1 − u 2 − v 2 2(1-u)(1-v)>1-u^2-v^2 2 ( 1 − u ) ( 1 − v ) > 1 − u 2 − v 2 . For attainment, V V V contains orthogonal primitive idempotents c 2 , c 3 c_2,c_3 c 2 , c 3 with V 23 ≠ 0 V_{23}\neq0 V 23 = 0 (F6) and a nonzero idempotent c 1 ⊥ c 2 + c 3 c_1\perp c_2+c_3 c 1 ⊥ c 2 + c 3 : take a frame of a simple component of rank ≥ 3 \geq3 ≥ 3 , or c 2 , c 3 c_2,c_3 c 2 , c 3 in a non-associative simple component (of rank ≥ 2 \geq2 ≥ 2 ) and c 1 c_1 c 1 the unit of another component. The Peirce algebra U = V ( c 2 + c 3 , 1 ) U=V(c_2+c_3,1) U = V ( c 2 + c 3 , 1 ) has rank two and contains V 23 ≠ 0 V_{23}\neq0 V 23 = 0 , so it is simple (a non-simple algebra of rank two is R ⊕ R \mathbb{R}\oplus\mathbb{R} R ⊕ R , where V 23 = 0 V_{23}=0 V 23 = 0 ), hence a spin factor of dimension ≥ 3 \geq3 ≥ 3 , and by the proof of (a) there are x ′ , y ′ ∈ U x',y'\in U x ′ , y ′ ∈ U with spectra { 1 ± u } , { 1 ± v } \{1\pm u\},\{1\pm v\} { 1 ± u } , { 1 ± v } and
λ min ( x ′ ∘ y ′ ) = min ϑ ( 1 + u v ϑ − u 2 + v 2 + 2 u v ϑ ) = 1 2 ( 1 − u 2 − v 2 ) , \lambda_{\min}(x'\circ y')=\min_{\vartheta}\left(1+uv\vartheta-\sqrt{u^2+v^2+2uv\vartheta}\right)=\frac{1}{2}(1-u^2-v^2), λ m i n ( x ′ ∘ y ′ ) = ϑ min ( 1 + uv ϑ − u 2 + v 2 + 2 uv ϑ ) = 2 1 ( 1 − u 2 − v 2 ) ,
the minimum being attained where u 2 + v 2 + 2 u v ϑ = 1 u^2+v^2+2uv\vartheta=1 u 2 + v 2 + 2 uv ϑ = 1 , i.e. at ϑ = 1 − u 2 − v 2 2 u v ∈ [ 0 , 1 ] \vartheta=\frac{1-u^2-v^2}{2uv}\in[0,1] ϑ = 2 uv 1 − u 2 − v 2 ∈ [ 0 , 1 ] , which is admissible because ( u + v ) 2 ≥ 1 (u+v)^2\geq1 ( u + v ) 2 ≥ 1 . Put r 0 = e − c 1 − c 2 − c 3 r_0=e-c_1-c_2-c_3 r 0 = e − c 1 − c 2 − c 3 and
x = ( 1 + u ) c 1 + x ′ + r 0 , y = ( 1 + v ) c 1 + y ′ + r 0 . x=(1+u)c_1+x'+r_0,\qquad y=(1+v)c_1+y'+r_0. x = ( 1 + u ) c 1 + x ′ + r 0 , y = ( 1 + v ) c 1 + y ′ + r 0 .
Then OMC ( x ) = u \operatorname{OMC}(x)=u OMC ( x ) = u , OMC ( y ) = v \operatorname{OMC}(y)=v OMC ( y ) = v , and by (F4) x ∘ y = ( 1 + u ) ( 1 + v ) c 1 + x ′ ∘ y ′ + r 0 x\circ y=(1+u)(1+v)c_1+x'\circ y'+r_0 x ∘ y = ( 1 + u ) ( 1 + v ) c 1 + x ′ ∘ y ′ + r 0 , whose spectrum is the union of { ( 1 + u ) ( 1 + v ) } \{(1+u)(1+v)\} {( 1 + u ) ( 1 + v )} , the spectrum of x ′ ∘ y ′ x'\circ y' x ′ ∘ y ′ in U U U , and { 1 } \{1\} { 1 } if r 0 ≠ 0 r_0\neq0 r 0 = 0 . As x ′ ∘ y ′ ≤ ( 1 + u ) ( 1 + v ) ( c 2 + c 3 ) x'\circ y'\leq(1+u)(1+v)(c_2+c_3) x ′ ∘ y ′ ≤ ( 1 + u ) ( 1 + v ) ( c 2 + c 3 ) by Lemma 6.8.1 and 1 2 ( 1 − u 2 − v 2 ) < 1 \frac{1}{2}(1-u^2-v^2)<1 2 1 ( 1 − u 2 − v 2 ) < 1 , we get λ max = ( 1 + u ) ( 1 + v ) \lambda_{\max}=(1+u)(1+v) λ m a x = ( 1 + u ) ( 1 + v ) and λ min = 1 2 ( 1 − u 2 − v 2 ) \lambda_{\min}=\frac{1}{2}(1-u^2-v^2) λ m i n = 2 1 ( 1 − u 2 − v 2 ) , so κ ∗ \kappa_* κ ∗ is attained.
Corollary 6.8.3. On Herm 2 ( F ) \operatorname{Herm}_2(\mathbb{F}) Herm 2 ( F ) , F = R , C , H , O \mathbb{F}=\mathbb{R},\mathbb{C},\mathbb{H},\mathbb{O} F = R , C , H , O (spin factors), the contrast of the Jordan product is subadditive exactly on the ellipse u 2 + u v + v 2 ≤ 1 u^2+uv+v^2\leq1 u 2 + uv + v 2 ≤ 1 ; on Herm n ( F ) \operatorname{Herm}_n(\mathbb{F}) Herm n ( F ) , F = R , C , H \mathbb{F}=\mathbb{R},\mathbb{C},\mathbb{H} F = R , C , H , n ≥ 3 n\geq3 n ≥ 3 , on the Albert algebra Herm 3 ( O ) \operatorname{Herm}_3(\mathbb{O}) Herm 3 ( O ) , and on R ⊕ V n \mathbb{R}\oplus V_n R ⊕ V n , V m ⊕ V n V_m\oplus V_n V m ⊕ V n (n ≥ 2 n\geq2 n ≥ 2 ) exactly on the triangle u + v ≤ 1 u+v\leq1 u + v ≤ 1 . In particular, for positive complex n × n n\times n n × n matrices the threshold changes between n = 2 n=2 n = 2 and n = 3 n=3 n = 3 .
The mechanism is the same as in Theorem 6.4.1 . Strang’s two bounds are attained by different configurations: the upper one by an idempotent under both E max E_{\max} E m a x , the lower one, for u + v > 1 u+v>1 u + v > 1 , by a non-commuting pair in a rank-two Peirce algebra. Fitting both needs a third orthogonal piece, i.e. room beyond a spin factor; in a spin factor the two extremes compete, and the region of subadditivity is larger. The quantity in the proof of (a) is the speed of x ∘ y x\circ y x ∘ y : for four-velocities with speeds u u u , v v v and angle θ \theta θ between them, OMC ( x ∘ y ) = ∣ u a ˉ + v b ˉ ∣ / ( 1 + u v cos θ ) \operatorname{OMC}(x\circ y)=\lvert u\bar a+v\bar b\rvert/(1+uv\cos\theta) OMC ( x ∘ y ) = ∣ u a ˉ + v b ˉ ∣ / ( 1 + uv cos θ ) is the Galilean sum of the velocities divided by the relativistic denominator.
The positivity criterion u 2 + v 2 < 1 u^2+v^2<1 u 2 + v 2 < 1 is [100 Cor. 4.4] and [89 ] , and the bounds of Lemma 6.8.1 for matrices are [89 , 99 ] ; de Bruijn [109 §4] uses a parametrisation close to the one in the proof of (a). The joint attainability of the two extremes is not discussed in [89 , 100 , 108 , 109 , 110 , 99 ] .
Remark 6.8.4 (Multiplicative commutators of spinor boosts). For positive matrices the additive commutator is controlled by Strang’s ∥ X Y − Y X ∥ ≤ 1 2 ( M − m ) ( N − n ) \lVert XY-YX\rVert\leq\frac{1}{2}(M-m)(N-n) ∥ X Y − Y X ∥ ≤ 2 1 ( M − m ) ( N − n ) [89 Thm. II] , attained in dimension two; in terms of contrasts, ∥ X Y − Y X ∥ ≤ 2 OMC ( X ) OMC ( Y ) μ ( X ) μ ( Y ) \lVert XY-YX\rVert\leq2\operatorname{OMC}(X)\operatorname{OMC}(Y)\mu(X)\mu(Y) ∥ X Y − Y X ∥ ≤ 2 OMC ( X ) OMC ( Y ) μ ( X ) μ ( Y ) . The multiplicative commutator K = X Y X − 1 Y − 1 K=XYX^{-1}Y^{-1} K = X Y X − 1 Y − 1 of two positive elements of SL ( 2 , C ) \operatorname{SL}(2,\mathbb{C}) SL ( 2 , C ) , i.e. of two spinor boosts X = exp ( α a ˉ ⋅ σ ⃗ ) X=\exp(\alpha\bar a\cdot\vec{\sigma}) X = exp ( α a ˉ ⋅ σ ) , Y = exp ( β b ˉ ⋅ σ ⃗ ) Y=\exp(\beta\bar b\cdot\vec{\sigma}) Y = exp ( β b ˉ ⋅ σ ) with OLC ( X ) = α \operatorname{OLC}(X)=\alpha OLC ( X ) = α , OLC ( Y ) = β \operatorname{OLC}(Y)=\beta OLC ( Y ) = β (boost rapidities 2 α 2\alpha 2 α , 2 β 2\beta 2 β , Remark 6.7.3 ) and angle θ \theta θ between the axes, satisfies
tr K = 2 − 4 sinh 2 α sinh 2 β sin 2 θ , \operatorname{tr}K=2-4\sinh^2\alpha\sinh^2\beta\sin^2\theta, tr K = 2 − 4 sinh 2 α sinh 2 β sin 2 θ ,
sinh OLC ( K ) = 1 2 sinh 2 α sinh 2 β ∣ sin θ ∣ ∣ tanh α + e i θ tanh β ∣ ; \sinh\operatorname{OLC}(K)=\frac{1}{2}\sinh2\alpha\sinh2\beta\lvert\sin\theta\rvert\lvert\tanh\alpha+e^{i\theta}\tanh\beta\rvert; sinh OLC ( K ) = 2 1 sinh 2 α sinh 2 β ∣ sin θ ∣ ∣ tanh α + e i θ tanh β ∣ ;
both follow from the expansion of K K K in the Pauli basis. Indeed, with ( s 0 + s ˉ ⋅ σ ⃗ ) ( t 0 + t ˉ ⋅ σ ⃗ ) = ( s 0 t 0 + s ˉ ⋅ t ˉ ) + ( s 0 t ˉ + t 0 s ˉ + i s ˉ × t ˉ ) ⋅ σ ⃗ (s_0+\bar s\cdot\vec{\sigma})(t_0+\bar t\cdot\vec{\sigma})=(s_0t_0+\bar s\cdot\bar t)+(s_0\bar t+t_0\bar s+i\bar s\times\bar t)\cdot\vec{\sigma} ( s 0 + s ˉ ⋅ σ ) ( t 0 + t ˉ ⋅ σ ) = ( s 0 t 0 + s ˉ ⋅ t ˉ ) + ( s 0 t ˉ + t 0 s ˉ + i s ˉ × t ˉ ) ⋅ σ for s 0 , t 0 ∈ C s_0,t_0\in\mathbb{C} s 0 , t 0 ∈ C , s ˉ , t ˉ ∈ C 3 \bar s,\bar t\in\mathbb{C}^3 s ˉ , t ˉ ∈ C 3 (bilinear products), and X ± 1 = cosh α ± sinh α a ˉ ⋅ σ ⃗ X^{\pm1}=\cosh\alpha\pm\sinh\alpha\bar a\cdot\vec{\sigma} X ± 1 = cosh α ± sinh α a ˉ ⋅ σ , Y ± 1 = cosh β ± sinh β b ˉ ⋅ σ ⃗ Y^{\pm1}=\cosh\beta\pm\sinh\beta\bar b\cdot\vec{\sigma} Y ± 1 = cosh β ± sinh β b ˉ ⋅ σ , a ˉ ⋅ b ˉ = cos θ \bar a\cdot\bar b=\cos\theta a ˉ ⋅ b ˉ = cos θ , one finds X Y = p 0 + ( r ˉ + i n ˉ ) ⋅ σ ⃗ XY=p_0+(\bar r+i\bar n)\cdot\vec{\sigma} X Y = p 0 + ( r ˉ + i n ˉ ) ⋅ σ and X − 1 Y − 1 = p 0 + ( − r ˉ + i n ˉ ) ⋅ σ ⃗ X^{-1}Y^{-1}=p_0+(-\bar r+i\bar n)\cdot\vec{\sigma} X − 1 Y − 1 = p 0 + ( − r ˉ + i n ˉ ) ⋅ σ , where p 0 = cosh α cosh β + sinh α sinh β cos θ p_0=\cosh\alpha\cosh\beta+\sinh\alpha\sinh\beta\cos\theta p 0 = cosh α cosh β + sinh α sinh β cos θ , r ˉ = sinh α cosh β a ˉ + cosh α sinh β b ˉ \bar r=\sinh\alpha\cosh\beta\bar a+\cosh\alpha\sinh\beta\bar b r ˉ = sinh α cosh β a ˉ + cosh α sinh β b ˉ and n ˉ = sinh α sinh β a ˉ × b ˉ \bar n=\sinh\alpha\sinh\beta\bar a\times\bar b n ˉ = sinh α sinh β a ˉ × b ˉ . Since r ˉ ⊥ n ˉ \bar r\perp\bar n r ˉ ⊥ n ˉ and det ( X Y ) = 1 \det(XY)=1 det ( X Y ) = 1 , p 0 2 = 1 + ∣ r ˉ ∣ 2 − ∣ n ˉ ∣ 2 p_0^2=1+\lvert\bar r\rvert^2-\lvert\bar n\rvert^2 p 0 2 = 1 + ∣ r ˉ ∣ 2 − ∣ n ˉ ∣ 2 , and multiplying out gives K = ( 1 − 2 ∣ n ˉ ∣ 2 ) + ( 2 i p 0 n ˉ − 2 r ˉ × n ˉ ) ⋅ σ ⃗ K=(1-2\lvert\bar n\rvert^2)+(2ip_0\bar n-2\bar r\times\bar n)\cdot\vec{\sigma} K = ( 1 − 2 ∣ n ˉ ∣ 2 ) + ( 2 i p 0 n ˉ − 2 r ˉ × n ˉ ) ⋅ σ . Hence tr K = 2 − 4 ∣ n ˉ ∣ 2 \operatorname{tr}K=2-4\lvert\bar n\rvert^2 tr K = 2 − 4 ∣ n ˉ ∣ 2 with ∣ n ˉ ∣ = sinh α sinh β ∣ sin θ ∣ \lvert\bar n\rvert=\sinh\alpha\sinh\beta\lvert\sin\theta\rvert ∣ n ˉ ∣ = sinh α sinh β ∣ sin θ ∣ ; and since the singular values of K ∈ SL ( 2 , C ) K\in\operatorname{SL}(2,\mathbb{C}) K ∈ SL ( 2 , C ) are e ± OLC ( K ) e^{\pm\operatorname{OLC}(K)} e ± OLC ( K ) , 4 sinh 2 OLC ( K ) = tr ( K ∗ K ) − 2 = 2 ( 1 − 2 ∣ n ˉ ∣ 2 + 8 ∣ n ˉ ∣ 2 p 0 2 + ∣ r ˉ ∣ 2 ) − 2 = 16 ∣ n ˉ ∣ 2 ∣ r ˉ ∣ 2 4\sinh^2\operatorname{OLC}(K)=\operatorname{tr}(K^*K)-2=2(1-2\lvert\bar n\rvert^2+8\lvert\bar n\rvert^2p_0^2+\lvert\bar r\rvert^2)-2=16\lvert\bar n\rvert^2\lvert\bar r\rvert^2 4 sinh 2 OLC ( K ) = tr ( K ∗ K ) − 2 = 2 ( 1 − 2 ∣ n ˉ ∣ 2 + 8 ∣ n ˉ ∣ 2 p 0 2 + ∣ r ˉ ∣ 2 ) − 2 = 16 ∣ n ˉ ∣ 2 ∣ r ˉ ∣ 2 , where ∣ r ˉ ∣ = cosh α cosh β ∣ tanh α + e i θ tanh β ∣ \lvert\bar r\rvert=\cosh\alpha\cosh\beta\lvert\tanh\alpha+e^{i\theta}\tanh\beta\rvert ∣ r ˉ ∣ = cosh α cosh β ∣ tanh α + e i θ tanh β ∣ . The trace formula is the classical trace identity for commutators in SL ( 2 ) \operatorname{SL}(2) SL ( 2 ) (Fricke; see e.g. [101 ] ), specialised to boosts. So for α , β > 0 \alpha,\beta>0 α , β > 0 and θ ∈ ( 0 , π ) \theta\in(0,\pi) θ ∈ ( 0 , π ) the element K K K is elliptic, i.e. conjugate to a non-trivial rotation, iff sinh α sinh β ∣ sin θ ∣ < 1 \sinh\alpha\sinh\beta\lvert\sin\theta\rvert<1 sinh α sinh β ∣ sin θ ∣ < 1 , and K K K has its spectrum on the unit circle for all angles iff sinh α sinh β ≤ 1 \sinh\alpha\sinh\beta\leq1 sinh α sinh β ≤ 1 ; the strict inequality is exactly u 2 + v 2 < 1 u^2+v^2<1 u 2 + v 2 < 1 with u = OMC ( X ) = tanh α u=\operatorname{OMC}(X)=\tanh\alpha u = OMC ( X ) = tanh α , v = OMC ( Y ) = tanh β v=\operatorname{OMC}(Y)=\tanh\beta v = OMC ( Y ) = tanh β , the positivity criterion for X ∘ Y X\circ Y X ∘ Y . This is not a coincidence, and the sufficiency holds for positive invertible X , Y X,Y X , Y in every unital C ∗ C^* C ∗ -algebra: with G = X − 1 Y − 1 + Y − 1 X − 1 G=X^{-1}Y^{-1}+Y^{-1}X^{-1} G = X − 1 Y − 1 + Y − 1 X − 1 and K ∗ = Y − 1 X − 1 Y X K^*=Y^{-1}X^{-1}YX K ∗ = Y − 1 X − 1 Y X one has G K = X − 1 Y − 1 X Y X − 1 Y − 1 + X − 1 Y − 1 GK=X^{-1}Y^{-1}XYX^{-1}Y^{-1}+X^{-1}Y^{-1} G K = X − 1 Y − 1 X Y X − 1 Y − 1 + X − 1 Y − 1 and hence K ∗ G K = X − 1 Y − 1 G K^*GK=X^{-1}Y^{-1}G K ∗ G K = X − 1 Y − 1 G . If u 2 + v 2 < 1 u^2+v^2<1 u 2 + v 2 < 1 , then G = 2 X − 1 ∘ Y − 1 G=2X^{-1}\circ Y^{-1} G = 2 X − 1 ∘ Y − 1 is positive and invertible by Lemma 6.8.1 , applied to X − 1 , Y − 1 X^{-1},Y^{-1} X − 1 , Y − 1 , which have contrasts v , w v,w v , w ; so G 1 / 2 K G − 1 / 2 G^{1/2}KG^{-1/2} G 1/2 K G − 1/2 is an invertible isometry, hence unitary, and the spectrum of K K K lies on the unit circle. For positive n × n n\times n n × n matrices with given contrasts the suprema of OLC ( K ) \operatorname{OLC}(K) OLC ( K ) and of the spectral radius of K K K are non-decreasing in n n n (K K K does not change under scaling of X X X and Y Y Y , so one may assume that 1 1 1 lies in both spectral intervals and pass to X ⊕ 1 X\oplus1 X ⊕ 1 , Y ⊕ 1 Y\oplus1 Y ⊕ 1 ); numerically they are the same for n = 3 , 4 n=3,4 n = 3 , 4 as for n = 2 n=2 n = 2 , so unlike the Jordan product these quantities seem to show no dimension threshold, but we have no proof.
Boosts in representations of SL(2,C)
The non-commuting equality case for 4 × 4 4\times4 4 × 4 positive matrices Theorem 1.3.3 raises the question whether it has a kinematic meaning, for instance in the Dirac representation. It has none, in the following sense.
Let ρ \rho ρ be a finite-dimensional continuous representation of S L ( 2 , C ) \mathrm{SL}(2,\mathbb{C}) SL ( 2 , C ) on a complex vector space W W W (a real representation, such as the vector representation on R 4 \mathbb{R}^{4} R 4 , is complexified; this does not change OLC). Continuous representations are smooth, and the real Lie algebra s l ( 2 , C ) \mathfrak{sl}(2,\mathbb{C}) sl ( 2 , C ) is semisimple, so ρ \rho ρ is completely reducible, with irreducible components of type ( j 1 , j 2 ) (j_{1},j_{2}) ( j 1 , j 2 ) , D j 1 ⊗ D ‾ j 2 D^{j_{1}}\otimes\overline{D}^{j_{2}} D j 1 ⊗ D j 2 , j 1 , j 2 ∈ 1 2 N 0 j_{1},j_{2}\in\frac{1}{2}\mathbb{N}_{0} j 1 , j 2 ∈ 2 1 N 0 [111 Ch. II] ; here D ‾ j ( g ) \overline{D}^{j}(g) D j ( g ) is the entrywise complex conjugate of D j ( g ) D^{j}(g) D j ( g ) in an orthonormal basis (unitarily equivalent to the physicists’ g ↦ D j ( g ∗ − 1 ) g\mapsto D^{j}(g^{*-1}) g ↦ D j ( g ∗− 1 ) ). We call an inner product on W W W admissible if
ρ ( g ∗ ) = ρ ( g ) ∗ ( g ∈ S L ( 2 , C ) ) . \rho(g^{*})=\rho(g)^{*}\qquad(g\in\mathrm{SL}(2,\mathbb{C})). ρ ( g ∗ ) = ρ ( g ) ∗ ( g ∈ SL ( 2 , C )) .
Admissible inner products exist: on D j 1 ⊗ D ‾ j 2 D^{j_{1}}\otimes\overline{D}^{j_{2}} D j 1 ⊗ D j 2 the tensor product of the SU(2)-invariant inner products is admissible, and an orthogonal sum of admissible inner products is admissible. On each irreducible block an admissible inner product is unique up to a positive factor (Schur), and for an admissible inner product invariant subspaces have invariant orthogonal complements. (Mere SU(2)-invariance is not enough: on the vector representation ⟨ x , y ⟩ = x 0 y 0 + 4 x ˉ ⋅ y ˉ \langle x,y\rangle=x_{0}y_{0}+4\bar{x}\cdot\bar{y} ⟨ x , y ⟩ = x 0 y 0 + 4 x ˉ ⋅ y ˉ is rotation-invariant, but boosts are not self-adjoint for it.) For admissible inner products, ρ ( b ) = ρ ( b 1 / 2 ) ∗ ρ ( b 1 / 2 ) \rho(b)=\rho(b^{1/2})^{*}\rho(b^{1/2}) ρ ( b ) = ρ ( b 1/2 ) ∗ ρ ( b 1/2 ) is positive definite for every positive b ∈ S L ( 2 , C ) b\in\mathrm{SL}(2,\mathbb{C}) b ∈ SL ( 2 , C ) , and ρ ( b ) 1 / 2 = ρ ( b 1 / 2 ) \rho(b)^{1/2}=\rho(b^{1/2}) ρ ( b ) 1/2 = ρ ( b 1/2 ) .
Let s ( ρ ) = max k ( j 1 ( k ) + j 2 ( k ) ) s(\rho)=\max_{k}(j_{1}^{(k)}+j_{2}^{(k)}) s ( ρ ) = max k ( j 1 ( k ) + j 2 ( k ) ) over the irreducible components, and, for a unit vector n ˉ ∈ R 3 \bar{n}\in\mathbb{R}^{3} n ˉ ∈ R 3 and η ≥ 0 \eta\geq0 η ≥ 0 , let b n ˉ ( η ) = exp ( η 2 n ˉ ⋅ σ ⃗ ) b_{\bar{n}}(\eta)=\exp(\frac{\eta}{2}\bar{n}\cdot\vec{\sigma}) b n ˉ ( η ) = exp ( 2 η n ˉ ⋅ σ ) , the pure boost of rapidity η \eta η along n ˉ \bar{n} n ˉ (Remark 6.7.3 ).
Lemma 6.9.1. For an admissible inner product, OLC ( ρ ( g ) ) = 2 s ( ρ ) OLC ( g ) \operatorname{OLC}(\rho(g))=2s(\rho)\operatorname{OLC}(g) OLC ( ρ ( g )) = 2 s ( ρ ) OLC ( g ) for every g ∈ S L ( 2 , C ) g\in\mathrm{SL}(2,\mathbb{C}) g ∈ SL ( 2 , C ) .
Proof. Write g = k 1 a k 2 g=k_{1}ak_{2} g = k 1 a k 2 with k 1 , k 2 ∈ S U ( 2 ) k_{1},k_{2}\in\mathrm{SU}(2) k 1 , k 2 ∈ SU ( 2 ) and a = diag ( e t / 2 , e − t / 2 ) a=\operatorname{diag}(e^{t/2},e^{-t/2}) a = diag ( e t /2 , e − t /2 ) , t = 2 OLC ( g ) ≥ 0 t=2\operatorname{OLC}(g)\geq0 t = 2 OLC ( g ) ≥ 0 (the KAK, i.e. singular value, decomposition). The operators ρ ( k i ) \rho(k_{i}) ρ ( k i ) are unitary, since ρ ( k ∗ ) = ρ ( k − 1 ) \rho(k^{*})=\rho(k^{-1}) ρ ( k ∗ ) = ρ ( k − 1 ) , so OLC ( ρ ( g ) ) = OLC ( ρ ( a ) ) \operatorname{OLC}(\rho(g))=\operatorname{OLC}(\rho(a)) OLC ( ρ ( g )) = OLC ( ρ ( a )) . On a component of type ( j 1 , j 2 ) (j_{1},j_{2}) ( j 1 , j 2 ) , ρ ( a ) = exp ( t K ) \rho(a)=\exp(tK) ρ ( a ) = exp ( t K ) with K = J 3 ( 1 ) ⊗ 1 + 1 ⊗ J 3 ( 2 ) K=J_{3}^{(1)}\otimes1+1\otimes J_{3}^{(2)} K = J 3 ( 1 ) ⊗ 1 + 1 ⊗ J 3 ( 2 ) , whose eigenvalues are m 1 + m 2 m_{1}+m_{2} m 1 + m 2 , ∣ m i ∣ ≤ j i |m_{i}|\leq j_{i} ∣ m i ∣ ≤ j i , so the spectrum of ρ ( a ) \rho(a) ρ ( a ) is { e t m } \{e^{tm}\} { e t m } with m m m ranging over a set symmetric about 0 0 0 whose largest element is s ( ρ ) s(\rho) s ( ρ ) . Hence OLC ( ρ ( a ) ) = 1 2 ln ( e t s / e − t s ) = t s = 2 s OLC ( g ) \operatorname{OLC}(\rho(a))=\frac{1}{2}\ln(e^{ts}/e^{-ts})=ts=2s\operatorname{OLC}(g) OLC ( ρ ( a )) = 2 1 ln ( e t s / e − t s ) = t s = 2 s OLC ( g ) . □
Theorem 6.9.2. Let ρ \rho ρ be as above with s ( ρ ) > 0 s(\rho)>0 s ( ρ ) > 0 and an admissible inner product, let η 1 , η 2 > 0 \eta_{1},\eta_{2}>0 η 1 , η 2 > 0 , and put X = ρ ( b n ˉ 1 ( η 1 ) ) X=\rho(b_{\bar{n}_{1}}(\eta_{1})) X = ρ ( b n ˉ 1 ( η 1 )) , Y = ρ ( b n ˉ 2 ( η 2 ) ) Y=\rho(b_{\bar{n}_{2}}(\eta_{2})) Y = ρ ( b n ˉ 2 ( η 2 )) . Then X , Y X,Y X , Y are positive definite, OLC ( X ) = s ( ρ ) η 1 \operatorname{OLC}(X)=s(\rho)\eta_{1} OLC ( X ) = s ( ρ ) η 1 , OLC ( Y ) = s ( ρ ) η 2 \operatorname{OLC}(Y)=s(\rho)\eta_{2} OLC ( Y ) = s ( ρ ) η 2 , and the following are equivalent:
(a) OLC ( X Y ) = OLC ( X ) + OLC ( Y ) \operatorname{OLC}(XY)=\operatorname{OLC}(X)+\operatorname{OLC}(Y) OLC ( X Y ) = OLC ( X ) + OLC ( Y ) ;
(b) OLC ( X 1 / 2 Y X 1 / 2 ) = OLC ( X ) + OLC ( Y ) \operatorname{OLC}(X^{1/2}YX^{1/2})=\operatorname{OLC}(X)+\operatorname{OLC}(Y) OLC ( X 1/2 Y X 1/2 ) = OLC ( X ) + OLC ( Y ) ;
(c) n ˉ 1 = n ˉ 2 \bar{n}_{1}=\bar{n}_{2} n ˉ 1 = n ˉ 2 .
Proof. Positivity was noted above, and OLC ( X ) = 2 s OLC ( b n ˉ 1 ( η 1 ) ) = s η 1 \operatorname{OLC}(X)=2s\operatorname{OLC}(b_{\bar{n}_{1}}(\eta_{1}))=s\eta_{1} OLC ( X ) = 2 s OLC ( b n ˉ 1 ( η 1 )) = s η 1 by Lemma 6.9.1. Since X Y = ρ ( b 1 b 2 ) XY=\rho(b_{1}b_{2}) X Y = ρ ( b 1 b 2 ) and X 1 / 2 Y X 1 / 2 = ρ ( b 1 1 / 2 b 2 b 1 1 / 2 ) X^{1/2}YX^{1/2}=\rho(b_{1}^{1/2}b_{2}b_{1}^{1/2}) X 1/2 Y X 1/2 = ρ ( b 1 1/2 b 2 b 1 1/2 ) , with b i = b n ˉ i ( η i ) b_{i}=b_{\bar{n}_{i}}(\eta_{i}) b i = b n ˉ i ( η i ) , Lemma 6.9.1 reduces (a) and (b) to the same statements for the 2 × 2 2\times2 2 × 2 matrices b 1 , b 2 b_{1},b_{2} b 1 , b 2 , i.e. to ρ = id \rho=\operatorname{id} ρ = id . For positive 2 × 2 2\times2 2 × 2 matrices the equality cases of X Y XY X Y and of X 1 / 2 Y X 1 / 2 X^{1/2}YX^{1/2} X 1/2 Y X 1/2 coincide (Remark after Theorem 6.3.4), and in Herm 2 ( C ) ≅ V 3 \operatorname{Herm}_{2}(\mathbb{C})\cong V_{3} Herm 2 ( C ) ≅ V 3 the latter is Theorem 6.7.2 (c): equality holds iff the space parts of b 1 , b 2 b_{1},b_{2} b 1 , b 2 , which are sinh ( η i / 2 ) n ˉ i \sinh(\eta_{i}/2)\bar{n}_{i} sinh ( η i /2 ) n ˉ i , are non-negatively proportional, i.e. iff n ˉ 1 = n ˉ 2 \bar{n}_{1}=\bar{n}_{2} n ˉ 1 = n ˉ 2 . □
Corollary 6.9.3. In no non-trivial finite-dimensional representation of S L ( 2 , C ) \mathrm{SL}(2,\mathbb{C}) SL ( 2 , C ) with an admissible inner product does a pair of pure boosts give a non-commuting equality case: equality forces equal axes, hence commutation. In particular no non-commuting equality pair of the kind of Theorem 1.3.3 arises from boosts of a Dirac spinor ( 1 2 , 0 ) ⊕ ( 0 , 1 2 ) (\frac{1}{2},0)\oplus(0,\frac{1}{2}) ( 2 1 , 0 ) ⊕ ( 0 , 2 1 ) (for the specific example of [2 ] this is already clear from its spectrum { 4 , 3 , 1 , 1 } \{4,3,1,1\} { 4 , 3 , 1 , 1 } , which is not of the form { a , a , a − 1 , a − 1 } \{a,a,a^{-1},a^{-1}\} { a , a , a − 1 , a − 1 } up to scaling).
What non-commuting equality pairs do realise is orthogonal blocks, i.e. direct sums (not tensor products, cf. Proposition 1.3.4 ). On two spinors, i.e. for S L ( 2 , C ) × S L ( 2 , C ) \mathrm{SL}(2,\mathbb{C}) \times\mathrm{SL}(2,\mathbb{C}) SL ( 2 , C ) × SL ( 2 , C ) acting on C 2 ⊕ C 2 \mathbb{C}^{2} \oplus\mathbb{C}^{2} C 2 ⊕ C 2 , let
X = b z ( 2 ln 2 ) ⊕ b z ( ln 2 ) , Y = b z ( 2 ln 2 ) ⊕ b x ( ln 2 ) . X=b_z(2\ln2)\oplus b_z(\ln2),\qquad Y=b_z(2\ln2)\oplus b_x(\ln2). X = b z ( 2 ln 2 ) ⊕ b z ( ln 2 ) , Y = b z ( 2 ln 2 ) ⊕ b x ( ln 2 ) .
Then λ max = 2 \lambda_{\max}=2 λ m a x = 2 , λ min = 1 2 \lambda_{\min}=\frac{1}{2} λ m i n = 2 1 for both, the first blocks give ∥ X Y ∥ = 4 \lVert XY\rVert=4 ∥ X Y ∥ = 4 and ∥ ( X Y ) − 1 ∥ = 4 \lVert(XY)^{-1}\rVert=4 ∥( X Y ) − 1 ∥ = 4 (the second blocks have norms at most 2 2 2 ), so OLC ( X Y ) = OLC ( X ) + OLC ( Y ) \operatorname{OLC}(XY)=\operatorname{OLC}(X)+\operatorname{OLC}(Y) OLC ( X Y ) = OLC ( X ) + OLC ( Y ) , and likewise for X 1 / 2 Y X 1 / 2 X^{1/2}YX^{1/2} X 1/2 Y X 1/2 , while X Y ≠ Y X XY\ne YX X Y = Y X : the first spinor is boosted collinearly by both, the second along different axes by a smaller rapidity, which does not reach the extremes. The example of Theorem 1.3.3 has the same structure. This is the structure behind Theorem 6.4.1 : the extremes are fixed by idempotents c ≤ E max c\le E_{\max} c ≤ E m a x , c ′ ≤ E min c'\le E_{\min} c ′ ≤ E m i n , and the non-commutation sits in an orthogonal rank-two piece. For non-positive elements equality without commutation is easy even for one spinor: g 1 = B U g_1=BU g 1 = B U , g 2 = U ∗ B g_2=U^{*}B g 2 = U ∗ B with B > 0 B>0 B > 0 , U U U unitary and [ U , B 2 ] ≠ 0 [U,B^{2}]\ne0 [ U , B 2 ] = 0 give g 1 g 2 = B 2 g_1g_2=B^{2} g 1 g 2 = B 2 and OLC ( g 1 g 2 ) = OLC ( g 1 ) + OLC ( g 2 ) \operatorname{OLC}(g_1g_2)=\operatorname{OLC}(g_1)+\operatorname{OLC}(g_2) OLC ( g 1 g 2 ) = OLC ( g 1 ) + OLC ( g 2 ) ; these lie outside the positive setting of Theorem 1.3.3 .
Infinite dimension: JB- and JBW-algebras
Let A A A be a unital JB-algebra [98 ] , A + A_{+} A + its cone of squares, and A + + A_{++} A ++ the set of invertible elements of A + A_{+} A + . For x ∈ A + + x\in A_{++} x ∈ A ++ put λ max ( x ) = max σ ( x ) \lambda_{\max}(x)=\max\sigma(x) λ m a x ( x ) = max σ ( x ) , λ min ( x ) = min σ ( x ) > 0 \lambda_{\min}(x)=\min\sigma(x)>0 λ m i n ( x ) = min σ ( x ) > 0 and OLC ( x ) = 1 2 ln ( λ max ( x ) / λ min ( x ) ) \operatorname{OLC}(x)=\frac{1}{2}\ln(\lambda_{\max}(x)/\lambda_{\min}(x)) OLC ( x ) = 2 1 ln ( λ m a x ( x ) / λ m i n ( x )) .
Proposition 6.10.1. For x , y ∈ A + + x,y\in A_{++} x , y ∈ A ++ , OLC ( P ( x 1 / 2 ) y ) ≤ OLC ( x ) + OLC ( y ) \operatorname{OLC}(P(x^{1/2})y)\le\operatorname{OLC}(x)+\operatorname{OLC}(y) OLC ( P ( x 1/2 ) y ) ≤ OLC ( x ) + OLC ( y ) .
Proof. The proof of Theorem 6.3.3 uses only that P ( a ) P(a) P ( a ) is a positive map [98 Prop. 3.3.6] and that λ min ( z ) e ≤ z ≤ λ max ( z ) e \lambda_{\min}(z)e\le z\le\lambda_{\max}(z)e λ m i n ( z ) e ≤ z ≤ λ m a x ( z ) e for z ∈ A + + z\in A_{++} z ∈ A ++ .
Let now M M M be a JBW-algebra and define its rank rk M ∈ { 1 , 2 , … , ∞ } \operatorname{rk} M\in\{1,2,\ldots,\infty\} rk M ∈ { 1 , 2 , … , ∞ } as the supremum of the cardinalities of families of pairwise orthogonal nonzero projections; for Euclidean Jordan algebras this is the usual rank. Elements x , y x,y x , y operator commute if [ T x , T y ] = 0 [T_x,T_y]=0 [ T x , T y ] = 0 , where T x = x ∘ ⋅ T_x=x\circ\cdot T x = x ∘ ⋅ .
Theorem 6.10.2. For a JBW-algebra M M M the following are equivalent:
(a) for all x , y ∈ M + + x,y\in M_{++} x , y ∈ M ++ , OLC ( P ( x 1 / 2 ) y ) = OLC ( x ) + OLC ( y ) \operatorname{OLC}(P(x^{1/2})y)=\operatorname{OLC}(x)+\operatorname{OLC}(y) OLC ( P ( x 1/2 ) y ) = OLC ( x ) + OLC ( y ) implies that x x x and y y y operator commute;
(b) rk M ≤ 3 \operatorname{rk} M\le3 rk M ≤ 3 , or M M M is associative.
Proof. (b)⇒ \Rightarrow ⇒ (a). If M M M is associative there is nothing to prove. Let rk M ≤ 3 \operatorname{rk} M\le3 rk M ≤ 3 . Then the centre of M M M has at most three minimal projections, and every factor summand has rank at most 3; factors of type II, III and I ∞ \mathrm{I}_{\infty} I ∞ contain infinite orthogonal families of projections, so every summand is of type I n \mathrm{I}_{n} I n , n ≤ 3 n\le3 n ≤ 3 . The JBW-factors of types I 1 \mathrm{I}_{1} I 1 , I 2 \mathrm{I}_{2} I 2 , I 3 \mathrm{I}_{3} I 3 are R \mathbb{R} R , the spin factors V H = R ⊕ H V_{\mathcal{H}}=\mathbb{R}\oplus\mathcal{H} V H = R ⊕ H over real Hilbert spaces H \mathcal{H} H of arbitrary dimension ≥ 2 \ge2 ≥ 2 , and Herm 3 ( F ) \operatorname{Herm}_{3}(\mathbb{F}) Herm 3 ( F ) , F = R , C , H , O \mathbb{F}=\mathbb{R},\mathbb{C},\mathbb{H},\mathbb{O} F = R , C , H , O [98 6.1.8, 6.3.1, Thm. 6.4.1] , [112 , 113 ] ; that a JBW-algebra of finite rank is a finite direct sum of algebras of type I n \mathrm{I}_{n} I n , n n n bounded by the rank, follows from [98 Lemma 5.3.4, Thm. 5.3.5] . So M M M is a direct sum, of total rank at most 3, of such algebras. Every summand carries a faithful associative trace ( tr ( α + v ) = 2 α (\operatorname{tr}(\alpha+v)=2\alpha ( tr ( α + v ) = 2 α on V H V_{\mathcal{H}} V H ), every element has a spectral decomposition over a Jordan frame of at most three minimal projections, the Peirce spaces V i i V_{ii} V ii of such a frame are R c i \mathbb{R}c_i R c i , and the Peirce rules hold. These are exactly the facts used in the proofs of Lemmas 6.3.1 , 6.3.2 , Theorem 6.3.4 and of Theorem 6.4.1 in the case rk V ≤ 3 \operatorname{rk}V\le3 rk V ≤ 3 , together with (F1)–(F8); hence these proofs apply verbatim.
(a)⇒ \Rightarrow ⇒ (b). Let M M M be non-associative with rk M ≥ 4 \operatorname{rk} M\ge4 rk M ≥ 4 . We construct pairwise orthogonal nonzero projections c 1 , c 2 , c 3 , c 4 c_1,c_2,c_3,c_4 c 1 , c 2 , c 3 , c 4 with sum e e e and a symmetry s s s with P ( s ) c 2 = c 3 P(s)c_2=c_3 P ( s ) c 2 = c 3 . Take nonzero
orthogonal projections q 1 , … , q 4 q_1,\ldots,q_4 q 1 , … , q 4 with sum e e e . Case A: some q i q_i q i , say q 1 q_1 q 1 , is not central. Then M 1 / 2 ( q 1 ) = ⨁ j ≠ 1 M 1 / 2 ( q 1 ) ∩ M 1 / 2 ( q j ) ≠ 0 M_{1/2}(q_1)=\bigoplus_{j\ne1}M_{1/2}(q_1)\cap M_{1/2}(q_j)\ne0 M 1/2 ( q 1 ) = ⨁ j = 1 M 1/2 ( q 1 ) ∩ M 1/2 ( q j ) = 0 , say the term with j = 2 j=2 j = 2 . In the JBW-algebra N = U q 1 + q 2 M N=U_{q_1+q_2}M N = U q 1 + q 2 M with unit q 1 + q 2 q_1+q_2 q 1 + q 2 , Theorem 5.2.11 of [98 ] writes q 1 = p 1 + p 2 q_1=p_1+p_2 q 1 = p 1 + p 2 , q 2 = p 1 ′ + p 2 ′ q_2=p'_1+p'_2 q 2 = p 1 ′ + p 2 ′ with p 1 , p 1 ′ p_1,p'_1 p 1 , p 1 ′ exchanged by a symmetry of N N N and c N ( p 2 ) ⊥ c N ( p 2 ′ ) c_N(p_2)\perp c_N(p'_2) c N ( p 2 ) ⊥ c N ( p 2 ′ ) ; if p 1 = 0 p_1=0 p 1 = 0 , then c N ( q 1 ) ⊥ c N ( q 2 ) c_N(q_1)\perp c_N(q_2) c N ( q 1 ) ⊥ c N ( q 2 ) , so q 1 q_1 q 1 would be central in N N N , contradicting M 1 / 2 ( q 1 ) ∩ N ≠ 0 M_{1/2}(q_1)\cap N\ne0 M 1/2 ( q 1 ) ∩ N = 0 . Adding e − q 1 − q 2 e-q_1-q_2 e − q 1 − q 2 to the symmetry, we get nonzero c 2 = p 1 ≤ q 1 c_2=p_1\le q_1 c 2 = p 1 ≤ q 1 , c 3 = p 1 ′ ≤ q 2 c_3=p'_1\le q_2 c 3 = p 1 ′ ≤ q 2 exchanged by a symmetry s s s of M M M , and we take c 1 = q 3 c_1=q_3 c 1 = q 3 , c 4 = q 4 + ( q 1 − c 2 ) + ( q 2 − c 3 ) c_4=q_4+(q_1-c_2)+(q_2-c_3) c 4 = q 4 + ( q 1 − c 2 ) + ( q 2 − c 3 ) . Case B: all q i q_i q i are central. Some summand U q 1 M U_{q_1}M U q 1 M is non-associative, hence contains a projection p p p that is not central in it (a JBW-algebra is generated by its projections [98 4.2.3] , and the centre is associative [98 2.5.3] ); [98, Thm. 5.2.11 applied in U q 1 M U_{q_1}M U q 1 M to p p p and q 1 − p q_1-p q 1 − p gives, as in Case A, nonzero c 2 ≤ p c_2\le p c 2 ≤ p , c 3 ≤ q 1 − p c_3\le q_1-p c 3 ≤ q 1 − p exchanged by a symmetry, and we take c 1 = q 2 c_1=q_2 c 1 = q 2 , c 4 = q 3 + q 4 + ( q 1 − c 2 − c 3 ) c_4=q_3+q_4+(q_1-c_2-c_3) c 4 = q 3 + q 4 + ( q 1 − c 2 − c 3 ) . Put u = 2 { c 2 s c 3 } u=2\{c_2\,s\,c_3\} u = 2 { c 2 s c 3 } (Jordan triple product). The unital Jordan subalgebra generated by s s s and c 2 c_2 c 2 contains c 3 = P ( s ) c 2 c_3=P(s)c_2 c 3 = P ( s ) c 2 and u u u , and is special by the Shirshov–Cohn theorem [98 2.4.14] ; computing there, with c 2 c 3 = 0 c_2c_3=0 c 2 c 3 = 0 and s c 3 s = c 2 sc_3s=c_2 s c 3 s = c 2 , gives u = c 2 s c 3 + c 3 s c 2 u=c_2sc_3+c_3sc_2 u = c 2 s c 3 + c 3 s c 2 , c 2 ∘ u = c 3 ∘ u = 1 2 u c_2\circ u=c_3\circ u=\frac{1}{2}u c 2 ∘ u = c 3 ∘ u = 2 1 u and u 2 = c 2 + c 3 u^2=c_2+c_3 u 2 = c 2 + c 3 . Now let
x = 4 c 1 + 3 c 2 + c 3 + c 4 , y = 4 c 1 + 2 ( c 2 + c 3 ) + u + c 4 . x=4c_1+3c_2+c_3+c_4,\qquad y=4c_1+2(c_2+c_3)+u+c_4. x = 4 c 1 + 3 c 2 + c 3 + c 4 , y = 4 c 1 + 2 ( c 2 + c 3 ) + u + c 4 .
As in the proof of Theorem 6.4.1 , y ∈ M + + y\in M_{++} y ∈ M ++ (the element 2 ( c 2 + c 3 ) + u 2(c_2+c_3)+u 2 ( c 2 + c 3 ) + u has spectrum { 1 , 3 } \{1,3\} { 1 , 3 } in U c 2 + c 3 M U_{c_2+c_3}M U c 2 + c 3 M ), the Peirce rules give P ( x 1 / 2 ) y = 16 c 1 + P ( ( x ′ ) 1 / 2 ) y ′ + c 4 P(x^{1/2})y=16c_1+P((x')^{1/2})y'+c_4 P ( x 1/2 ) y = 16 c 1 + P (( x ′ ) 1/2 ) y ′ + c 4 with x ′ = 3 c 2 + c 3 x'=3c_2+c_3 x ′ = 3 c 2 + c 3 , y ′ = 2 ( c 2 + c 3 ) + u y'=2(c_2+c_3)+u y ′ = 2 ( c 2 + c 3 ) + u , and by Proposition 6.10.1 the spectrum of P ( ( x ′ ) 1 / 2 ) y ′ P((x')^{1/2})y' P (( x ′ ) 1/2 ) y ′ in U c 2 + c 3 M U_{c_2+c_3}M U c 2 + c 3 M lies in [ 1 , 9 ] [1,9] [ 1 , 9 ] ; so λ max = 16 = 4 ⋅ 4 \lambda_{\max}=16=4\cdot4 λ m a x = 16 = 4 ⋅ 4 , λ min = 1 \lambda_{\min}=1 λ m i n = 1 , and equality holds. Finally, a direct computation with the Peirce rules gives T x T y c 2 = 6 c 2 + u T_xT_yc_2=6c_2+u T x T y c 2 = 6 c 2 + u and T y T x c 2 = 6 c 2 + 3 2 u T_yT_xc_2=6c_2+\frac{3}{2}u T y T x c 2 = 6 c 2 + 2 3 u , so [ T x , T y ] c 2 = − 1 2 u ≠ 0 [T_x,T_y]c_2=-\frac{1}{2}u\ne0 [ T x , T y ] c 2 = − 2 1 u = 0 : x x x and y y y do not operator commute. □
Remark 6.10.3. In V H V_{\mathcal H} V H over an infinite-dimensional Hilbert space everything in § 9.8 holds verbatim: the boosts are those of S O + ( 1 , H ) \mathrm{SO}^+(1,\mathcal H) SO + ( 1 , H ) , and equality holds iff the space parts are non-negatively proportional. For algebras of infinite rank, such as B ( H ) s a B(\mathcal H)_{\mathrm{sa}} B ( H ) sa , the extreme points of the spectrum need not be eigenvalues. Thus the finite-dimensional criterion in Theorem 6.3.4 , expressed using common extreme eigenvectors, does not extend to arbitrary infinite-rank algebras without additional hypotheses.
A characterisation of traces
A positive linear functional φ \varphi φ on V V V is a trace if φ ( x ∘ ( y ∘ z ) ) = φ ( ( x ∘ y ) ∘ z ) \varphi(x\circ(y\circ z))=\varphi((x\circ y)\circ z) φ ( x ∘ ( y ∘ z )) = φ (( x ∘ y ) ∘ z ) ; for V = ⨁ k V k V=\bigoplus_k V_k V = ⨁ k V k (simple components) the traces are exactly φ = ∑ k α k tr k \varphi=\sum_k\alpha_k\operatorname{tr}_k φ = ∑ k α k tr k , α k ≥ 0 \alpha_k\ge0 α k ≥ 0 (for simple V V V this is [96 Prop. III.4.1] ; the direct-sum case follows by restricting to components). A symmetry is s ∈ V s\in V s ∈ V with s 2 = e s^2=e s 2 = e ; then P ( s ) P(s) P ( s ) is an automorphism of V V V and P ( s ) e = e P(s)e=e P ( s ) e = e .
Theorem 6.11.1. Let φ ≠ 0 \varphi\ne0 φ = 0 be a positive linear functional on V V V and p > 0 p>0 p > 0 . The following are equivalent:
(a) φ \varphi φ is a trace;
(b) φ ( P ( s ) x ) = φ ( x ) \varphi(P(s)x)=\varphi(x) φ ( P ( s ) x ) = φ ( x ) for all symmetries s s s and all x ∈ V x\in V x ∈ V ;
(c) φ ( a p ) φ ( P ( s ) a − p ) ≥ φ ( e ) 2 \varphi(a^p)\varphi(P(s)a^{-p})\ge\varphi(e)^2 φ ( a p ) φ ( P ( s ) a − p ) ≥ φ ( e ) 2 for all a ∈ Ω a\in\Omega a ∈ Ω and all symmetries s s s .
Proof. (a)⇒ \Rightarrow ⇒ (b): P ( s ) x = 2 s ∘ ( s ∘ x ) − s 2 ∘ x = 2 s ∘ ( s ∘ x ) − x P(s)x=2s\circ(s\circ x)-s^2\circ x=2s\circ(s\circ x)-x P ( s ) x = 2 s ∘ ( s ∘ x ) − s 2 ∘ x = 2 s ∘ ( s ∘ x ) − x , and by associativity of φ \varphi φ , φ ( s ∘ ( s ∘ x ) ) = φ ( s 2 ∘ x ) = φ ( x ) \varphi(s\circ(s\circ x))=\varphi(s^2\circ x)=\varphi(x) φ ( s ∘ ( s ∘ x )) = φ ( s 2 ∘ x ) = φ ( x ) . (b)⇒ \Rightarrow ⇒ (c): φ ( P ( s ) a − p ) = φ ( a − p ) \varphi(P(s)a^{-p})=\varphi(a^{-p}) φ ( P ( s ) a − p ) = φ ( a − p ) , and ( u , v ) ↦ φ ( u ∘ v ) (u,v)\mapsto\varphi(u\circ v) ( u , v ) ↦ φ ( u ∘ v ) is positive semidefinite, so by Cauchy–Schwarz with u = a p / 2 u=a^{p/2} u = a p /2 , v = a − p / 2 v=a^{-p/2} v = a − p /2 , u ∘ v = e u\circ v=e u ∘ v = e : φ ( e ) 2 ≤ φ ( a p ) φ ( a − p ) \varphi(e)^2\le\varphi(a^p)\varphi(a^{-p}) φ ( e ) 2 ≤ φ ( a p ) φ ( a − p ) . (c)⇒ \Rightarrow ⇒ (b): for h ∈ V h\in V h ∈ V and a symmetry s s s let f ( ϵ ) = φ ( e ϵ h ) φ ( P ( s ) e − ϵ h ) − φ ( e ) 2 ≥ 0 f(\epsilon)=\varphi(e^{\epsilon h})\varphi(P(s)e^{-\epsilon h})-\varphi(e)^2\ge0 f ( ϵ ) = φ ( e ϵ h ) φ ( P ( s ) e − ϵ h ) − φ ( e ) 2 ≥ 0 . Then f ( 0 ) = 0 f(0)=0 f ( 0 ) = 0 , so 0 = f ′ ( 0 ) = p φ ( e ) ( φ ( h ) − φ ( P ( s ) h ) ) 0=f'(0)=p\varphi(e)(\varphi(h)-\varphi(P(s)h)) 0 = f ′ ( 0 ) = pφ ( e ) ( φ ( h ) − φ ( P ( s ) h )) , and φ ( e ) > 0 \varphi(e)>0 φ ( e ) > 0 (if φ ( e ) = 0 \varphi(e)=0 φ ( e ) = 0 , then 0 ≤ x ≤ λ max ( x ) e 0 \le x \le\lambda_{\max}(x)e 0 ≤ x ≤ λ m a x ( x ) e gives φ = 0 \varphi= 0 φ = 0 on Ω \Omega Ω , which spans V V V ). (b)⇒ \Rightarrow ⇒ (a): let V k V_k V k be a simple component of rank r k ≥ 2 r_k \ge2 r k ≥ 2 and c i ⊥ c j c_i \perp c_j c i ⊥ c j primitive in V k V_k V k . With u u u from (F6), s = ( e − c i − c j ) + u s = (e-c_i-c_j)+u s = ( e − c i − c j ) + u is a symmetry (( e − c i − c j ) ∘ u = 0 (e-c_i-c_j)\circ u=0 ( e − c i − c j ) ∘ u = 0 by (F4), u 2 = c i + c j u^2=c_i+c_j u 2 = c i + c j ) and P ( s ) c i = P ( u ) c i = c j P(s)c_i=P(u)c_i=c_j P ( s ) c i = P ( u ) c i = c j , the mixed terms vanishing by (F4). Hence φ ( c i ) = φ ( c j ) \varphi(c_i)=\varphi(c_j) φ ( c i ) = φ ( c j ) ; every primitive idempotent of V k V_k V k belongs to a Jordan frame of V k V_k V k , so φ ( c ) = φ ( e k ) / r k \varphi(c)=\varphi(e_k)/r_k φ ( c ) = φ ( e k ) / r k for all primitive c ∈ V k c\in V_k c ∈ V k , and by the spectral theorem φ ∣ V k = φ ( e k ) tr e k tr k \varphi|_{V_k}=\frac{\varphi(e_k)}{\operatorname{tr}e_k}\operatorname{tr}_k φ ∣ V k = tr e k φ ( e k ) tr k . Components of rank 1 are R \mathbb{R} R , where every functional is a multiple of tr \operatorname{tr} tr .
Remark 6.11.2. With s = e s=e s = e only, (c) holds for every positive functional (Cauchy–Schwarz).
Open questions
The equality case of Theorem 6.3.4 in JBW-algebras of infinite rank, such as B ( H ) s a B(\mathcal{H})_{\mathrm{sa}} B ( H ) sa , where the extreme points of the spectrum need not be eigenvalues.
Schatten contrasts in JBW-algebras with a faithful normal tracial state: is the defect at most 1 p ln ( 1 / ε ) \frac{1}{p}\ln(1/\varepsilon) p 1 ln ( 1/ ε ) , with ε \varepsilon ε the infimum of the traces of nonzero projections?
For which other symmetric cones, besides the Lorentz cones, does the Wigner rotation κ \kappa κ of Lemma 6.5.1 have a geometric meaning, e.g. as a holonomy of the Riemannian symmetric space G ( Ω ) / Aut ( V ) G(\Omega)/\operatorname{Aut}(V) G ( Ω ) / Aut ( V ) ?
The analogue of Theorem 6.8.2 for the spectral contrasts C Φ C_{\Phi} C Φ , which needs a defect (already C p ( x 2 ) > 2 C p ( x ) C_p(x^2)>2C_p(x) C p ( x 2 ) > 2 C p ( x ) ); the region for fixed rank ≥ 3 \ge3 ≥ 3 is not known.
Bibliography
Bibliography