Introduction
Uniform retention of ℓ \ell ℓ out of r r r particles from a homogeneous Lorentzian law contracts matrix entropy by the sharp universal factor ℓ / r \ell/r ℓ / r , at every positive field and in every finite matrix dimension. Repeated labels are allowed. For r ≥ 2 r \ge2 r ≥ 2 , every matrix extremizer extends uniquely from its two-particle marginal; its noncommutative structure is confined to labels that never occur together.
For matroids, complete contraction gives nonlinear subgaussian concentration and matrix Chernoff and Bernstein bounds. The equality theory identifies all extremal down-up eigenfunctions and, for the unit-edge basis graph, the spectral separation Spec ( L G ) ⊆ { 0 , 2 } ∪ [ 3 , ∞ ) \operatorname{Spec}(L_G) \subseteq\{0,2\} \cup[3,\infty) Spec ( L G ) ⊆ { 0 , 2 } ∪ [ 3 , ∞ ) . The eigenvalue-two space consists exactly of binary direct-factor contrasts, and the optimal universal complete entropy-decay rate is four.
Contraction, profiles and trace duality form the common entrance (§ 2 –§ 4 ). The concentration bounds and all-field marginal region can then be read directly. For spectral applications, continue through scalar translations to § 9 and § 10 ; the quantum equality route passes through reconstruction and § 7 and § 8 .
Complete contraction
Let f ( z ) = ∑ ∣ α ∣ = r a α z α f(z) = \sum_{|\alpha|=r} a_\alpha z^\alpha f ( z ) = ∑ ∣ α ∣ = r a α z α be a nonzero homogeneous Lorentzian polynomial of degree r ≥ 1 r \ge1 r ≥ 1 , and fix a positive field λ \lambda λ . Its coefficient law is
μ ( α ) = a α λ α f ( λ ) . \mu(\alpha) = \frac{a_\alpha\lambda^\alpha}{f(\lambda)}. μ ( α ) = f ( λ ) a α λ α .
An occupation α \alpha α contains α i \alpha_i α i particles of label i i i . Retain a uniformly chosen set of ℓ \ell ℓ particles, write μ ℓ \mu_\ell μ ℓ for the retained law, and let F ℓ ( β ) = E [ F ( α ) ∣ β ] F_\ell(\beta) = \mathbb{E}[F(\alpha) \mid\beta] F ℓ ( β ) = E [ F ( α ) ∣ β ] under the reference deletion coupling. For a positive semidefinite matrix function put
H μ ( F ) = E μ Tr ( F log F ) − Tr [ ( E μ F ) log ( E μ F ) ] . \mathcal{H}_\mu(F) = \mathbb{E}_\mu\operatorname{Tr}(F \log F) - \operatorname{Tr}[(\mathbb{E}_\mu F)\log(\mathbb{E}_\mu F)]. H μ ( F ) = E μ Tr ( F log F ) − Tr [( E μ F ) log ( E μ F )] .
No trace normalization is imposed.
Theorem 1.1 (Complete particle-deletion contraction). For every 0 ≤ ℓ ≤ r 0 \le\ell\le r 0 ≤ ℓ ≤ r and every positive semidefinite matrix function of any finite size,
H μ ℓ ( F ℓ ) ≤ ℓ r H μ ( F ) . (1) \mathcal{H}_{\mu_\ell}(F_\ell) \le\frac{\ell}{r}\mathcal{H}_\mu(F). \tag*{(1)} H μ ℓ ( F ℓ ) ≤ r ℓ H μ ( F ) . (1) The coefficient is uniform in the positive field and matrix size, and is optimal over the class of Lorentzian laws.
The word complete records the uniformity in matrix size. For a normalized classical–quantum state, the entropy is relative entropy from μ ⊗ ρ Q \mu\otimes\rho_Q μ ⊗ ρ Q , using the actual quantum marginal ρ Q \rho_Q ρ Q . The input classical marginal may differ from μ \mu μ . If it equals μ \mu μ , the entropy is mutual information. Deleting one particle and refilling conditionally gives complete modified logarithmic Sobolev constant at least 1 / r 1/r 1/ r for the generator I − K I-K I − K .
Lieb’s three-matrix inequality turns the Lorentzian Hessian signature into a quadratic entropy inequality. Conditioning on two unmarked particles identifies this defect with the second difference of the deletion profile. Convexity then gives contraction, while the same defects control strictness and equality.
Cryan–Guo–Mousa established scalar deletion contraction and scalar concentration for homogeneous strongly log-concave subset laws [17 ] . The scalar entropy framework also includes [7 , 4 ] ; Lee–Seo give the one-particle occupation version with repeated labels [33 Lemma 3.5 and Remark 3.6] . The quadratic converse is exact: scalar contraction, complete contraction and one positive eigenvalue are equivalent. Infinitesimal tests on coordinate derivatives at one positive field therefore recognize Lorentzianity (Corollary 4.4 ).
Quadratic models for entropy equality
For r ≥ 2 r \ge2 r ≥ 2 , absorb the positive field and normalize f ( 1 ) = 1 f(1) = 1 f ( 1 ) = 1 . Its translation space is
N f = { h : D h f = 0 } = ker ∇ 2 f ( 1 ) . \mathcal{N}_f = \{h : D_h f = 0\} = \ker\nabla^2 f(1). N f = { h : D h f = 0 } = ker ∇ 2 f ( 1 ) .
It determines all scalar extremizers. For matrices, the mean and the one-particle conditional densities also determine every extremizer, by Gibbs reconstruction. The following theorem constructs every equality density from its two-particle marginal.
Theorem 1.2 (Quadratic extension). Restrict to active labels and put W = ∇ 2 f ( 1 ) / ( r ( r − 1 ) ) W = \nabla^2 f(1)/(r(r-1)) W = ∇ 2 f ( 1 ) / ( r ( r − 1 )) and f 2 ( z ) = z T W z f_2(z) = z^{\mathsf{T}}Wz f 2 ( z ) = z T W z . For each finite matrix dimension, two-particle conditional expectation is a bijection between the complete entropy equality densities of f f f and those of f 2 f_2 f 2 , normalized by Tr E F = 1 \operatorname{Tr} EF = 1 Tr E F = 1 . Given a quadratic extremizer E i j E_{ij} E ij , set σ = ∑ i j w i j E i j \sigma= \sum_{ij} w_{ij}E_{ij} σ = ∑ ij w ij E ij . On supp σ \sigma σ , put
G i = ∑ j W i j E i j ( W 1 ) i , B i = σ − 1 / 2 G i σ − 1 / 2 . G_i = \frac{\sum_j W_{ij}E_{ij}}{(W\mathbf{1})_i}, \qquad B_i = \sigma^{-1/2}G_i\sigma^{-1/2}. G i = ( W 1 ) i ∑ j W ij E ij , B i = σ − 1/2 G i σ − 1/2 . Its unique extension is
F α = σ 1 / 2 ( ∏ i B i α i ) σ 1 / 2 . F_\alpha= \sigma^{1/2}\left(\prod_i B_i^{\alpha_i}\right)\sigma^{1/2}. F α = σ 1/2 ( i ∏ B i α i ) σ 1/2 . Factors occurring in the same occupation commute. The extension has H μ ( F ) / r = H W ( E ) / 2 H_\mu(F)/r = H_W(E)/2 H μ ( F ) / r = H W ( E ) /2 . More generally, any two Lorentzian laws with the same normalized W W W have canonically identified equality families, even when their degrees differ.
The finite marginal test and normal form (Theorem 7.1 ) combine Petz recovery, Hayden–Jozsa–Petz–Winter’s Markov structure theorem [27 ] , and Lorentzian translations. For matroid basis laws, the quantum register is independent of the simplified basis; only parallel-class choices carry equality information (Corollary 7.7). When every two distinct active labels can occur together, logarithmic extremizers are simultaneously diagonalizable. In this case, fixing the classical marginal leaves only constant equality encodings, although a singular Hessian makes ℓ / r \ell/r ℓ / r the sharp supremum, approached by qubit encodings. In general, informative fixed-marginal equality exists exactly when two distinct active labels never co-occur, and a qubit suffices.
Power equality for 1 < p < 2 1 < p < 2 1 < p < 2 places the exclusive classes in orthogonal reducing subspaces; those inputs have linear profiles for every trace function (Theorem 8.2 ). For affine matrix translation families, a quartic sum of squares decides exact equality for the whole family (Theorem 8.3 ). This criterion applies to every homogeneous occupation law with the stated translation identities.
Extremal matroid walks and spectral separation
For the spectral statements, let M M M be a loopless rank-r r r matroid with at least two bases, r ≥ 1 r \ge1 r ≥ 1 , and let μ \mu μ be uniform on B ( M ) \mathcal{B}(M) B ( M ) . The down-up walk P d u P_{\mathrm{du}} P du removes one uniformly chosen element and chooses a uniform basis completing the retained face, allowing the removed element to return. Write C C C for the parallel classes. The basis graph G G G joins bases differing by one exchange, and its positive unit-edge Laplacian is
( L G g ) ( B ) = ∑ B ′ ∼ B ( g ( B ) − g ( B ′ ) ) . (L_Gg)(B) = \sum_{B' \sim B}\left(g(B)-g(B')\right). ( L G g ) ( B ) = B ′ ∼ B ∑ ( g ( B ) − g ( B ′ ) ) .
The constant-edge Feder–Mihail walk is P F M = I − p L G P_{\mathrm{FM}} = I - pL_G P FM = I − p L G , where 0 < p ≤ 1 / Δ ( G ) 0 < p \le1/\Delta(G) 0 < p ≤ 1/Δ ( G ) and Δ ( G ) \Delta(G) Δ ( G ) is the maximum degree. We use the gap 1 − λ 2 ( P ) 1-\lambda_2(P) 1 − λ 2 ( P ) , where λ 2 ( P ) \lambda_2(P) λ 2 ( P ) is the second-largest eigenvalue.
Theorem 1.3 (Complete classification of spectral equality). With these normalizations, the following statements hold.
(1) The down-up gap is at least 1 / r 1/r 1/ r , and
ker ( P d u − ( 1 − 1 / r ) I ) = { B ↦ ∑ i ∈ B h i : ∑ i ∈ C h i = 0 for every C } . (2) \ker(P_{\mathrm{du}}-(1-1/r)I)=\left\{B\mapsto\sum_{i\in B}h_i:\sum_{i\in C}h_i=0\text{ for every }C\right\}. \tag*{(2)} ker ( P du − ( 1 − 1/ r ) I ) = { B ↦ i ∈ B ∑ h i : i ∈ C ∑ h i = 0 for every C } . (2) This space has dimension ∑ C ( ∣ C ∣ − 1 ) \sum_C(|C|-1) ∑ C ( ∣ C ∣ − 1 ) . Consequently the gap equals 1 / r 1/r 1/ r if and only if M M M has a nontrivial parallel class.
(2) If M M M has k k k direct summands isomorphic to U 1 , 2 U_{1,2} U 1 , 2 , then
Spec ( L G ) ⊆ { 0 , 2 } ∪ [ 3 , ∞ ) , dim ker ( L G − 2 I ) = k . (3) \operatorname{Spec}(L_G)\subseteq\{0,2\}\cup[3,\infty),\qquad\dim\ker(L_G-2I)=k. \tag*{(3)} Spec ( L G ) ⊆ { 0 , 2 } ∪ [ 3 , ∞ ) , dim ker ( L G − 2 I ) = k . (3) The eigenvalue-two space is spanned by the k k k binary contrasts 1 { a ∈ B } − 1 { b ∈ B } \mathbf{1}_{\{a\in B\}}-\mathbf{1}_{\{b\in B\}} 1 { a ∈ B } − 1 { b ∈ B } on these summands. The Feder–Mihail gap equals 2 p 2p 2 p exactly when k > 0 k>0 k > 0 ; if k = 0 k=0 k = 0 , it is at least 3 p 3p 3 p . The threshold three is sharp.
Part (1) proves more than the existence of a degree-one extremizer asked for in [2 Conjecture 8.1] : it identifies every extremizer. It gives an exact criterion on the original matroid for the structural question in [2, Conjecture 8.2]. Part (2) answers [2 Problem 8.4] . Parallel elements alone suffice for down-up equality; the unit-edge walk singles out binary direct components. The proofs appear in Corollaries 5.8 and 10.11 and Theorem 10.7 .
The extremal-space theorem holds at every proper retained level for all positive-field Lorentzian laws: the modes are precisely the Hessian-null linear counts. A weighted upper-star inequality and parallel-class decomposition give the unit-edge separation; triangles attain the threshold three.
The same decomposition expresses all retained-level matroid walks through contraction minors and shows that scalar equality tilts preserve their spectra simultaneously (Corollary 9.5). The general splitting theorem uses derivative laws, permits repeated particles, and needs no Lorentzian hypothesis (Theorem 9.3 ).
A general matrix Laplace theorem
Let L L L be the positive generator of a finite irreducible reversible chain with invariant law μ \mu μ and rates q ( x , y ) q(x,y) q ( x , y ) . For a Hermitian observable A A A , define
Γ L ( A ) ( x ) = 1 2 ∑ y q ( x , y ) ( A ( x ) − A ( y ) ) 2 . \Gamma_L(A)(x)=\frac{1}{2}\sum_y q(x,y)(A(x)-A(y))^2. Γ L ( A ) ( x ) = 2 1 y ∑ q ( x , y ) ( A ( x ) − A ( y ) ) 2 .
Suppose that α > 0 \alpha>0 α > 0 satisfies the complete entropy-production inequality
α H μ ( F ) ≤ E μ Tr [ ( L F ) log F ] ( F > 0 ) . \alpha H_\mu(F)\leq E_\mu\operatorname{Tr}[(LF)\log F]\qquad(F>0). α H μ ( F ) ≤ E μ Tr [( L F ) log F ] ( F > 0 ) .
Theorem 1.4 (Matrix Laplace comparison). If Γ L ( A ) ( x ) ⪯ V \Gamma_L(A)(x)\preceq V Γ L ( A ) ( x ) ⪯ V for a fixed matrix V ⪰ 0 V\succeq0 V ⪰ 0 and every x x x , then for every Hermitian offset K K K and every real θ \theta θ ,
E μ Tr e K + θ A ≤ Tr exp ( K + θ E μ A + θ 2 α V ) . (4) E_\mu\operatorname{Tr}e^{K+\theta A}\leq\operatorname{Tr}\exp\left(K+\theta E_\mu A+\frac{\theta^2}{\alpha}V\right). \tag*{(4)} E μ Tr e K + θ A ≤ Tr exp ( K + θ E μ A + α θ 2 V ) . (4)
The transport estimate retains Tr ( M V ) \operatorname{Tr}(MV) Tr ( M V ) until the matrix mean M M M is optimized. For a centered observable, an affine variance bound Γ L ( A ) ⪯ a A + B \Gamma_L(A)\preceq aA+B Γ L ( A ) ⪯ a A + B , with a ≥ 0 a\geq0 a ≥ 0 and B ⪰ 0 B\succeq0 B ⪰ 0 , yields a Bernstein comparison with θ 2 B / ( α − a θ ) \theta^2B/(\alpha-a\theta) θ 2 B / ( α − a θ ) for θ < α / a \theta<\alpha/a θ < α / a ; when a = 0 a=0 a = 0 , all real θ \theta θ are allowed (Theorem 11.3 ). For a rank-r r r matroid, the product-field basis law μ ( B ) ∝ ∏ i ∈ B λ i \mu(B)\propto\prod_{i\in B}\lambda_i μ ( B ) ∝ ∏ i ∈ B λ i , λ i > 0 \lambda_i>0 λ i > 0 , is Lorentzian [13 ] . If a d × d d\times d d × d Hermitian function A A A changes by at most a > 0 a>0 a > 0 in operator norm across every basis exchange, then, for t > 0 t>0 t > 0 ,
Pr { λ max ( A − E A ) ≥ t } ≤ d exp ( − t 2 2 r a 2 ) . (5) \Pr\{\lambda_{\max}(A-EA)\geq t\}\leq d\exp\left(-\frac{t^2}{2ra^2}\right). \tag*{(5)} Pr { λ m a x ( A − E A ) ≥ t } ≤ d exp ( − 2 r a 2 t 2 ) . (5)
Componentwise complementation replaces rank by the sum of the smaller rank or corank, with componentwise exchange bounds.
Entropy–trace duality also gives a Laplace comparison for subconfiguration averages (Theorem 12.1 ), yielding matrix Chernoff bounds and Bernstein variance ∑ i Pr ( i ∈ B ) A i 2 − ( E Y ) 2 / r \sum_i \Pr(i \in B)A_i^2 - (\mathbb{E}Y)^2/r ∑ i Pr ( i ∈ B ) A i 2 − ( E Y ) 2 / r for Y = ∑ i ∈ B A i Y = \sum_{i\in B} A_i Y = ∑ i ∈ B A i .
Kathuria explicitly proposed extending nonlinear matrix concentration to strongly log-concave laws and arbitrary matroids, and asked for a matrix modified logarithmic Sobolev inequality and a matrix Herbst argument [29 Section 5] . Complete replacement and the finite-chain comparison give this route for every Lorentzian law; (1.5) gives the proposed all-matroid conclusion.
Dependent matrix concentration was developed for strongly Rayleigh measures [32 , 29 ] , under ℓ ∞ \ell_\infty ℓ ∞ -independence [30 ] , and for nonlinear observables under stochastic covering assumptions [1 ] . Harvey–Olver proved matrix Chernoff bounds for pipage-rounding outputs on matroid base polytopes [26 ] . The distribution in the present results is the product-field basis law itself, for every matroid. These estimates include the Fano basis law, whose generating polynomial fails real stability [16 Example 11.1] . Scalar-to-matrix Poincaré methods already give exponential concentration for Lorentzian measures [23 ] . Transport and concentration from complete logarithmic Sobolev inequalities are developed in [21 , 22 , 18 ] ; Huang–Tropp obtain nonlinear subgaussian matrix concentration from a curvature criterion [28 ] .
Sharp basis-graph inequalities
Complete entropy factors over exchange edges. For the uniform law on the N N N bases of a matroid and the unit-edge Laplacian ( L G g ) ( B ) = ∑ B ′ ∼ B ( g ( B ) − g ( B ′ ) ) (L_Gg)(B)=\sum_{B'\sim B}(g(B)-g(B')) ( L G g ) ( B ) = ∑ B ′ ∼ B ( g ( B ) − g ( B ′ )) , we obtain
∣ ∂ U ∣ ≥ a log 2 ( N / a ) , a = min { ∣ U ∣ , N − ∣ U ∣ } > 0 , |\partial U| \ge a\log_2(N/a), \qquad a=\min\{|U|,N-|U|\}>0, ∣ ∂ U ∣ ≥ a log 2 ( N / a ) , a = min { ∣ U ∣ , N − ∣ U ∣ } > 0 ,
Ent μ ( g 2 ) ≤ ⟨ g , L G g ⟩ μ , H μ ( e − t L G F ) ≤ e − 4 t H μ ( F ) . \operatorname{Ent}_{\mu}(g^2)\le\langle g,L_Gg\rangle_{\mu}, \qquad\mathcal{H}_{\mu}(e^{-tL_G}F)\le e^{-4t}\mathcal{H}_{\mu}(F). Ent μ ( g 2 ) ≤ ⟨ g , L G g ⟩ μ , H μ ( e − t L G F ) ≤ e − 4 t H μ ( F ) .
All three constants are sharp uniformly over matroids. The small-set bound extends the cube’s classical logarithmic edge-isoperimetric inequality [24 , 9 , 25 ] . Montenegro–Son proved this bound for balanced matroids [37 ] . Anari–Liu–Oveis Gharan–Vinzant proved the matroid case of the Mihail–Vazirani conjecture: every matroid basis graph has expansion at least one [5 ] . Here equality sets are exactly cylinders fixing binary direct factors, up to complementation, and the optimal Gross constant is one exactly when such a factor exists. Spectral separation also bounds distance from the binary eigenspace.
Under the full transverse gap (15) , the complete unit-edge rate improves to 4 r ( 1 + κ ) / ( r + ( r − 1 ) κ ) 4r(1+\kappa)/(r+(r-1)\kappa) 4 r ( 1 + κ ) / ( r + ( r − 1 ) κ ) , with corresponding Gross and small-set improvements. Every simple matroid of rank at least two therefore has a guaranteed complete rate greater than four.
The exact marginal region
Suppose a homogeneous multiaffine generating polynomial and its complement are both Lorentzian. The exact region of weights for which the complete marginal inequality holds at every positive field is
W ( M ) = { w ≥ 0 : w i + w j ≤ 1 for distinct i , j in the same nontrivial component of M } . \mathcal{W}(M)=\{w\ge0:w_i+w_j\le1\text{ for distinct }i,j\text{ in the same nontrivial component of }M\}. W ( M ) = { w ≥ 0 : w i + w j ≤ 1 for distinct i , j in the same nontrivial component of M } .
For X i = 1 i ∈ B X_i=\mathbf{1}_{i\in B} X i = 1 i ∈ B , this means
∑ i w i H μ i ( E [ F ∣ X i ] ) ≤ H μ ( F ) \sum_i w_i\mathcal{H}_{\mu_i}\bigl(\mathbb{E}[F\mid X_i]\bigr)\le\mathcal{H}_{\mu}(F) i ∑ w i H μ i ( E [ F ∣ X i ] ) ≤ H μ ( F )
in every matrix dimension. The same region is exact for the matrix Hölder inequality required at every positive field, with an arbitrary Hermitian offset (Theorem 13.3 and Corollary 13.5). It depends only on the component partition, and includes every product-field matroid basis law.
For the half vector, fixed-field sharpness is governed by the common translation directions of the law and its complement. On connected rank-m m m support, a sharp field exists exactly when n = 2 m n=2m n = 2 m and N f \mathcal{N}_f N f contains a vector v v v with no zero coordinate. That translation space is then one-dimensional, and every sharp field rescales ∣ h ∣ |h| ∣ h ∣ by one constant on its positive half and another on its negative half. A single translation already forces complement reciprocity and independent Bernoulli coordinates within both halves (Theorems 14.3 and 14.5 ). The optimal coefficient of a product law is the maximum of its component coefficients. Unless a binary direct factor exists, the half bound is strict on an open dense set of positive fields, uniformly in matrix size (Theorem 14.6 ). The sharp-field locus is unchanged for every trace-power entropy 1 < p ≤ 2 1 < p \le2 1 < p ≤ 2 , with the same complete strictness bound (Theorem 14.1 ).
The matrix equality family has scalar parameters on those sharp components, with noncommuting factors confined to binary direct choices (Theorem 14.7 ). For a connected rank-three sharp law, the Lorentzian Hessian condition becomes the triangle condition for a unitary matrix, giving a complete determinantal classification (Theorem 14.9 ). In every higher rank, stable matching-mixture laws and determinantal sharp laws are incomparable (Proposition 14.10 ). The corresponding local entropy spectrum is explicit and independent of the field parameter; its matrix statement excludes the full eigenvalue-one space of the second variation.
Profiles, equality, and reconstruction
The profiles H k = H μ k ( F k ) H_k = \mathcal{H}_{\mu_k}(F_k) H k = H μ k ( F k ) form exactly the cone with H 0 = 0 H_0 = 0 H 0 = 0 , nonnegative initial slope and discrete convexity. Classical encodings on partition matroids realize the whole cone (Appendix § 16 ); second differences decompose every deletion deficit into nonnegative terms.
If E μ Tr F = 1 \mathbb{E}_{\mu}\operatorname{Tr} F = 1 E μ Tr F = 1 , its classical–quantum state ρ \rho ρ admits a Gibbs reconstruction τ ℓ \tau_\ell τ ℓ at each level 1 ≤ ℓ < r 1 \le\ell< r 1 ≤ ℓ < r , with normalizing mass 0 < Z ℓ ≤ 1 0 < Z_\ell\le1 0 < Z ℓ ≤ 1 and
H r − r ℓ H ℓ = D ( ρ ∥ τ ℓ ) − log Z ℓ , ∥ ρ − τ ℓ ∥ 1 2 ≤ 2 r ℓ ( ℓ r H r − H ℓ ) . H_r - \frac{r}{\ell} H_\ell= D(\rho\|\tau_\ell) - \log Z_\ell,\qquad\|\rho-\tau_\ell\|_1^2 \le\frac{2r}{\ell}\left(\frac{\ell}{r}H_r-H_\ell\right). H r − ℓ r H ℓ = D ( ρ ∥ τ ℓ ) − log Z ℓ , ∥ ρ − τ ℓ ∥ 1 2 ≤ ℓ 2 r ( r ℓ H r − H ℓ ) .
The coefficient is optimal. Support compression treats singular inputs, and trace saturation forces self-consistency. A separate one-particle specialization reconstructs from one conditional matrix per active label, even when the known deficit is at a higher retained level.
§ 15 combines scalar translations with the quantum normal form to classify fixed-field marginal equality.
Complete Lorentzian entropy contraction
Matrix entropy and the three-matrix inequality. Lieb’s trace inequality converts the Lorentzian quadratic signature into a matrix entropy bound. We fix the notation first.
For a probability measure π \pi π on a finite set and positive semidefinite matrices F ( x ) F(x) F ( x ) of one fixed size, write
H π ( F ) = ∑ x π ( x ) Tr [ F ( x ) log F ( x ) ] − Tr [ F ‾ log F ‾ ] , F ‾ = ∑ x π ( x ) F ( x ) . \mathcal{H}_{\pi}(F)=\sum_x\pi(x)\operatorname{Tr}[F(x)\log F(x)]-\operatorname{Tr}[\overline{F}\log\overline{F}],\qquad\overline{F}=\sum_x\pi(x)F(x). H π ( F ) = x ∑ π ( x ) Tr [ F ( x ) log F ( x )] − Tr [ F log F ] , F = x ∑ π ( x ) F ( x ) .
We use 0 log 0 = 0 0\log0=0 0 log 0 = 0 . No normalization of the trace of F ‾ \overline{F} F is needed. When Tr F ‾ = 1 \operatorname{Tr}\overline{F}=1 Tr F = 1 , the classical–quantum state ρ X Q = ∑ x π ( x ) ∣ x ⟩ ⟨ x ∣ ⊗ F ( x ) \rho_{XQ}=\sum_x\pi(x)|x\rangle\langle x|\otimes F(x) ρ X Q = ∑ x π ( x ) ∣ x ⟩ ⟨ x ∣ ⊗ F ( x ) satisfies
H π ( F ) = D ( ρ X Q ∥ π ⊗ ρ Q ) . (6) \mathcal{H}_{\pi}(F)=D(\rho_{XQ}\Vert\pi\otimes\rho_Q). \tag*{(6)} H π ( F ) = D ( ρ X Q ∥ π ⊗ ρ Q ) . (6)
Its quantum marginal is ρ Q = F ‾ \rho_Q=\overline{F} ρ Q = F . Thus the reference on the right uses the actual quantum marginal; the classical marginal of ρ X Q \rho_{XQ} ρ X Q is allowed to differ from π \pi π .
For a positive definite matrix σ \sigma σ , define the Fréchet derivative of the logarithm by
T σ ( A ) = D log ( σ ) [ A ] = ∫ 0 ∞ ( σ + t I ) − 1 A ( σ + t I ) − 1 d t . T_{\sigma}(A)=D\log(\sigma)[A]=\int_0^\infty(\sigma+tI)^{-1}A(\sigma+tI)^{-1}\,dt. T σ ( A ) = D log ( σ ) [ A ] = ∫ 0 ∞ ( σ + t I ) − 1 A ( σ + t I ) − 1 d t .
It is self-adjoint and strictly positive as an operator on the real Hilbert space of Hermitian matrices, and T σ ( σ ) = I T_{\sigma}(\sigma)=I T σ ( σ ) = I . In particular ⟨ A , B ⟩ σ = Tr [ A T σ ( B ) ] \langle A,B\rangle_{\sigma}=\operatorname{Tr}[AT_{\sigma}(B)] ⟨ A , B ⟩ σ = Tr [ A T σ ( B )] is a real inner product. We use Lieb’s three-matrix inequality [34 ] :
Tr exp ( log A − log B + log C ) ≤ Tr [ A T B ( C ) ] ( A , B , C > 0 ) . (7) \operatorname{Tr}\exp(\log A-\log B+\log C)\leq\operatorname{Tr}[AT_B(C)]\qquad(A,B,C>0). \tag*{(7)} Tr exp ( log A − log B + log C ) ≤ Tr [ A T B ( C )] ( A , B , C > 0 ) . (7)
Occupation laws and complete contraction
The quadratic inequality propagates by conditioning on two unmarked particles. Occupation notation includes repeated labels.
For degree r ≥ 2 r\geq2 r ≥ 2 , a homogeneous polynomial with nonnegative coefficients is Lorentzian when its support is M M M -convex and every order-( r − 2 ) (r-2) ( r − 2 ) partial derivative has a Hessian with at most one positive eigenvalue. Nonnegative homogeneous polynomials of degree at most one are Lorentzian. Here M M M -convexity means that, for support points α , β \alpha,\beta α , β and α i > β i \alpha_i>\beta_i α i > β i , some j j j with α j < β j \alpha_j<\beta_j α j < β j has both α − e i + e j \alpha-e_i+e_j α − e i + e j and β + e i − e j \beta+e_i-e_j β + e i − e j in the support. For homogeneous nonnegative polynomials this class is equivalent to strong log-concavity; see [13 Theorems 2.25 and 2.30] .
Let
f ( z ) = ∑ α ∈ N n , ∣ α ∣ = r a α z α ( a α ≥ 0 ) , f ≠ 0 , f(z)=\sum_{\alpha\in\mathbb{N}^n,\ |\alpha|=r}a_\alpha z^\alpha\qquad(a_\alpha\geq0),\qquad f\ne0, f ( z ) = α ∈ N n , ∣ α ∣ = r ∑ a α z α ( a α ≥ 0 ) , f = 0 ,
be Lorentzian and homogeneous of degree r ≥ 1 r\geq1 r ≥ 1 . For a positive field λ \lambda λ , define the occupation law
μ λ ( α ) = a α λ α f ( λ ) . \mu_\lambda(\alpha)=\frac{a_\alpha\lambda^\alpha}{f(\lambda)}. μ λ ( α ) = f ( λ ) a α λ α .
An occupation α \alpha α describes r r r particles, of which α i \alpha_i α i have label i i i . Retaining a uniformly chosen set of ℓ \ell ℓ particles defines the classical channel
D r → ℓ ( β ∣ α ) = ∏ i ( α i β i ) ( r ℓ ) , ∣ β ∣ = ℓ , β ≤ α . (8) D_{r\to\ell}(\beta\mid\alpha)=\frac{\prod_i\binom{\alpha_i}{\beta_i}}{\binom{r}{\ell}},\qquad|\beta|=\ell,\quad\beta\leq\alpha. \tag*{(8)} D r → ℓ ( β ∣ α ) = ( ℓ r ) ∏ i ( β i α i ) , ∣ β ∣ = ℓ , β ≤ α . (8)
Write μ ℓ = μ λ D r → ℓ \mu_\ell=\mu_\lambda D_{r\to\ell} μ ℓ = μ λ D r → ℓ . Every classical channel below acts as the identity on the quantum register.
Theorem 2.1 (Complete Lorentzian downsampling). For every finite quantum register Q Q Q , every classical–quantum state ρ A Q \rho_{AQ} ρ A Q whose classical support is contained in supp μ λ \operatorname{supp}\mu_\lambda supp μ λ , and every 0 ≤ ℓ ≤ r 0\leq\ell\leq r 0 ≤ ℓ ≤ r ,
D ( ( D r → ℓ ⊗ id Q ) ρ A Q ∥ μ ℓ ⊗ ρ Q ) ≤ ℓ r D ( ρ A Q ∥ μ λ ⊗ ρ Q ) . (9) D\bigl((D_{r\to\ell}\otimes\operatorname{id}_Q)\rho_{AQ}\,\big\|\,\mu_\ell\otimes\rho_Q\bigr)\leq\frac{\ell}{r}D\bigl(\rho_{AQ}\,\big\|\,\mu_\lambda\otimes\rho_Q\bigr). \tag*{(9)} D ( ( D r → ℓ ⊗ id Q ) ρ A Q μ ℓ ⊗ ρ Q ) ≤ r ℓ D ( ρ A Q μ λ ⊗ ρ Q ) . (9) Equivalently, the matrix entropy of every positive semidefinite density contracts by ℓ / r \ell/r ℓ / r under conditional expectation onto the retained occupation. The coefficient has no dependence on n n n or $\dim Q. It is the best universal coefficient for the class of Lorentzian laws.
Proof. The retained laws. Absorb the field and normalization into f f f , so that f ( 1 ) = 1 f(1)=1 f ( 1 ) = 1 and its coefficients are the probabilities of μ λ \mu_\lambda μ λ . Positive diagonal substitutions preserve the Lorentzian property. With D 1 = ∑ i ∂ i D_1=\sum_i \partial_i D 1 = ∑ i ∂ i , the generating polynomial of the level-k k k law is
f k = k ! r ! D 1 r − k f . (10) f_k=\frac{k!}{r!}D_1^{r-k}f. \tag*{(10)} f k = r ! k ! D 1 r − k f . (10)
Indeed the coefficient of z β z^\beta z β in this polynomial is ∑ α ≥ β μ λ ( α ) ∏ i ( α i β i ) / ( r k ) \sum_{\alpha\ge\beta}\mu_\lambda(\alpha)\prod_i\binom{\alpha_i}{\beta_i}/\binom{r}{k} ∑ α ≥ β μ λ ( α ) ∏ i ( β i α i ) / ( k r ) . Every f k f_k f k is normalized and Lorentzian. The closure under nonnegative directional derivatives used here is [13 Corollary 2.11] .
For a positive semidefinite matrix function F F F on level r r r , let F k ( β ) = E [ F ( A ) ∣ B k = β ] F_k(\beta)=\mathbb{E}[F(A)\mid B_k=\beta] F k ( β ) = E [ F ( A ) ∣ B k = β ] under the base law and uniform deletion. All its means equal a fixed matrix M M M . Put
H k ( F k ) = E μ k Tr ( F k log F k ) − Tr ( M log M ) . H_k(F_k)=\mathbb{E}_{\mu_k}\operatorname{Tr}(F_k\log F_k)-\operatorname{Tr}(M\log M). H k ( F k ) = E μ k Tr ( F k log F k ) − Tr ( M log M ) .
We prove that ( H k ) k = 0 r (H_k)_{k=0}^r ( H k ) k = 0 r is convex. Since H 0 = 0 H_0=0 H 0 = 0 , its endpoint chord is the required contraction inequality.
The quadratic step. At level two, give the two remaining particles a uniform random order ( U , V ) (U,V) ( U , V ) . The symmetric ordered-pair law W W W has entries
W i i = μ 2 ( 2 e i ) , W i j = 1 2 μ 2 ( e i + e j ) ( i ≠ j ) . W_{ii}=\mu_2(2e_i),\qquad W_{ij}=\frac{1}{2}\mu_2(e_i+e_j)\quad(i\ne j). W ii = μ 2 ( 2 e i ) , W ij = 2 1 μ 2 ( e i + e j ) ( i = j ) .
Thus 2 W = ∇ 2 f 2 2W=\nabla^2 f_2 2 W = ∇ 2 f 2 and W W W has at most one positive eigenvalue. Its row sums p i = ∑ j W i j p_i=\sum_jW_{ij} p i = ∑ j W ij form the one-particle law. Write F i j = F 2 ( e i + e j ) F_{ij}=F_2(e_i+e_j) F ij = F 2 ( e i + e j ) . Set G i = E [ F 2 ∣ U = i ] G_i=\mathbb{E}[F_2\mid U=i] G i = E [ F 2 ∣ U = i ] and M = ∑ i p i G i M=\sum_ip_iG_i M = ∑ i p i G i , and regularize to make the matrix values positive definite. Conditional expectation rewrites the entropy difference as
H 2 ( F 2 ) − 2 H 1 ( G ) = ∑ i , j W i j Tr F i j ( log F i j − log G i − log G j + log M ) . H_2(F_2)-2H_1(G)=\sum_{i,j}W_{ij}\operatorname{Tr}F_{ij}(\log F_{ij}-\log G_i-\log G_j+\log M). H 2 ( F 2 ) − 2 H 1 ( G ) = i , j ∑ W ij Tr F ij ( log F ij − log G i − log G j + log M ) .
Apply Klein’s inequality with reference exp ( log G i + log G j − log M ) \exp(\log G_i+\log G_j-\log M) exp ( log G i + log G j − log M ) , then Lieb’s inequality (7) . These two steps replace the noncommuting logarithms by the quadratic form determined by W W W :
H 2 ( F 2 ) − 2 H 1 ( G ) ≥ Tr M − ∑ i , j W i j Tr exp ( log G i + log G j − log M ) ≥ Tr M − ∑ i , j W i j Tr G i T M ( G j ) ≥ 0. \begin{aligned}
H_2(F_2)-2H_1(G)\ge\operatorname{Tr}M-\sum_{i,j}W_{ij}\operatorname{Tr}\exp(\log G_i+\log G_j-\log M) \\
&\ge\operatorname{Tr}M-\sum_{i,j}W_{ij}\operatorname{Tr}G_i\mathcal{T}_M(G_j)\ge0.
\end{aligned} H 2 ( F 2 ) − 2 H 1 ( G ) ≥ Tr M − i , j ∑ W ij Tr exp ( log G i + log G j − log M ) ≥ Tr M − i , j ∑ W ij Tr G i T M ( G j ) ≥ 0.
Here T M \mathcal{T}_M T M is the logarithmic derivative defined above. To justify the last inequality, the form ⟨ A , B ⟩ M = Tr A T M ( B ) \langle A,B\rangle_M=\operatorname{Tr}A\mathcal{T}_M(B) ⟨ A , B ⟩ M = Tr A T M ( B ) is a positive definite real inner product on Hermitian matrices. The one-positive-eigenvalue condition and 1 T W 1 = 1 \mathbf{1}^{T}W\mathbf{1}=1 1 T W 1 = 1 give c T W c ≤ 0 c^{T}Wc\le0 c T W c ≤ 0 whenever p T c = 0 p^{T}c=0 p T c = 0 : otherwise the span of 1 \mathbf{1} 1 and c c c would be a two-dimensional positive subspace. Expand G i − M G_i-M G i − M in an orthonormal basis for this inner product. Every coefficient vector is p p p -centered. All cross terms with M M M vanish, while T M ( M ) = I \mathcal{T}_M(M)=I T M ( M ) = I , proving the assertion. Positive regularization followed by continuity covers singular matrix values. This proves the base case with repeated labels included.
Two-particle links and convexity. Fix 1 ≤ j < r 1\le j<r 1 ≤ j < r and mark j − 1 j-1 j − 1 uniformly chosen particles in a level-( j + 1 ) (j+1) ( j + 1 ) occupation. If their occupation is γ \gamma γ , the two unmarked particles have normalized generating polynomial
f γ ( z ) = ∂ γ f j + 1 ( z ) ∂ γ f j + 1 ( 1 ) . f^\gamma(z)=\frac{\partial^\gamma f_{j+1}(z)}{\partial^\gamma f_{j+1}(1)}. f γ ( z ) = ∂ γ f j + 1 ( 1 ) ∂ γ f j + 1 ( z ) .
It is Lorentzian. Give this link the matrix value F j + 1 ( γ + η ) F_{j+1}(\gamma+\eta) F j + 1 ( γ + η ) . Its mean is F j − 1 ( γ ) F_{j-1}(\gamma) F j − 1 ( γ ) ; after retaining one unmarked particle, its conditional mean is F j ( γ + e i ) F_j(\gamma+e_i) F j ( γ + e i ) . The latter assertion includes repeated labels: the identity
( α γ ) ( α i − γ i ) = ( γ i + 1 ) ( α γ + e i ) (11) \binom{\alpha}{\gamma}(\alpha_i-\gamma_i)=(\gamma_i+1)\binom{\alpha}{\gamma+e_i} \tag*{(11)} ( γ α ) ( α i − γ i ) = ( γ i + 1 ) ( γ + e i α ) (11)
shows that, once γ + e i \gamma+ e_i γ + e i is fixed, remembering the mark introduces a factor independent of the original occupation.
Write H 2 γ , H 1 γ H_2^\gamma, H_1^\gamma H 2 γ , H 1 γ for the link entropies, relative to their common mean. The matrix entropy chain rule gives
E γ H 2 γ = H j + 1 − H j − 1 , E γ H 1 γ = H j − H j − 1 . \mathbb{E}_\gamma H_2^\gamma= H_{j+1} - H_{j-1}, \qquad\mathbb{E}_\gamma H_1^\gamma= H_j - H_{j-1}. E γ H 2 γ = H j + 1 − H j − 1 , E γ H 1 γ = H j − H j − 1 .
The quadratic step therefore proves the exact identity
H j + 1 − 2 H j + H j − 1 = E γ ( H 2 γ − 2 H 1 γ ) ≥ 0. (12) H_{j+1} - 2H_j + H_{j-1} = \mathbb{E}_\gamma(H_2^\gamma- 2H_1^\gamma) \ge0. \tag*{(12)} H j + 1 − 2 H j + H j − 1 = E γ ( H 2 γ − 2 H 1 γ ) ≥ 0. (12)
Thus the profile is convex and H ℓ ≤ ( ℓ / r ) H r H_\ell\le(\ell/r)H_r H ℓ ≤ ( ℓ / r ) H r .
Finally, write ρ A Q = ∑ α μ λ ( α ) ∣ α ⟩ ⟨ α ∣ ⊗ F ( α ) \rho_{AQ} = \sum_\alpha\mu_\lambda(\alpha)|\alpha\rangle\langle\alpha| \otimes F(\alpha) ρ A Q = ∑ α μ λ ( α ) ∣ α ⟩ ⟨ α ∣ ⊗ F ( α ) . Then M = ρ Q M = \rho_Q M = ρ Q and H r H_r H r is its relative entropy from μ λ ⊗ ρ Q \mu_\lambda\otimes\rho_Q μ λ ⊗ ρ Q . The output density is F ℓ F_\ell F ℓ , so the matrix entropy inequality is exactly (9) .
Sharpness. Take the product polynomial ∏ j = 1 r ( z j , 0 + z j , 1 ) \prod_{j=1}^r(z_{j,0}+z_{j,1}) ∏ j = 1 r ( z j , 0 + z j , 1 ) , with its uniform product law, and a nonconstant positive scalar density of mean one depending only on the choice in the first pair. The retained occupation reveals that choice with probability ℓ / r \ell/r ℓ / r ; otherwise its conditional density is one. Thus its entropy is exactly ℓ / r \ell/r ℓ / r times the input entropy. □
Corollary 2.2 (Complete replacement contraction). Delete one uniformly chosen particle and then draw its replacement from the conditional law induced by μ λ \mu_\lambda μ λ and that deletion kernel. If K K K is this reversible heat-bath kernel, then
D ( ( K ∗ ⊗ id Q ) ρ ∥ μ λ ⊗ ρ Q ) ≤ ( 1 − r − 1 ) D ( ρ ∥ μ λ ⊗ ρ Q ) . D\bigl((K_* \otimes\operatorname{id}_Q)\rho\,\|\, \mu_\lambda\otimes\rho_Q\bigr) \le(1-r^{-1})D(\rho\,\|\, \mu_\lambda\otimes\rho_Q). D ( ( K ∗ ⊗ id Q ) ρ ∥ μ λ ⊗ ρ Q ) ≤ ( 1 − r − 1 ) D ( ρ ∥ μ λ ⊗ ρ Q ) . For r = 1 r=1 r = 1 the replacement draws a fresh particle and the left side is zero. With generator I − K I-K I − K , continuous time gives decay at least e − t / r e^{-t/r} e − t / r in this relative entropy. Equivalently, with the convention Ent μ F ≤ c − 1 E μ Tr [ ( I − K ) F log F ] \operatorname{Ent}_{\mu}F \le c^{-1}\mathbb{E}_{\mu}\operatorname{Tr}[(I-K)F\log F] Ent μ F ≤ c − 1 E μ Tr [( I − K ) F log F ] , the complete modified logarithmic Sobolev constant is at least 1 / r 1/r 1/ r .
Proof. Apply Theorem 2.1 with ℓ = r − 1 \ell=r-1 ℓ = r − 1 , then quantum data processing for the conditional replacement channel. Its reference output is μ λ \mu_\lambda μ λ , and its quantum marginal remains ρ Q \rho_Q ρ Q . Convexity gives H ( F ) − H ( K F ) ≤ E Tr [ ( I − K ) F log F ] H(F)-H(KF) \le\mathbb{E}\operatorname{Tr}[(I-K)F\log F] H ( F ) − H ( K F ) ≤ E Tr [( I − K ) F log F ] . Differentiation along the semigroup and this inequality prove the continuous-time assertion. □
Entropy profiles and strict contraction
The marked-link identity controls every chord of the entropy profile. Its second differences give the exact deficits and equality conditions. Appendix § 16 proves that their nonnegativity gives all universal profile constraints.
Put H k = H μ k ( F k ) H_k = H_{\mu_k}(F_k) H k = H μ k ( F k ) and H 0 = 0 H_0=0 H 0 = 0 .
Theorem 3.1 (Convexity and the exact deficit decomposition). For every positive semidefinite matrix input,
d j : = H j + 1 − 2 H j + H j − 1 ≥ 0 ( 1 ≤ j < r ) . d_j := H_{j+1} - 2H_j + H_{j-1} \ge0 \qquad(1 \le j < r). d j := H j + 1 − 2 H j + H j − 1 ≥ 0 ( 1 ≤ j < r ) . Each d j d_j d j is the average complete quadratic entropy defect after marking j − 1 j-1 j − 1 particles in the level-( j + 1 ) (j+1) ( j + 1 ) reference law. Consequently
H ℓ − H j ≤ ℓ − j r − j ( H r − H j ) ( 0 ≤ j < ℓ < r ) . (13) H_\ell- H_j \le\frac{\ell-j}{r-j}(H_r-H_j) \qquad(0 \le j < \ell< r). \tag*{(13)} H ℓ − H j ≤ r − j ℓ − j ( H r − H j ) ( 0 ≤ j < ℓ < r ) . (13) and
ℓ r H r − H ℓ = ∑ j = 1 r − 1 min ( ℓ , j ) [ r − max ( ℓ , j ) ] r d j . (14) \frac{\ell}{r}H_r-H_\ell
= \sum_{j=1}^{r-1}\frac{\min(\ell,j)[r-\max(\ell,j)]}{r}\,d_j. \tag*{(14)} r ℓ H r − H ℓ = j = 1 ∑ r − 1 r min ( ℓ , j ) [ r − max ( ℓ , j )] d j . (14) Equality in deletion at one proper retained level is equivalent to H k = k H 1 H_k=kH_1 H k = k H 1 for every k k k . More generally, equality in any chord (3.1) is equivalent to the profile being affine on the whole interval { j , … , r } \{j,\ldots,r\} { j , … , r } . It is also equivalent to equality in deletion from r − j r-j r − j to ℓ − j \ell-j ℓ − j in every positive-mass link obtained by marking j j j particles.
Proof. The proof of Theorem 2.1 established convexity and its link interpretation in (12) . Summing the second differences gives
H k = k H 1 + ∑ j = 1 k − 1 ( k − j ) d j . H_k = kH_1 + \sum_{j=1}^{k-1}(k-j)d_j. H k = k H 1 + j = 1 ∑ k − 1 ( k − j ) d j .
Substitution proves (14) . Applying the same identity to the shifted profile ( H j + k − H j ) k = 0 r − j (H_{j+k}-H_j)_{k=0}^{r-j} ( H j + k − H j ) k = 0 r − j proves the chord bound; its deficit is
∑ t = j + 1 r [ min ( ℓ , t ) − j ] [ r − max ( ℓ , t ) ] r − j d t . \sum_{t=j+1}^{r} \frac{[\min(\ell,t)-j][r-\max(\ell,t)]}{r-j}d_t. t = j + 1 ∑ r r − j [ min ( ℓ , t ) − j ] [ r − max ( ℓ , t )] d t .
All coefficients are positive. This proves both affine-profile characterizations.
Finally, mark j j j particles at level r r r . The chain rule expresses H r − H j H_r-H_j H r − H j and H ℓ − H j H_\ell-H_j H ℓ − H j as the averages of the input and output entropies in the resulting links. The required conditional means follow from the general identity
( α γ ) ( α − γ η ) = ( α γ + η ) ( γ + η γ ) . \binom{\alpha}{\gamma}\binom{\alpha-\gamma}{\eta}
=
\binom{\alpha}{\gamma+\eta}\binom{\gamma+\eta}{\gamma}. ( γ α ) ( η α − γ ) = ( γ + η α ) ( γ γ + η ) .
where γ \gamma γ is marked and η \eta η is retained outside the mark. Their deletion deficits are nonnegative and average to the chord deficit. The average vanishes exactly when every positive-mass link has zero deficit.
Corollary 3.2 (Information after partial disclosure). Suppose the classical input marginal equals the reference law μ \mu μ . For nested uniform retentions A r → A ℓ → A j A_r \to A_\ell\to A_j A r → A ℓ → A j ,
I ( A ℓ : Q ∣ A j ) ≤ ℓ − j r − j I ( A r : Q ∣ A j ) ( 0 ≤ j < ℓ < r ) . I(A_\ell:Q\mid A_j)\leq\frac{\ell-j}{r-j}I(A_r:Q\mid A_j)\qquad(0\leq j<\ell<r). I ( A ℓ : Q ∣ A j ) ≤ r − j ℓ − j I ( A r : Q ∣ A j ) ( 0 ≤ j < ℓ < r ) .
In particular, H j = 0 H_j=0 H j = 0 implies H ℓ ≤ ( ℓ − j ) H r / ( r − j ) H_\ell\leq(\ell-j)H_r/(r-j) H ℓ ≤ ( ℓ − j ) H r / ( r − j ) .
Proof. Here H k = I ( A k : Q ) H_k=I(A_k:Q) H k = I ( A k : Q ) . The Markov property gives I ( A k : Q ∣ A j ) = H k − H j I(A_k:Q\mid A_j)=H_k-H_j I ( A k : Q ∣ A j ) = H k − H j for k ≥ j k\geq j k ≥ j ; apply (13) .
A quadratic criterion and quantitative strictness
The quadratic signature is the exact local criterion: scalar contraction, complete contraction, and one positive eigenvalue are equivalent. Its strict version controls the entire deletion chain. Only the unconditioned two-particle law needs a quantitative gap; the conditioned links retain the qualitative Lorentzian hypothesis. All transition eigenvalues below are real by reversibility.
Theorem 3.3 (Quadratic entropy criterion). Let W W W be a symmetric entrywise nonnegative matrix with 1 T W 1 = 1 \mathbf{1}^{T}W\mathbf{1}=1 1 T W 1 = 1 , and remove its zero rows. Put p = W 1 p=W\mathbf{1} p = W 1 . Regard W W W as the law of an ordered pair ( U , V ) (U,V) ( U , V ) , and let F i j = F j i ⪰ 0 F_{ij}=F_{ji}\succeq0 F ij = F j i ⪰ 0 be a matrix function of its unordered occupation. The following conditions are equivalent.
(1) W W W has exactly one positive eigenvalue.
(2) For scalar F ≥ 0 F\geq0 F ≥ 0 , H p ( E [ F ∣ U ] ) ≤ 1 2 H W ( F ) \mathcal{H}_p(\mathbb{E}[F\mid U])\leq\frac{1}{2}\mathcal{H}_W(F) H p ( E [ F ∣ U ]) ≤ 2 1 H W ( F ) .
(3) The inequality in (2) holds in every finite matrix dimension.
Thus a scalar signature condition is sufficient for the complete entropy inequality, with the same constant.
Proof. The proof of Theorem 2.1 gives (1)⇒ \Rightarrow ⇒ (3), and (3)⇒ \Rightarrow ⇒ (2) is immediate. For necessity, put P = diag ( p ) − 1 W P=\operatorname{diag}(p)^{-1}W P = diag ( p ) − 1 W . This stochastic matrix is self-adjoint on L 2 ( p ) L^2(p) L 2 ( p ) and has spectrum in [ − 1 , 1 ] [-1,1] [ − 1 , 1 ] . If W W W has a second positive eigenvalue, the restriction of P P P to the orthogonal complement of the constants has an eigenvector h h h with eigenvalue λ > 0 \lambda>0 λ > 0 . For sufficiently small real ε \varepsilon ε , set F i j = 1 + ε ( h i + h j ) F_{ij}=1+\varepsilon(h_i+h_j) F ij = 1 + ε ( h i + h j ) . Its mean is one and E [ F ∣ U = i ] = 1 + ε ( 1 + λ ) h i \mathbb{E}[F\mid U=i]=1+\varepsilon(1+\lambda)h_i E [ F ∣ U = i ] = 1 + ε ( 1 + λ ) h i . Expansion at the constant density gives
H W ( F ) = ε 2 ( 1 + λ ) ∥ h ∥ L 2 ( p ) 2 + O ( ε 3 ) , 2 H p ( E [ F ∣ U ] ) = ε 2 ( 1 + λ ) 2 ∥ h ∥ L 2 ( p ) 2 + O ( ε 3 ) . \mathcal{H}_W(F)=\varepsilon^2(1+\lambda)\lVert h\rVert_{L^2(p)}^2+O(\varepsilon^3),\qquad
2\mathcal{H}_p(\mathbb{E}[F\mid U])=\varepsilon^2(1+\lambda)^2\lVert h\rVert_{L^2(p)}^2+O(\varepsilon^3). H W ( F ) = ε 2 ( 1 + λ ) ∥ h ∥ L 2 ( p ) 2 + O ( ε 3 ) , 2 H p ( E [ F ∣ U ]) = ε 2 ( 1 + λ ) 2 ∥ h ∥ L 2 ( p ) 2 + O ( ε 3 ) .
These contradict (2). Finally 1 T W 1 = 1 \mathbf{1}^{T}W\mathbf{1}=1 1 T W 1 = 1 ensures that at least one eigenvalue is positive.
For a normalized degree-r r r Lorentzian polynomial with r ≥ 2 r \ge2 r ≥ 2 , let W W W be the ordered two-particle law after uniform deletion, and let p = W 1 p = W\mathbf{1} p = W 1 . Write κ ≥ 0 \kappa\ge0 κ ≥ 0 for any number satisfying
h T W h ≤ − κ ∑ i p i h i 2 whenever ∑ i p i h i = 0. (15) h^{\mathsf T}Wh \le-\kappa\sum_i p_i h_i^2 \qquad\text{whenever } \sum_i p_i h_i = 0. \tag*{(15)} h T W h ≤ − κ i ∑ p i h i 2 whenever i ∑ p i h i = 0. (15)
Equivalently, the nonconstant spectrum of diag ( p ) − 1 W \operatorname{diag}(p)^{-1}W diag ( p ) − 1 W is bounded above by − κ -\kappa − κ . The value κ = 0 \kappa= 0 κ = 0 is always available. For k ≥ 1 k \ge1 k ≥ 1 put
a k = k + ( k − 1 ) κ . a_k = k + (k-1)\kappa. a k = k + ( k − 1 ) κ .
Theorem 3.4 (The two-particle gap controls every deletion level). Under (3.3), in every finite matrix dimension,
H μ ℓ ( F ℓ ) ≤ a ℓ a r H μ r ( F ) , 1 ≤ ℓ ≤ r . (16) \mathcal{H}_{\mu_\ell}(F_\ell) \le\frac{a_\ell}{a_r}\mathcal{H}_{\mu_r}(F), \qquad1 \le\ell\le r. \tag*{(16)} H μ ℓ ( F ℓ ) ≤ a r a ℓ H μ r ( F ) , 1 ≤ ℓ ≤ r . (16) For deletion of one particle followed by conditional replacement, the complete modified logarithmic Sobolev constant is at least
1 + κ r + ( r − 1 ) κ . \frac{1+\kappa}{r+(r-1)\kappa}. r + ( r − 1 ) κ 1 + κ . The theorem uses only the gap of the unmarked two-particle law; no quantitative bound on the conditioned links is required.
Proof. We first improve the rank-two estimate. Put G i = E [ F 2 ∣ U = i ] G_i = \mathbb{E}[F_2 \mid U=i] G i = E [ F 2 ∣ U = i ] , M = ∑ i p i G i M = \sum_i p_iG_i M = ∑ i p i G i , and Z i = G i − M Z_i = G_i-M Z i = G i − M . Regularize by adding ε I \varepsilon I ε I if necessary. The proof of Theorem 2.1, together with (3.3), gives
H 2 − 2 H 1 ≥ − ∑ i j W i j ⟨ Z i , Z j ⟩ M ≥ κ ∑ i p i ⟨ Z i , Z i ⟩ M , H_2-2H_1 \ge-\sum_{ij}W_{ij}\langle Z_i,Z_j\rangle_M \ge\kappa\sum_i p_i\langle Z_i,Z_i\rangle_M, H 2 − 2 H 1 ≥ − ij ∑ W ij ⟨ Z i , Z j ⟩ M ≥ κ i ∑ p i ⟨ Z i , Z i ⟩ M ,
where H k = H μ k ( F k ) H_k = \mathcal{H}_{\mu_k}(F_k) H k = H μ k ( F k ) and ⟨ A , B ⟩ M = Tr A T M ( B ) \langle A,B\rangle_M = \operatorname{Tr} A\mathcal{T}_M(B) ⟨ A , B ⟩ M = Tr A T M ( B ) . Operator concavity of the logarithm implies
log A − log M ⪯ T M ( A − M ) . \log A-\log M \preceq\mathcal{T}_M(A-M). log A − log M ⪯ T M ( A − M ) .
Taking the trace against A A A , and using T M ( M ) = I \mathcal{T}_M(M)=I T M ( M ) = I , yields
Tr [ A ( log A − log M ) − A + M ] ≤ Tr [ ( A − M ) T M ( A − M ) ] . \operatorname{Tr}[A(\log A-\log M)-A+M] \le\operatorname{Tr}[(A-M)\mathcal{T}_M(A-M)]. Tr [ A ( log A − log M ) − A + M ] ≤ Tr [( A − M ) T M ( A − M )] .
Averaging over A = G i A=G_i A = G i proves H 2 ≥ ( 2 + κ ) H 1 H_2 \ge(2+\kappa)H_1 H 2 ≥ ( 2 + κ ) H 1 . The entropy-profile chord inequality with lower endpoint one gives, for k ≥ 2 k \ge2 k ≥ 2 ,
H k + 1 ≥ k H k − H 1 k − 1 . (17) H_{k+1} \ge\frac{kH_k-H_1}{k-1}. \tag*{(17)} H k + 1 ≥ k − 1 k H k − H 1 . (17)
Starting with H 2 ≥ a 2 H 1 H_2 \ge a_2H_1 H 2 ≥ a 2 H 1 , induction in (3.5) gives H k ≥ a k H 1 H_k \ge a_kH_1 H k ≥ a k H 1 for every k k k . Substitution of this last estimate back into the same recursion gives
H k + 1 ≥ ( k k − 1 − 1 ( k − 1 ) a k ) H k = a k + 1 a k H k . H_{k+1} \ge\left(\frac{k}{k-1}-\frac{1}{(k-1)a_k}\right)H_k = \frac{a_{k+1}}{a_k}H_k. H k + 1 ≥ ( k − 1 k − ( k − 1 ) a k 1 ) H k = a k a k + 1 H k .
Together with the rank-two estimate, these inequalities telescope to (3.4). Conditional replacement cannot increase matrix entropy, so its one-step coefficient is at most a r − 1 / a r a_{r-1}/a_r a r − 1 / a r . The supporting-hyperplane inequality for Tr A log A \operatorname{Tr} A\log A Tr A log A then gives the asserted entropy-production constant. □
Corollary 3.5 (Exact structural criterion for the universal coefficient). Let M M M be a loopless matroid of rank r ≥ 2 r \ge2 r ≥ 2 , with a fixed positive product field, and let 1 ≤ ℓ < r 1 \le\ell< r 1 ≤ ℓ < r . Its complete deletion coefficient equals ℓ / r \ell/r ℓ / r if and only if M M M has a parallel class containing at least two elements. If M M M is simple, the coefficient is strictly smaller than ℓ / r \ell/r ℓ / r , uniformly over the quantum dimension.
Proof. For a simple matroid, the Hessian of its basis polynomial at every positive point has signature ( + , − , … , − ) (+,-,\ldots,-) ( + , − , … , − ) by Murai–Nagaoka–Yazawa [38 ] . After absorbing the field, the Hessian of the two-particle polynomial is a positive scalar multiple of that Hessian. Indeed, for a homogeneous polynomial of degree r r r ,
∇ 2 ( 2 r ! D 1 r − 2 f ) = 2 r ( r − 1 ) ∇ 2 f ( 1 ) . \nabla^2\left(\frac{2}{r!}D_1^{r-2}f\right)=\frac{2}{r(r-1)}\nabla^2 f(\mathbf{1}). ∇ 2 ( r ! 2 D 1 r − 2 f ) = r ( r − 1 ) 2 ∇ 2 f ( 1 ) .
Its restriction to p ⊥ p^{\perp} p ⊥ is negative definite, so (3.3) holds with κ > 0 \kappa> 0 κ > 0 . Since ℓ < r \ell< r ℓ < r , a ℓ / a r < ℓ / r a_{\ell}/a_{r} < \ell/r a ℓ / a r < ℓ / r .
Conversely, let C C C be a nontrivial parallel class and put q i = λ i / ∑ j ∈ C λ j q_i = \lambda_i/\sum_{j\in C}\lambda_j q i = λ i / ∑ j ∈ C λ j . Choose nonconstant positive matrices A i A_i A i with mean A = ∑ i ∈ C q i A i A = \sum_{i\in C}q_iA_i A = ∑ i ∈ C q i A i , and set
F ( B ) = { A i if B ∩ C = { i } , A if B ∩ C = ∅ . F(B) =
\begin{cases}
A_i & \text{if } B \cap C = \{i\},\\
A & \text{if } B \cap C = \varnothing.
\end{cases} F ( B ) = { A i A if B ∩ C = { i } , if B ∩ C = ∅ .
Conditional on the other elements of a basis meeting C C C , its element in C C C has law q q q . If a retained occupation misses C C C , the conditional mean of F F F is therefore A A A ; if it contains i i i , the conditional mean is A i A_i A i . The probability of retaining the class element is ℓ / r \ell/r ℓ / r times its original probability. The output matrix entropy is consequently exactly ℓ / r \ell/r ℓ / r times the input entropy. Scalar nonconstant A i A_i A i already suffice.
Trace duality and the characterization of Lorentzian laws
Entropy contraction and exponential trace comparison are dual forms of the same finite-channel inequality. We keep an arbitrary Hermitian offset in the exponential. This is the form that can be iterated under conditioning in the concentration arguments.
Let μ \mu μ be a probability measure on X \mathcal{X} X , let D ( y ∣ x ) D(y \mid x) D ( y ∣ x ) be a stochastic kernel, and put ν = μ D \nu= \mu D ν = μ D . For a matrix function F F F on X \mathcal{X} X , define
T F ( y ) = E [ F ( X ) ∣ Y = y ] , D ∗ H ( x ) = ∑ y D ( y ∣ x ) H ( y ) . TF(y) = \mathbb{E}[F(X) \mid Y = y], \qquad D^{*}H(x) = \sum_y D(y \mid x)H(y). T F ( y ) = E [ F ( X ) ∣ Y = y ] , D ∗ H ( x ) = y ∑ D ( y ∣ x ) H ( y ) .
Zero reference probabilities are omitted. The means of F F F and T F TF T F agree.
Theorem 4.1 (Complete entropy–trace duality). Fix a matrix dimension and a real number c ≥ 1 c \ge1 c ≥ 1 . The following assertions are equivalent.
For every F ⪰ 0 F \succeq0 F ⪰ 0 , H ν ( T F ) ≤ c − 1 H μ ( F ) H_{\nu}(TF) \le c^{-1}H_{\mu}(F) H ν ( T F ) ≤ c − 1 H μ ( F ) .
For every positive definite matrix family A y A_y A y , with M = E ν A M = \mathbb{E}_{\nu}A M = E ν A , and every Hermitian K K K ,
E μ Tr exp ( K + c D ∗ ( log A ) ) ≤ Tr exp ( K + c log M ) . (18) \mathbb{E}_{\mu}\operatorname{Tr}\exp\left(K + cD^{*}(\log A)\right) \le\operatorname{Tr}\exp(K + c\log M). \tag*{(18)} E μ Tr exp ( K + c D ∗ ( log A ) ) ≤ Tr exp ( K + c log M ) . (18) For every such family,
E μ Tr exp ( c D ∗ ( log A ) − ( c − 1 ) log M ) ≤ Tr M . (19) \mathbb{E}_{\mu}\operatorname{Tr}\exp\left(cD^{*}(\log A) - (c-1)\log M\right) \le\operatorname{Tr}M. \tag*{(19)} E μ Tr exp ( c D ∗ ( log A ) − ( c − 1 ) log M ) ≤ Tr M . (19) The statements hold at every matrix dimension simultaneously whenever one of them does.
Proof. The same exponential family proves both directions. Write D r a w ( X ∥ Y ) = Tr X ( log X − log Y ) D_{\mathrm{raw}}(X\mathbin{\|}Y) = \operatorname{Tr}X(\log X - \log Y) D raw ( X ∥ Y ) = Tr X ( log X − log Y ) . This function is jointly convex, and Klein’s inequality states D r a w ( X ∥ Y ) ≥ Tr X − Tr Y D_{\mathrm{raw}}(X\mathbin{\|}Y) \ge\operatorname{Tr}X - \operatorname{Tr}Y D raw ( X ∥ Y ) ≥ Tr X − Tr Y . Assume (1), and set
F x = exp ( K + c D ∗ ( log A ) ( x ) ) , J y = T F ( y ) , N = E μ F = E ν J . F_x = \exp\left(K + cD^{*}(\log A)(x)\right), \qquad J_y = TF(y), \qquad N = \mathbb{E}_{\mu}F = \mathbb{E}_{\nu}J. F x = exp ( K + c D ∗ ( log A ) ( x ) ) , J y = T F ( y ) , N = E μ F = E ν J .
Adjointness gives
E μ Tr F log F = Tr N K + c E ν Tr J log A . \mathbb{E}_{\mu}\operatorname{Tr}F\log F = \operatorname{Tr}NK + c\mathbb{E}_{\nu}\operatorname{Tr}J\log A. E μ Tr F log F = Tr N K + c E ν Tr J log A .
Consequently the entropy inequality implies
c E ν D r a w ( J ∥ A ) ≤ Tr N K + ( c − 1 ) Tr N log N . c\mathbb{E}_{\nu}D_{\mathrm{raw}}(J\mathbin{\|}A) \le\operatorname{Tr}NK + (c-1)\operatorname{Tr}N\log N. c E ν D raw ( J ∥ A ) ≤ Tr N K + ( c − 1 ) Tr N log N .
Joint convexity bounds the left side below by c D r a w ( N ∥ M ) cD_{\mathrm{raw}}(N\mathbin{\|}M) c D raw ( N ∥ M ) . Rearrangement and Klein’s inequality against exp ( K + c log M ) \exp(K + c\log M) exp ( K + c log M ) prove (2). Taking K = − ( c − 1 ) log M K = -(c-1)\log M K = − ( c − 1 ) log M proves (3).
Conversely, first take positive definite F F F , put J = T F J = TF J = T F and N = E F N = \mathbb{E}F N = E F , and use (3) with A = J A = J A = J . Define
E x = exp ( c D ∗ ( log J ) ( x ) − ( c − 1 ) log N ) . E_x = \exp\left(cD^{*}(\log J)(x) - (c-1)\log N\right). E x = exp ( c D ∗ ( log J ) ( x ) − ( c − 1 ) log N ) .
The averaged Klein inequality gives
E μ Tr F log F ≥ E μ Tr F log E + Tr N − E μ Tr E ≥ c E ν Tr J log J − ( c − 1 ) Tr N log N . \begin{aligned}
\mathbb{E}_{\mu}\operatorname{Tr}F\log F \ge\mathbb{E}_{\mu}\operatorname{Tr}F\log E + \operatorname{Tr}N - \mathbb{E}_{\mu}\operatorname{Tr}E \\
&\ge c\mathbb{E}_{\nu}\operatorname{Tr}J\log J - (c-1)\operatorname{Tr}N\log N.
\end{aligned} E μ Tr F log F ≥ E μ Tr F log E + Tr N − E μ Tr E ≥ c E ν Tr J log J − ( c − 1 ) Tr N log N .
This is (1). Replacing F F F by F + ε I F + \varepsilon I F + ε I and passing to the limit handles positive semidefinite functions.
This explicit duality is a finite-channel form of quantum Brascamp–Lieb duality [10 ] . For the Lorentzian deletion kernel it gives an inequality for matrices indexed by arbitrary subconfigurations.
Corollary 4.2 (Trace inequalities at every order). Let μ \mu μ be a degree-r r r Lorentzian occupation law and let 1 ≤ ℓ ≤ r 1 \le\ell\le r 1 ≤ ℓ ≤ r . For positive definite matrices A β A_\beta A β indexed by retained ℓ \ell ℓ -occupations and any Hermitian K K K ,
E μ Tr exp ( K + r ℓ ∑ β D r → ℓ ( β ∣ α ) log A β ) ≤ Tr exp ( K + r ℓ log E μ ℓ A β ) . (20) \mathbb{E}_{\mu}\operatorname{Tr}\exp\left(K+\frac{r}{\ell}\sum_{\beta}D_{r\to\ell}(\beta\mid\alpha)\log A_{\beta}\right)\leq\operatorname{Tr}\exp\left(K+\frac{r}{\ell}\log\mathbb{E}_{\mu_{\ell}}A_{\beta}\right). \tag*{(20)} E μ Tr exp K + ℓ r β ∑ D r → ℓ ( β ∣ α ) log A β ≤ Tr exp ( K + ℓ r log E μ ℓ A β ) . (20) The coefficient r / ℓ r/\ell r / ℓ is the largest universally valid coefficient in this trace inequality over the class of Lorentzian laws. Under the two-particle gap hypothesis (15) of § 3.2 , it may be increased to a r / a ℓ a_r/a_\ell a r / a ℓ , where a k = k + ( k − 1 ) κ a_k=k+(k-1)\kappa a k = k + ( k − 1 ) κ .
Proof. Apply Theorem 4.1 to Theorems 2.1 and 3.4 . The product-of-pairs equality example in Theorem 2.1 , together with the converse direction of the duality, proves universal optimality.
For ℓ = 1 \ell=1 ℓ = 1 , writing p i = E α i / r p_i=\mathbb{E}\alpha_i/r p i = E α i / r , the inequality becomes
E μ Tr exp ( K + ∑ i α i log A i ) ≤ Tr exp ( K + r log ∑ i p i A i ) . \mathbb{E}_{\mu}\operatorname{Tr}\exp\left(K+\sum_i\alpha_i\log A_i\right)\leq\operatorname{Tr}\exp\left(K+r\log\sum_i p_iA_i\right). E μ Tr exp ( K + i ∑ α i log A i ) ≤ Tr exp ( K + r log i ∑ p i A i ) .
For r = 1 r=1 r = 1 this is Lieb’s concavity inequality. The general theorem allows arbitrary noncommuting matrices on every retained level.
The scalar trace comparison has a geometric interpretation: f 1 / r f^{1/r} f 1/ r lies below each of its tangent affine functions. This recovers concavity and, after taking derivatives, Lorentzianity. For multiaffine polynomials, the scalar all-field equivalence below is the α = 1 \alpha=1 α = 1 case of the entropic-independence characterization in [4 Theorem 4] . We include the tangent-majorant argument to handle general homogeneous polynomials and to identify the equivalent complete conditions.
Theorem 4.3 (Entropic characterization of Lorentzian polynomials). Let f f f be a nonzero homogeneous polynomial with nonnegative coefficients. The following are equivalent.
f f f is Lorentzian.
Every nonzero partial derivative g = ∂ γ f g=\partial^\gamma f g = ∂ γ f of positive degree k k k , at every positive external field, has scalar selected-particle entropy contraction at most 1 / k 1/k 1/ k .
Every law in (2) has complete selected-particle contraction at most 1 / k 1/k 1/ k .
Every law in (2) has the complete deletion inequalities at every retained level, with coefficients ℓ / k \ell/k ℓ / k .
If f f f is multiaffine of positive degree r r r , it suffices in (2) to test f f f itself at every positive field, with coefficient 1 / r 1/r 1/ r . The same reduction applies to the complete conditions (3) and (4).
Proof. Lorentzian polynomials are closed under partial derivatives and positive diagonal substitutions. Theorem 2.1 therefore gives (1)⇒(4)⇒(3)⇒(2).
Assume (2). Fix a nonzero derivative g g g of degree k k k and a positive point a a a . Its field law has selected-particle distribution p i = a i ∂ i g ( a ) / ( k g ( a ) ) p_i=a_i\partial_i g(a)/(kg(a)) p i = a i ∂ i g ( a ) / ( k g ( a )) . Apply the scalar version of Theorem 4.1 with A i = b i / a i A_i=b_i/a_i A i = b i / a i . For every b > 0 b>0 b > 0 ,
g ( b ) g ( a ) ≤ ( ∑ i b i ∂ i g ( a ) k g ( a ) ) k . \frac{g(b)}{g(a)}\leq\left(\frac{\sum_i b_i\partial_i g(a)}{k g(a)}\right)^k. g ( a ) g ( b ) ≤ ( k g ( a ) ∑ i b i ∂ i g ( a ) ) k .
Thus ϕ = g 1 / k \phi=g^{1/k} ϕ = g 1/ k lies below its tangent affine function at every positive point:
ϕ ( b ) ≤ ϕ ( a ) + ∇ ϕ ( a ) ⋅ ( b − a ) . \phi(b)\leq\phi(a)+\nabla\phi(a)\cdot(b-a). ϕ ( b ) ≤ ϕ ( a ) + ∇ ϕ ( a ) ⋅ ( b − a ) .
The function ϕ \phi ϕ is the infimum of these affine majorants, so it is concave. In particular g g g is log-concave on the positive orthant. This holds for every nonzero partial derivative, proving strong log-concavity. For homogeneous polynomials with nonnegative coefficients, strong log-concavity is equivalent to Lorentzianity by Brändén–Huh [13 Theorem 2.30] .
For multiaffine f f f , the test on f f f alone first proves its log-concavity by the same tangent argument. For every variable,
∂ i f = lim u → ∞ u − 1 f ∣ z i = u . \partial_i f=\lim_{u\to\infty}u^{-1}f\big|_{z_i=u}. ∂ i f = u → ∞ lim u − 1 f z i = u .
Positive pointwise limits preserve log-concavity. Iterating this identity over distinct variables proves log-concavity of every nonzero partial derivative; repeated derivatives are zero. Thus f f f is strongly log-concave and hence Lorentzian. □
Corollary 4.4 (Recognition at one positive field). Fix one positive field x x x . A nonzero homogeneous polynomial f f f with nonnegative coefficients is Lorentzian if and only if every nonzero coordinate derivative g g g of degree k ≥ 2 k \ge2 k ≥ 2 , at that field, satisfies
Var ( T k → 1 u ) ≤ k − 1 Var ( u ) \operatorname{Var}(T_{k\to1}u)\le k^{-1}\operatorname{Var}(u) Var ( T k → 1 u ) ≤ k − 1 Var ( u ) for every linear count observable u ( α ) = α ⋅ h u(\alpha)=\alpha\cdot h u ( α ) = α ⋅ h . Thus these infinitesimal scalar tests imply all the complete inequalities at every positive field.
Proof. The forward direction follows by linearizing complete deletion. For the converse absorb x x x into g g g , normalize g ( 1 ) = 1 g(\mathbf{1})=1 g ( 1 ) = 1 , and put W = ∇ 2 g ( 1 ) / ( k ( k − 1 ) ) W=\nabla^2g(\mathbf{1})/(k(k-1)) W = ∇ 2 g ( 1 ) / ( k ( k − 1 )) , p = W 1 p=W\mathbf{1} p = W 1 and P = diag ( p ) P=\operatorname{diag}(p) P = diag ( p ) on the active labels. Here T 1 ∗ h ( α ) = α ⋅ h / k T_{1}^{*}h(\alpha)=\alpha\cdot h/k T 1 ∗ h ( α ) = α ⋅ h / k and T 1 T 1 ∗ = [ I + ( k − 1 ) P ] / k T_{1}T_{1}^{*}=[I+(k-1)P]/k T 1 T 1 ∗ = [ I + ( k − 1 ) P ] / k : after conditioning on a tagged particle, a uniformly selected particle is the tag with probability 1 / k 1/k 1/ k , and otherwise follows the pair transition P P P . A positive nonconstant eigenvalue of P P P would violate the test on the linear function T 1 ∗ h T_{1}^{*}h T 1 ∗ h . Hence each derivative Hessian has only one positive eigenvalue. It is indecomposable: a decomposition in disjoint active variables would give two positive Hessian blocks. The quadratic derivative Hessians are constant, so their log-concavity holds everywhere. The recognition theorem of Anari–Liu–Oveis Gharan–Vinzant [6 Theorem 3.2] now gives complete log-concavity, hence Lorentzianity. Degrees zero and one are automatic. □
Translation directions and all scalar equality cases
Every scalar equality density comes from a translation that leaves the generating polynomial unchanged. For a Lorentzian polynomial, these global translations are exactly the null directions of its Hessian at any positive point. This identifies all equality cases and decides when the deletion coefficient is sharp.
Absorb the positive reference field into the variables and normalize the Lorentzian generating polynomial as
f ( z ) = ∑ ∣ α ∣ = r μ ( α ) z α , f ( 1 ) = 1 , r ≥ 2. f(z)=\sum_{|\alpha|=r}\mu(\alpha)z^\alpha,\qquad f(\mathbf{1})=1,\qquad r\ge2. f ( z ) = ∣ α ∣ = r ∑ μ ( α ) z α , f ( 1 ) = 1 , r ≥ 2.
Remove variables with ∂ i f = 0 \partial_i f=0 ∂ i f = 0 . Put
N f = { h ∈ R n : D h f = 0 } , D h = ∑ i h i ∂ i . \mathcal{N}_f=\{h\in\mathbb{R}^n:D_hf=0\},\qquad D_h=\sum_i h_i\partial_i. N f = { h ∈ R n : D h f = 0 } , D h = i ∑ h i ∂ i .
Equivalently, h ∈ N f h\in\mathcal{N}_f h ∈ N f means f ( z + t h ) = f ( z ) f(z+th)=f(z) f ( z + t h ) = f ( z ) as a polynomial identity in z z z and t t t . The space N f \mathcal{N}_f N f is therefore intrinsic to f f f .
Lemma 5.1 (The Hessian kernel is the translation space). For every positive x x x ,
ker ∇ 2 f ( x ) = N f . \ker\nabla^2f(x)=\mathcal{N}_f. ker ∇ 2 f ( x ) = N f . Consequently the Hessian rank is constant on the positive orthant.
Proof. Differentiating D h f = 0 D_hf=0 D h f = 0 gives one inclusion. For the converse, the Lorentzian signature forces a Hessian null vector to lie in the kernel of every derivative Hessian. Iteration then turns a null vector at one point into a polynomial identity. This is the kernel argument of Brändén–Huh [13 Lemma 2.15] , also expressed by the degree-one Hodge–Riemann theorem [38 Theorem 3.8] ; the proof below includes identically zero derivatives.
Induct on the degree, starting with the constant Hessian of a quadratic. For degree d ≥ 3 d \ge3 d ≥ 3 , Euler’s identities give
( d − 2 ) ∇ 2 f ( x ) = ∑ i x i ∇ 2 ( ∂ i f ) ( x ) , x T ∇ 2 ( ∂ i f ) ( x ) = ( d − 2 ) e i T ∇ 2 f ( x ) . (d-2)\nabla^2 f(x)=\sum_i x_i\nabla^2(\partial_i f)(x),\qquad x^{\mathsf T}\nabla^2(\partial_i f)(x)=(d-2)e_i^{\mathsf T}\nabla^2f(x). ( d − 2 ) ∇ 2 f ( x ) = i ∑ x i ∇ 2 ( ∂ i f ) ( x ) , x T ∇ 2 ( ∂ i f ) ( x ) = ( d − 2 ) e i T ∇ 2 f ( x ) .
Let h ∈ ker ∇ 2 f ( x ) h\in\ker\nabla^2f(x) h ∈ ker ∇ 2 f ( x ) . The second identity makes h h h orthogonal to x x x with respect to each derivative Hessian H i H_i H i : x T H i h = 0 x^{\mathsf T}H_i h=0 x T H i h = 0 . For every nonzero derivative,
x T H i x = ( d − 1 ) ( d − 2 ) ∂ i f ( x ) > 0. x^{\mathsf T}H_i x=(d-1)(d-2)\partial_i f(x)>0. x T H i x = ( d − 1 ) ( d − 2 ) ∂ i f ( x ) > 0.
Thus H i H_i H i has exactly one positive eigenvalue. Its restriction to this orthogonal complement is therefore nonpositive, with zero square only on its kernel. The first identity now expresses 0 0 0 as a sum of these nonpositive squares with strictly positive coefficients x i x_i x i . Every square vanishes, so h h h lies in every derivative Hessian kernel.
Induction gives D h ∂ i f = 0 D_h\partial_i f=0 D h ∂ i f = 0 for each nonzero derivative; the identity holds for the zero derivatives as well. Hence every partial derivative of D h f D_hf D h f vanishes. Since D h f D_hf D h f is homogeneous of positive degree, it is zero. Thus h h h generates a global translation of f f f . □
Theorem 5.2 (All scalar equality densities). Let f f f be as above and let 1 ≤ ℓ < r 1\le\ell<r 1 ≤ ℓ < r . For a probability law ν \nu ν supported on supp μ \operatorname{supp}\mu supp μ , equality in
D ( ν D r → ℓ ∥ μ D r → ℓ ) ≤ ℓ r D ( ν ∥ μ ) D(\nu D_{r\to\ell}\Vert\mu D_{r\to\ell})\le\frac{\ell}{r}D(\nu\Vert\mu) D ( ν D r → ℓ ∥ μ D r → ℓ ) ≤ r ℓ D ( ν ∥ μ ) holds if and only if there is an a ∈ R ≥ 0 n a\in\mathbb{R}_{\ge0}^{n} a ∈ R ≥ 0 n such that
a − 1 ∈ N f , ν ( α ) = μ ( α ) a α . (21) a-1\in\mathcal{N}_f,\qquad\nu(\alpha)=\mu(\alpha)a^\alpha. \tag*{(21)} a − 1 ∈ N f , ν ( α ) = μ ( α ) a α . (21) These densities attain equality at every retained level. Writing p i = ∂ i f ( 1 ) / r p_i=\partial_i f(1)/r p i = ∂ i f ( 1 ) / r , their entropies are
H μ k ( a β ) = k ∑ i p i a i log a i ( 0 ≤ k ≤ r ) , (22) \mathcal{H}_{\mu_k}(a^\beta)=k\sum_i p_i a_i\log a_i\qquad(0\le k\le r), \tag*{(22)} H μ k ( a β ) = k i ∑ p i a i log a i ( 0 ≤ k ≤ r ) , (22) with 0 log 0 = 0 0\log0=0 0 log 0 = 0 and 0 0 = 1 0^0=1 0 0 = 1 .
Proof. Write F = ν / μ F=\nu/\mu F = ν / μ and H k = H μ k ( F k ) H_k=\mathcal{H}_{\mu_k}(F_k) H k = H μ k ( F k ) . Equality at one proper retained level makes the entire entropy profile linear, by Theorem 3.1 : H k = k H 1 H_k=kH_1 H k = k H 1 for every k k k . The quadratic level will identify the translation direction, and H r = r H 1 H_r=rH_1 H r = r H 1 will identify the density.
Set a i = F 1 ( i ) a_i=F_1(i) a i = F 1 ( i ) and h = a − 1 h=a-1 h = a − 1 . Then p T h = 0 p^{\mathsf T}h=0 p T h = 0 . The quadratic proof of Theorem 2.1 gives
0 = H 2 − 2 H 1 ≥ − h T W h ≥ 0 , W = ∇ 2 f ( 1 ) r ( r − 1 ) . 0=H_2-2H_1\ge-h^{\mathsf T}Wh\ge0,\qquad W=\frac{\nabla^2f(1)}{r(r-1)}. 0 = H 2 − 2 H 1 ≥ − h T W h ≥ 0 , W = r ( r − 1 ) ∇ 2 f ( 1 ) .
The form of W W W is nonpositive on p ⊥ p^\perp p ⊥ . Polarizing at the zero square h T W h = 0 h^{\mathsf T}Wh=0 h T W h = 0 shows that W h Wh W h annihilates p ⊥ p^\perp p ⊥ , so W h Wh W h is proportional to p p p . But 1 T W h = p T h = 0 1^{\mathsf T}Wh=p^{\mathsf T}h=0 1 T W h = p T h = 0 and 1 T p = 1 1^{\mathsf T}p=1 1 T p = 1 , forcing W h = 0 Wh=0 W h = 0 . Lemma 5.1 therefore gives h ∈ N f h\in\mathcal{N}_f h ∈ N f .
Translation invariance normalizes the candidate extremizer: f ( a ) = f ( 1 ) = 1 f(a)=f(1)=1 f ( a ) = f ( 1 ) = 1 , so μ a ( α ) = μ ( α ) a α \mu_a(\alpha)=\mu(\alpha)a^\alpha μ a ( α ) = μ ( α ) a α is a probability law. It contains the support of ν \nu ν : whenever F ( α ) > 0 F(\alpha)>0 F ( α ) > 0 and α i > 0 \alpha_i>0 α i > 0 , that occupation contributes positively to F 1 ( i ) = a i F_1(i)=a_i F 1 ( i ) = a i . Relative entropy now gives the remaining rigidity:
D ( ν ∥ μ a ) = H r − ∑ i E ν α i log a i = H r − r ∑ i p i a i log a i = H r − r H 1 = 0. D(\nu\Vert\mu_a)=H_r-\sum_i\mathbb{E}_{\nu}\alpha_i\log a_i=H_r-r\sum_i p_i a_i\log a_i=H_r-rH_1=0. D ( ν ∥ μ a ) = H r − i ∑ E ν α i log a i = H r − r i ∑ p i a i log a i = H r − r H 1 = 0.
Vanishing relative entropy forces ν = μ a \nu=\mu_a ν = μ a , proving necessity.
Conversely, suppose (5.1) holds. Translation invariance passes to every partial derivative of f f f . The conditional-mean formula for deletion therefore gives, for each retained occupation β \beta β ,
E μ [ a α ∣ β ] = a β ∂ β f ( a ) ∂ β f ( 1 ) = a β . \mathbb{E}_{\mu}[a^\alpha\mid\beta]=a^\beta\frac{\partial^\beta f(a)}{\partial^\beta f(1)}=a^\beta. E μ [ a α ∣ β ] = a β ∂ β f ( 1 ) ∂ β f ( a ) = a β .
The identity extends to a i = 0 a_i=0 a i = 0 by polynomial continuity. The normalized generating polynomial f k f_k f k of μ k \mu_k μ k is a positive multiple of D 1 r − k f D_1^{r-k}f D 1 r − k f , so f k ( a ) = 1 f_k(a)=1 f k ( a ) = 1 and ∂ i f k ( a ) = ∂ i f k ( 1 ) = k p i \partial_i f_k(a)=\partial_i f_k(1)=kp_i ∂ i f k ( a ) = ∂ i f k ( 1 ) = k p i . Its tilted mean occupation is consequently k p i a i kp_i a_i k p i a i . Expanding log a β \log a^\beta log a β , with the stated boundary convention, gives (5.2) and equality at every level. □
For an arbitrary density, its one-particle marginal still selects a canonical field tilt. The entropy deficit splits exactly into relative entropy from that tilt and a normalization defect.
Proposition 5.3 (The deficit controls an explicit field tilt). For any probability law ν \nu ν on supp μ \operatorname{supp}\mu supp μ , set G i = ( ν / μ ) 1 ( i ) G_i=(\nu/\mu)_1(i) G i = ( ν / μ ) 1 ( i ) , Z = f ( G ) Z=f(G) Z = f ( G ) and μ G ( α ) = Z − 1 μ ( α ) G α \mu_G(\alpha)=Z^{-1}\mu(\alpha)G^\alpha μ G ( α ) = Z − 1 μ ( α ) G α . Then 0 < Z ≤ 1 0<Z\le1 0 < Z ≤ 1 and
D ( ν ∥ μ ) − r D ( ν D r → 1 ∥ μ D r → 1 ) = D ( ν ∥ μ G ) − log Z . (23) D(\nu\Vert\mu)-rD(\nu D_{r\to1}\Vert\mu D_{r\to1})=D(\nu\Vert\mu_G)-\log Z. \tag*{(23)} D ( ν ∥ μ ) − r D ( ν D r → 1 ∥ μ D r → 1 ) = D ( ν ∥ μ G ) − log Z . (23) Writing H k = D ( ν D r → k ∥ μ D r → k ) H_k=D(\nu D_{r\to k}\Vert\mu D_{r\to k}) H k = D ( ν D r → k ∥ μ D r → k ) , every 1 ≤ ℓ < r 1\le\ell<r 1 ≤ ℓ < r also satisfies
ℓ r H r − H ℓ ≥ r − ℓ r ( r − 1 ) ( D ( ν ∥ μ G ) − log Z ) . (24) \frac{\ell}{r}H_r-H_\ell\ge\frac{r-\ell}{r(r-1)}\bigl(D(\nu\Vert\mu_G)-\log Z\bigr). \tag*{(24)} r ℓ H r − H ℓ ≥ r ( r − 1 ) r − ℓ ( D ( ν ∥ μ G ) − log Z ) . (24) Both terms in parentheses are nonnegative.
Proof. Any occupation of positive ν \nu ν mass has G i > 0 G_i>0 G i > 0 wherever its exponent is positive, so Z > 0 Z>0 Z > 0 . The scalar bound in Corollary 4.2, extended to the boundary by continuity, gives
Z ≤ ( ∑ i p i G i ) r = 1. Z\le\left(\sum_i p_iG_i\right)^r=1. Z ≤ ( i ∑ p i G i ) r = 1.
Expansion of D ( ν ∥ μ G ) D(\nu\Vert\mu_G) D ( ν ∥ μ G ) gives (5.3). For ℓ > 1 \ell>1 ℓ > 1 , the chord from level one to level r r r gives
H ℓ ≤ ℓ − 1 r − 1 H r + r − ℓ r − 1 H 1 . H_\ell\le\frac{\ell-1}{r-1}H_r+\frac{r-\ell}{r-1}H_1. H ℓ ≤ r − 1 ℓ − 1 H r + r − 1 r − ℓ H 1 .
Rearranging and substituting (5.3) proves (5.4); for ℓ = 1 \ell=1 ℓ = 1 it is an identity. □
Corollary 5.4 (Exact sharpness for every Lorentzian law). For 1 ≤ ℓ < r 1\le\ell<r 1 ≤ ℓ < r , the complete deletion coefficient of a fixed positive-field Lorentzian law equals ℓ / r \ell/r ℓ / r if and only if its active-variable Hessian is singular. Equivalently, its nonzero first partial derivatives are linearly dependent. In the singular case, nonconstant positive scalar densities attain the coefficient. In the nonsingular case, the improvement in Theorem 3.4 is strict in every matrix dimension. This dichotomy is independent of the positive reference field.
Proof. A nonsingular Hessian is negative definite on p ⊥ p^\perp p ⊥ . Thus (3.3) holds with κ > 0 \kappa>0 κ > 0 , and the complete contraction coefficient is strictly below ℓ / r \ell/r ℓ / r . If 0 ≠ h ∈ N f 0\ne h\in N_f 0 = h ∈ N f , take a = 1 + t h > 0 a=1+th>0 a = 1 + t h > 0 for small nonzero t t t . Euler’s identity gives p T h = 0 p^{\mathsf T}h=0 p T h = 0 . Since every p i > 0 p_i>0 p i > 0 , strict convexity of u log u u\log u u log u gives ∑ i p i a i log a i > 0 \sum_i p_i a_i\log a_i>0 ∑ i p i a i log a i > 0 . The density a α a^\alpha a α is therefore nonconstant and attains the complete upper bound. Finally, the translation space is independent of the reference point, while positive fields act by invertible diagonal coordinate changes. The dichotomy holds throughout the field family. □
Corollary 5.5 (Sharpness over the whole field family). Let f f f be multiaffine Lorentzian of degree r ≥ 2 r\ge2 r ≥ 2 with at least two supported bases. For 1 ≤ ℓ < r 1\le\ell<r 1 ≤ ℓ < r ,
sup λ > 0 η c o m p l e t e ( μ λ , D r → ℓ ) = ℓ / r . \sup_{\lambda>0}\eta_{\mathrm{complete}}(\mu_\lambda,D_{r\to\ell})=\ell/r. λ > 0 sup η complete ( μ λ , D r → ℓ ) = ℓ / r . Thus strictness at every fixed positive field need not give an improvement uniform over fields.
Proof. Expose an edge of the support matroid’s base polytope and send the positive field along an exposing functional. The laws converge to a nondegenerate law on two bases differing by one exchange. On this pair, deletion retains the distinguishing particle with probability ℓ / r \ell/r ℓ / r and otherwise erases the choice. Relative entropy therefore contracts by exactly ℓ / r \ell/r ℓ / r .
Fix a nonconstant scalar tilt of this limiting two-point law and view it as a probability law ν \nu ν on the original support. Every point of positive ν \nu ν mass has positive limiting reference mass, and the same holds after deletion. Thus the scalar relative entropies converge to their two-point values. Their output-to-input ratio tends to ℓ / r \ell/r ℓ / r , giving the lower bound on the supremum. Complete deletion gives the upper bound. □
Equality for product-field matroid bases
For a matroid, the translation space has a concrete meaning: one may redistribute mass inside a parallel class while leaving the class itself unchanged. The equality theorem below shows that these redistributions exhaust all scalar extremizers.
Let M M M be a loopless rank-r r r matroid, let λ i > 0 \lambda_i>0 λ i > 0 , and write μ \mu μ for its product-field basis law. Partition the ground set into parallel classes C C C . Put
w C = ∑ i ∈ C λ i , q i = λ i / w C . w_C=\sum_{i\in C}\lambda_i,\qquad q_i=\lambda_i/w_C. w C = i ∈ C ∑ λ i , q i = λ i / w C .
Singleton classes are included. A basis contains at most one element from each class. On the classes it induces the product-field basis law of the simplification M ‾ \overline{M} M , with fields w C w_C w C . Conditioned on this simplified basis, the chosen labels in its classes are independent and have probabilities q i q_i q i .
Theorem 5.6 (Complete description of scalar equality). Suppose r ≥ 2 r\ge2 r ≥ 2 and 1 ≤ ℓ < r 1\le\ell<r 1 ≤ ℓ < r . For any probability law ν \nu ν on the bases of M M M , equality in
D ( ν D r → ℓ ∥ μ D r → ℓ ) ≤ ℓ r D ( ν ∥ μ ) D(\nu_{D_{r\to\ell}}\Vert\mu_{D_{r\to\ell}})\le\frac{\ell}{r}D(\nu\Vert\mu) D ( ν D r → ℓ ∥ μ D r → ℓ ) ≤ r ℓ D ( ν ∥ μ ) holds if and only if there are numbers a i ≥ 0 a_i\ge0 a i ≥ 0 such that
∑ i ∈ C q i a i = 1 for every parallel class C , ν ( B ) = μ ( B ) ∏ i ∈ B a i . (25) \sum_{i\in C}q_i a_i=1\quad\text{for every parallel class }C,\qquad\nu(B)=\mu(B)\prod_{i\in B}a_i. \tag*{(25)} i ∈ C ∑ q i a i = 1 for every parallel class C , ν ( B ) = μ ( B ) i ∈ B ∏ a i . (25) Every such density attains equality simultaneously at all retained levels. Thus equality changes only the distribution of the label inside each parallel class; it preserves the law of the simplified basis. For a simple matroid, equality forces ν = μ \nu=\mu ν = μ .
Proof. Let L L L be the class-sum map ( L h ) C = ∑ i ∈ C λ i h i (Lh)_C=\sum_{i\in C}\lambda_i h_i ( L h ) C = ∑ i ∈ C λ i h i . The field-absorbed basis polynomial factors through L L L , and its Hessian is a positive scalar multiple of L T ∇ 2 f M ‾ ( w ) L L^{\mathsf T}\nabla^2 f_{\overline{M}}(w)L L T ∇ 2 f M ( w ) L . The map L L L is onto. The Hessian of the simple matroid M ‾ \overline{M} M is nonsingular by Murai–Nagaoka–Yazawa [38 ] , so the field-absorbed Hessian has kernel exactly ker L \ker L ker L . Lemma 5.1 identifies this with the translation space of the field-absorbed polynomial. The condition a − 1 ∈ ker L a-1\in\ker L a − 1 ∈ ker L is exactly ∑ i ∈ C q i a i = 1 \sum_{i\in C}q_i a_i=1 ∑ i ∈ C q i a i = 1 for each class. Theorem 5.2 proves the assertion, including zero coordinates and equality at every retained level.
The full extremal eigenspace
Translation directions determine every eigenfunction that attains the contraction bound ℓ / r \ell/r ℓ / r . A one-particle Gram identity detects these directions, and the variance chord forces the same equality space at every retained level. This also identifies the slow part of the deletion–replacement heat flow.
Keep a normalized Lorentzian law of degree r ≥ 2 r\ge2 r ≥ 2 on its active variables. Write T ℓ F = E [ F ∣ β ] T_\ell F=\mathbb{E}[F\mid\beta] T ℓ F = E [ F ∣ β ] and K ℓ = T ℓ ∗ T ℓ K_\ell=T_\ell^{*}T_\ell K ℓ = T ℓ ∗ T ℓ , the operator that deletes to level ℓ \ell ℓ and then conditionally refills. All spaces use the reference inner products. Put
W = ∇ 2 f ( 1 ) r ( r − 1 ) , p = W 1 , P = diag ( p ) − 1 W . W=\frac{\nabla^2 f(1)}{r(r-1)},\qquad p=W\mathbf{1},\qquad P=\operatorname{diag}(p)^{-1}W. W = r ( r − 1 ) ∇ 2 f ( 1 ) , p = W 1 , P = diag ( p ) − 1 W .
Theorem 5.7 (The entire extremal eigenspace). For 1 ≤ ℓ < r 1\le\ell<r 1 ≤ ℓ < r ,
ker ( K ℓ − ℓ / r ) = { α ↦ α ⋅ h : h ∈ N f } . \ker(K_\ell-\ell/r)=\{\alpha\mapsto\alpha\cdot h:h\in\mathcal{N}_f\}. ker ( K ℓ − ℓ / r ) = { α ↦ α ⋅ h : h ∈ N f } . Its dimension is the active Hessian nullity. If 0 < κ ≤ 1 / ( r − 1 ) 0<\kappa\le1/(r-1) 0 < κ ≤ 1/ ( r − 1 ) bounds the strictly negative nonconstant spectrum of P P P above by − κ -\kappa − κ , then, on the mean-zero orthogonal complement of these linear modes,
K ℓ ⪯ ℓ − ( r − ℓ ) κ r I . K_\ell\preceq\frac{\ell-(r-\ell)\kappa}{r}I. K ℓ ⪯ r ℓ − ( r − ℓ ) κ I . If there is no strictly negative nonconstant eigenvalue, take κ = 1 / ( r − 1 ) \kappa=1/(r-1) κ = 1/ ( r − 1 ) .
Here κ \kappa κ bounds only the strictly negative spectrum; zero modes are removed explicitly. In contrast, (15) bounds the whole nonconstant spectrum and permits κ > 0 \kappa> 0 κ > 0 only when that kernel is zero.
Proof. Condition on one tagged particle, then select a particle uniformly from the full configuration. With probability 1 / r 1/r 1/ r this selects the tag; otherwise the label transition is P P P . Thus
T 1 T 1 ∗ = I + ( r − 1 ) P r . (26) T_1T_1^*=\frac{I+(r-1)P}{r}. \tag*{(26)} T 1 T 1 ∗ = r I + ( r − 1 ) P . (26)
The nonconstant spectrum of P P P is nonpositive, and ker P = N f \ker P=\mathcal{N}_f ker P = N f by Lemma 5.1. The Gram identity therefore identifies the 1 / r 1/r 1/ r -eigenspace of K 1 K_1 K 1 with the adjoint image of N f \mathcal{N}_f N f . Since T 1 ∗ h ( α ) = α ⋅ h / r T_1^*h(\alpha)=\alpha\cdot h/r T 1 ∗ h ( α ) = α ⋅ h / r , these are the linear counts g h ( α ) = α ⋅ h g_h(\alpha)=\alpha\cdot h g h ( α ) = α ⋅ h . Their means vanish. The factorial-second-moment formula and W h = 0 Wh=0 W h = 0 give
Var ( g h ) = r ∑ i p i h i 2 . \operatorname{Var}(g_h)=r\sum_i p_i h_i^2. Var ( g h ) = r i ∑ p i h i 2 .
All active p i p_i p i are positive, so the map h ↦ g h h\mapsto g_h h ↦ g h is injective.
To transfer this rigidity to level ℓ \ell ℓ , linearize (13) at the constant density. For V j = Var ( T j F ) V_j=\operatorname{Var}(T_jF) V j = Var ( T j F ) , the result is
V ℓ ≤ ℓ − 1 r − 1 V r + r − ℓ r − 1 V 1 . V_\ell\leq\frac{\ell-1}{r-1}V_r+\frac{r-\ell}{r-1}V_1. V ℓ ≤ r − 1 ℓ − 1 V r + r − 1 r − ℓ V 1 .
Combining this with V 1 ≤ V r / r V_1\leq V_r/r V 1 ≤ V r / r gives V ℓ ≤ ℓ V r / r V_\ell\leq\ell V_r/r V ℓ ≤ ℓ V r / r . Equality at level ℓ < r \ell<r ℓ < r forces equality in the one-particle bound, so every extremal mode belongs to the linear space already identified. Conversely, translation invariance gives T ℓ g h ( β ) = β ⋅ h T_\ell g_h(\beta)=\beta\cdot h T ℓ g h ( β ) = β ⋅ h , and uniform retention gives T ℓ ∗ ( β ⋅ h ) = ℓ g h / r T_\ell^*(\beta\cdot h)=\ell g_h/r T ℓ ∗ ( β ⋅ h ) = ℓ g h / r . This proves the eigenspace identity at every level.
On the mean-zero orthogonal complement of these modes, (26) improves the one-particle bound to V 1 ≤ [ 1 − ( r − 1 ) κ ] V r / r V_1\leq[1-(r-1)\kappa]V_r/r V 1 ≤ [ 1 − ( r − 1 ) κ ] V r / r ; the remaining kernel of K 1 K_1 K 1 contributes only eigenvalue zero. Substitution in the same variance chord proves the residual estimate. □
Corollary 5.8 (Down-up equality; Alaniz Pintos’s degree-one conjecture). Let M M M be a loopless rank-r r r matroid with at least two bases, r ≥ 1 r\geq1 r ≥ 1 , and let μ λ \mu_\lambda μ λ be its positive product-field basis law. Let P d u P_{\mathrm{du}} P du delete a uniformly chosen element and then refill conditionally according to μ λ \mu_\lambda μ λ . Writing C C C for the parallel classes, one has
ker ( P d u − ( 1 − 1 / r ) I ) = { B ↦ ∑ i ∈ B h i : ∑ i ∈ C λ i h i = 0 for every C } . (27) \ker\left(P_{\mathrm{du}}-(1-1/r)I\right)
=\left\{B\mapsto\sum_{i\in B}h_i:\sum_{i\in C}\lambda_i h_i=0\text{ for every }C\right\}. \tag*{(27)} ker ( P du − ( 1 − 1/ r ) I ) = { B ↦ i ∈ B ∑ h i : i ∈ C ∑ λ i h i = 0 for every C } . (27) Its dimension is ∑ C ( ∣ C ∣ − 1 ) \sum_C(|C|-1) ∑ C ( ∣ C ∣ − 1 ) . The down-up gap equals 1 / r 1/r 1/ r if and only if M M M has a nontrivial parallel class; otherwise it is strictly larger. In particular every extremal eigenfunction is linear, resolving [2 Conjecture 8.1] for the uniform law and extending the conclusion to all positive fields.
Proof. For r ≥ 2 r\geq2 r ≥ 2 , one has P d u = K r − 1 P_{\mathrm{du}}=K_{r-1} P du = K r − 1 . The proof of Theorem 5.6 identifies the translation space with the kernel of the weighted class-sum map. Substitution in Theorem 5.7 proves the displayed identity and its dimension. The same theorem gives the gap bound and makes equality equivalent to a nonzero translation direction. If r = 1 r=1 r = 1 , all active elements form one parallel class and each step resamples the entire basis. Thus P d u P_{\mathrm{du}} P du is projection onto constants; its zero eigenspace is precisely the weighted mean-zero functions on the singleton bases, which are the displayed linear functions. □
The exact parallel-class criterion gives the classification sought in [2 Conjecture 8.2] , on the original matroid. Having a parallel pair in some minor is insufficient: U 2 , 3 U_{2,3} U 2 , 3 contracts to U 1 , 2 U_{1,2} U 1 , 2 , but its down-up gap is 3 / 4 > 1 / 2 3/4>1/2 3/4 > 1/2 . Indeed, its three-state transition matrix has diagonal entries 1 / 2 1/2 1/2 and off-diagonal entries 1 / 4 1/4 1/4 .
The gap characterization also follows by combining the product bound of Alev–Lau [3 Theorem 1.5] with the strict Hessian theorem of Murai–Nagaoka–Yazawa [38 ] . Theorem 5.7 identifies every extremal mode and extends the conclusion to general Lorentzian occupation laws.
The residual estimate also separates the slow part of the continuous flow. Let Π N \Pi_N Π N be the orthogonal projection onto the linear modes in Theorem 5.7 . For t ≥ 0 t \ge0 t ≥ 0 and every mean-zero g g g ,
∥ e t ( K r − 1 − I ) g − e − t / r Π N g ∥ 2 ≤ e − ( 1 + κ ) t / r ∥ g − Π N g ∥ 2 . \left\|e^{t(K_{r-1}-I)}g-e^{-t/r}\Pi_N g\right\|_2 \le e^{-(1+\kappa)t/r}\left\|g-\Pi_N g\right\|_2. e t ( K r − 1 − I ) g − e − t / r Π N g 2 ≤ e − ( 1 + κ ) t / r ∥ g − Π N g ∥ 2 .
Indeed, the positive generator I − K r − 1 I-K_{r-1} I − K r − 1 equals 1 / r 1/r 1/ r on the projected space and is at least ( 1 + κ ) / r (1+\kappa)/r ( 1 + κ ) / r on its orthogonal complement; apply its spectral decomposition.
Retained-level reconstruction and sharp stability
Average the retained conditional logarithms and correct for their common mean. The resulting Gibbs candidate expresses the deletion deficit as relative entropy plus a normalization defect, giving sharp stability and self-consistency even for singular inputs.
Let f f f be Lorentzian of degree r ≥ 2 r \ge2 r ≥ 2 , absorb the positive field, normalize f ( 1 ) = 1 f(1)=1 f ( 1 ) = 1 , and remove inactive variables. Write
ρ = ∑ α μ ( α ) ∣ α ⟩ ⟨ α ∣ ⊗ F α , σ = E μ F , Tr σ = 1 , \rho=\sum_{\alpha}\mu(\alpha)|\alpha\rangle\langle\alpha|\otimes F_\alpha,\qquad\sigma=\mathbb{E}_{\mu}F,\qquad\operatorname{Tr}\sigma=1, ρ = α ∑ μ ( α ) ∣ α ⟩ ⟨ α ∣ ⊗ F α , σ = E μ F , Tr σ = 1 ,
and put G i = F 1 ( i ) G_i=F_1(i) G i = F 1 ( i ) and p i = μ 1 ( i ) p_i=\mu_1(i) p i = μ 1 ( i ) . The exponential variational identities below belong to the framework of quantum entropy remainders and Brascamp–Lieb duality [14 , 10 ] . Here all their matrix parameters are the observed conditional means.
For positive definite conditional matrices, reconstruction at level ℓ \ell ℓ with reciprocal contraction coefficient c c c has the form
E α = exp ( c ∑ β D r → ℓ ( β ∣ α ) log F ℓ ( β ) − ( c − 1 ) log σ ) . E_\alpha=\exp\left(c\sum_{\beta}D_{r\to\ell}(\beta\mid\alpha)\log F_\ell(\beta)-(c-1)\log\sigma\right). E α = exp c β ∑ D r → ℓ ( β ∣ α ) log F ℓ ( β ) − ( c − 1 ) log σ .
For semidefinite inputs, the same formula is evaluated on the common support of the conditional matrices seen from α \alpha α . Compress the register to supp σ \operatorname{supp}\sigma supp σ , so that σ > 0 \sigma>0 σ > 0 . For A ⪰ 0 A\succeq0 A ⪰ 0 , write log s A \log_s A log s A for its logarithm on supp A \operatorname{supp}A supp A , extended by zero on the kernel. For a projection P P P , exp P \exp_P exp P denotes the exponential on ran P \operatorname{ran}P ran P ; its extension to the whole register is zero on ker P \ker P ker P .
Theorem 6.1 (Reconstruction from a retained level). Fix 1 ≤ ℓ < r 1\le\ell<r 1 ≤ ℓ < r , and let c ≥ 1 c\ge1 c ≥ 1 be a valid reciprocal complete contraction coefficient, meaning H r ( F ) ≥ c H ℓ ( F ℓ ) H_r(F)\ge cH_\ell(F_\ell) H r ( F ) ≥ c H ℓ ( F ℓ ) for every matrix input. For each occupation α \alpha α , let P α P_\alpha P α project onto
⋂ β : D r → ℓ ( β ∣ α ) > 0 supp F ℓ ( β ) , \bigcap_{\beta:\,D_{r\to\ell}(\beta\mid\alpha)>0}\operatorname{supp}F_\ell(\beta), β : D r → ℓ ( β ∣ α ) > 0 ⋂ supp F ℓ ( β ) , and define
K α ( ℓ , c ) = c ∑ β D r → ℓ ( β ∣ α ) log s F ℓ ( β ) − ( c − 1 ) log σ , K_\alpha^{(\ell,c)}=c\sum_{\beta}D_{r\to\ell}(\beta\mid\alpha)\log_s F_\ell(\beta)-(c-1)\log\sigma, K α ( ℓ , c ) = c β ∑ D r → ℓ ( β ∣ α ) log s F ℓ ( β ) − ( c − 1 ) log σ , E α ( ℓ , c ) = P α exp P α ( P α K α ( ℓ , c ) P α ) P α , E_\alpha^{(\ell,c)}=P_\alpha\exp_{P_\alpha}\left(P_\alpha K_\alpha^{(\ell,c)}P_\alpha\right)P_\alpha, E α ( ℓ , c ) = P α exp P α ( P α K α ( ℓ , c ) P α ) P α , Z ℓ , c = E μ Tr E α ( ℓ , c ) , τ ℓ , c = Z ℓ , c − 1 ∑ α μ ( α ) ∣ α ⟩ ⟨ α ∣ ⊗ E α ( ℓ , c ) . Z_{\ell,c}=\mathbb{E}_{\mu}\operatorname{Tr}E_\alpha^{(\ell,c)},\qquad
\tau_{\ell,c}=Z_{\ell,c}^{-1}\sum_{\alpha}\mu(\alpha)|\alpha\rangle\langle\alpha|\otimes E_\alpha^{(\ell,c)}. Z ℓ , c = E μ Tr E α ( ℓ , c ) , τ ℓ , c = Z ℓ , c − 1 α ∑ μ ( α ) ∣ α ⟩ ⟨ α ∣ ⊗ E α ( ℓ , c ) . Then 0 < Z ℓ , c ≤ 1 0<Z_{\ell,c}\le1 0 < Z ℓ , c ≤ 1 and
H r − c H ℓ = D ( ρ ∥ τ ℓ , c ) − log Z ℓ , c . (28) H_r-cH_\ell=D(\rho\Vert\tau_{\ell,c})-\log Z_{\ell,c}. \tag*{(28)} H r − c H ℓ = D ( ρ ∥ τ ℓ , c ) − log Z ℓ , c . (28) Consequently
∥ ρ − τ ℓ , c ∥ 1 ≤ 2 ( H r − c H ℓ ) , Z ℓ , c ≥ e − ( H r − c H ℓ ) . \|\rho-\tau_{\ell,c}\|_1\le\sqrt{2(H_r-cH_\ell)},\qquad Z_{\ell,c}\ge e^{-(H_r-cH_\ell)}. ∥ ρ − τ ℓ , c ∥ 1 ≤ 2 ( H r − c H ℓ ) , Z ℓ , c ≥ e − ( H r − c H ℓ ) . Equality H r = c H ℓ H_r=cH_\ell H r = c H ℓ holds exactly when Z ℓ , c = 1 Z_{\ell,c}=1 Z ℓ , c = 1 and ρ = τ ℓ , c \rho=\tau_{\ell,c} ρ = τ ℓ , c . For positive definite inputs, every P α = I P_\alpha=I P α = I and the formula is the ordinary exponential of K α ( ℓ , c ) K_\alpha^{(\ell,c)} K α ( ℓ , c ) .
For the universal choice c = r / ℓ c=r/\ell c = r / ℓ , writing δ ℓ = ℓ H r / r − H ℓ \delta_\ell=\ell H_r/r-H_\ell δ ℓ = ℓ H r / r − H ℓ , this gives
∥ ρ − τ ℓ , r / ℓ ∥ 1 ≤ 2 r ℓ δ ℓ . (29) \|\rho-\tau_{\ell,r/\ell}\|_1\le\sqrt{\frac{2r}{\ell}\,\delta_\ell}. \tag*{(29)} ∥ ρ − τ ℓ , r / ℓ ∥ 1 ≤ ℓ 2 r δ ℓ . (29) Under (3.3), one may instead use c = a r / a ℓ c=a_r/a_\ell c = a r / a ℓ and the improved deficit H r − ( a r / a ℓ ) H ℓ H_r-(a_r/a_\ell)H_\ell H r − ( a r / a ℓ ) H ℓ .
Conversely, start with any positive semidefinite matrices G β G_\beta G β of mean σ = E μ ℓ G \sigma= \mathbb{E}_{\mu_\ell}G σ = E μ ℓ G , with Tr σ = 1 \operatorname{Tr}\sigma= 1 Tr σ = 1 , and form the same exponentials with F ℓ ( β ) F_\ell(\beta) F ℓ ( β ) replaced by G β G_\beta G β . If Z ℓ , c = 1 Z_{\ell,c} = 1 Z ℓ , c = 1 , then
E μ E ( ℓ , c ) = σ , T ℓ E ( ℓ , c ) ( β ) = G β . \mathbb{E}_\mu E^{(\ell,c)} = \sigma,\qquad T_\ell E^{(\ell,c)}(\beta) = G_\beta. E μ E ( ℓ , c ) = σ , T ℓ E ( ℓ , c ) ( β ) = G β .
Thus trace saturation is sufficient for self-consistency.
Proof. The conditional mean F ℓ ( β ) F_\ell(\beta) F ℓ ( β ) is a positive mixture containing F α F_\alpha F α whenever D r → ℓ ( β ∣ α ) > 0 D_{r\to\ell}(\beta\mid\alpha) > 0 D r → ℓ ( β ∣ α ) > 0 . Therefore supp F α ⊆ ran P α \operatorname{supp} F_\alpha\subseteq\operatorname{ran} P_\alpha supp F α ⊆ ran P α . At least one F α F_\alpha F α is nonzero, so Z ℓ , c > 0 Z_{\ell,c} > 0 Z ℓ , c > 0 .
For a Hermitian H H H and a projection P P P ,
Tr P exp P ( P H P ) ≤ Tr exp H . \operatorname{Tr}_P \exp_P(PHP) \leq\operatorname{Tr}\exp H. Tr P exp P ( P H P ) ≤ Tr exp H .
Indeed, in an eigenbasis of P H P PHP P H P , scalar Jensen gives e ⟨ u , H u ⟩ ≤ ⟨ u , e H u ⟩ e^{\langle u,Hu\rangle} \leq\langle u,e^H u\rangle e ⟨ u , H u ⟩ ≤ ⟨ u , e H u ⟩ ; summing proves the inequality. Apply it to the exponent obtained by replacing every F ℓ ( β ) F_\ell(\beta) F ℓ ( β ) with F ℓ ( β ) + ε I F_\ell(\beta) + \varepsilon I F ℓ ( β ) + ε I and σ \sigma σ with σ + ε I \sigma+ \varepsilon I σ + ε I . The normalized trace inequality of Theorem 4.1 bounds the sum of the uncompressed exponential traces by 1 + ε dim Q 1 + \varepsilon\dim Q 1 + ε dim Q . On ran P α \operatorname{ran} P_\alpha ran P α , the compressed logarithms converge to those defining K α ( ℓ , c ) K_\alpha^{(\ell,c)} K α ( ℓ , c ) . Passing to the limit therefore proves Z ℓ , c ≤ 1 Z_{\ell,c} \leq1 Z ℓ , c ≤ 1 .
The support inclusion also gives Tr ( F α log E α ( ℓ , c ) ) = Tr ( F α K α ( ℓ , c ) ) \operatorname{Tr}(F_\alpha\log E_\alpha^{(\ell,c)}) = \operatorname{Tr}(F_\alpha K_\alpha^{(\ell,c)}) Tr ( F α log E α ( ℓ , c ) ) = Tr ( F α K α ( ℓ , c ) ) . Adjointness of deletion now gives
E μ Tr ( F α log E α ( ℓ , c ) ) = c E μ ℓ Tr ( F ℓ log F ℓ ) − ( c − 1 ) Tr ( σ log σ ) . \mathbb{E}_\mu\operatorname{Tr}(F_\alpha\log E_\alpha^{(\ell,c)}) = c\mathbb{E}_{\mu_\ell}\operatorname{Tr}(F_\ell\log F_\ell) - (c-1)\operatorname{Tr}(\sigma\log\sigma). E μ Tr ( F α log E α ( ℓ , c ) ) = c E μ ℓ Tr ( F ℓ log F ℓ ) − ( c − 1 ) Tr ( σ log σ ) .
Expanding the relative entropy proves (28) . Its two terms on the right are nonnegative. Quantum Pinsker and their separate bounds prove the stability, normalization, and equality assertions. The choices of c c c follow from Theorems 2.1 and 3.4 .
For the converse abbreviate E = E ( ℓ , c ) E = E^{(\ell,c)} E = E ( ℓ , c ) , put N = E μ E N = \mathbb{E}_\mu E N = E μ E and J = T ℓ E J = T_\ell E J = T ℓ E , and use D g ( A ∥ B ) = Tr [ A ( log A − log B ) − A + B ] D_g(A\mathbin{\|}B) = \operatorname{Tr}[A(\log A - \log B) - A + B] D g ( A ∥ B ) = Tr [ A ( log A − log B ) − A + B ] . By construction supp J β ⊆ supp G β \operatorname{supp} J_\beta\subseteq\operatorname{supp} G_\beta supp J β ⊆ supp G β , so these relative entropies are finite. Since Tr N = Z ℓ , c = 1 \operatorname{Tr} N = Z_{\ell,c} = 1 Tr N = Z ℓ , c = 1 , the same expansion gives
H r ( E ) − c H ℓ ( J ) = − c E μ ℓ D g ( J β ∥ G β ) + ( c − 1 ) D g ( N ∥ σ ) ≤ − D g ( N ∥ σ ) . H_r(E) - cH_\ell(J) = -c\mathbb{E}_{\mu_\ell}D_g(J_\beta\mathbin{\|}G_\beta) + (c-1)D_g(N\mathbin{\|}\sigma) \leq-D_g(N\mathbin{\|}\sigma). H r ( E ) − c H ℓ ( J ) = − c E μ ℓ D g ( J β ∥ G β ) + ( c − 1 ) D g ( N ∥ σ ) ≤ − D g ( N ∥ σ ) .
The last inequality is joint convexity. Complete contraction makes the left side nonnegative, so N = σ N = \sigma N = σ . The identity then forces D g ( J β ∥ G β ) = 0 D_g(J_\beta\mathbin{\|}G_\beta) = 0 D g ( J β ∥ G β ) = 0 for every positive-mass output, which gives J β = G β J_\beta= G_\beta J β = G β . □
Corollary 6.2 (Sharpness of retained-level stability). For every r ≥ 2 r \geq2 r ≥ 2 and 1 ≤ ℓ < r 1 \leq\ell< r 1 ≤ ℓ < r , the coefficient 2 r / ℓ 2r/\ell 2 r / ℓ in (29) is optimal uniformly over Lorentzian laws, even when the classical marginal is fixed to the reference and the encoding register is a classical bit.
Proof. Take the uniform basis law of a partition matroid with r r r binary classes, identified with independent unbiased bits X = ( X 1 , … , X r ) X = (X_1,\ldots,X_r) X = ( X 1 , … , X r ) . Put χ ( X ) = ( − 1 ) X 1 + ⋯ + X r \chi(X) = (-1)^{X_1+\cdots+X_r} χ ( X ) = ( − 1 ) X 1 + ⋯ + X r and, for 0 < t < 1 0 < t < 1 0 < t < 1 , encode X = x X = x X = x by
F x = 1 2 ( 1 + t χ ( x ) 0 0 1 − t χ ( x ) ) . F_x = \frac{1}{2}\begin{pmatrix} 1+t\chi(x) & 0 \\ 0 & 1-t\chi(x) \end{pmatrix}. F x = 2 1 ( 1 + t χ ( x ) 0 0 1 − t χ ( x ) ) . The mean is σ = I / 2 \sigma= I/2 σ = I /2 . Every proper retention omits at least one independent bit, so F ℓ = σ F_\ell= \sigma F ℓ = σ and H ℓ = 0 H_\ell= 0 H ℓ = 0 for ℓ < r \ell< r ℓ < r . Its reconstruction is exactly τ ℓ , r / ℓ = μ ⊗ σ \tau_{\ell,r/\ell} = \mu\otimes\sigma τ ℓ , r / ℓ = μ ⊗ σ . Therefore
∥ ρ − τ ℓ , r / ℓ ∥ 1 = t , H r = ( 1 + t ) log ( 1 + t ) + ( 1 − t ) log ( 1 − t ) 2 = t 2 2 + O ( t 4 ) . \lVert\rho-\tau_{\ell,r/\ell}\rVert_1 = t,\qquad H_r = \frac{(1+t)\log(1+t)+(1-t)\log(1-t)}{2} = \frac{t^2}{2}+O(t^4). ∥ ρ − τ ℓ , r / ℓ ∥ 1 = t , H r = 2 ( 1 + t ) log ( 1 + t ) + ( 1 − t ) log ( 1 − t ) = 2 t 2 + O ( t 4 ) . Since δ ℓ = ℓ H r / r \delta_\ell= \ell H_r/r δ ℓ = ℓ H r / r , the ratio of the squared trace distance to δ ℓ \delta_\ell δ ℓ tends to 2 r / ℓ 2r/\ell 2 r / ℓ as t ↓ 0 t\downarrow0 t ↓ 0 . □
Reconstruction at level ℓ \ell ℓ uses the whole conditional family F ℓ F_\ell F ℓ . The next theorem uses only the one-particle family ( G i ) i (G_i)_i ( G i ) i , even when the given deficit is measured at a higher level.
Theorem 6.3 (Gibbs reconstruction and stability). For a positive definite input define
E α = exp ( ∑ i α i log G i − ( r − 1 ) log σ ) , Z = E μ Tr E α , τ = Z − 1 ∑ α μ ( α ) ∣ α ⟩ ⟨ α ∣ ⊗ E α . E_\alpha= \exp\left(\sum_i \alpha_i\log G_i-(r-1)\log\sigma\right),\qquad Z = \mathbb{E}_\mu\operatorname{Tr}E_\alpha,\qquad\tau= Z^{-1}\sum_\alpha\mu(\alpha)|\alpha\rangle\langle\alpha|\otimes E_\alpha. E α = exp ( i ∑ α i log G i − ( r − 1 ) log σ ) , Z = E μ Tr E α , τ = Z − 1 α ∑ μ ( α ) ∣ α ⟩ ⟨ α ∣ ⊗ E α .
For a general input use the support-compressed version of this formula from Theorem 6.1 ; its common support at α \alpha α is ⋂ α i > 0 supp G i \bigcap_{\alpha_i>0}\operatorname{supp}G_i ⋂ α i > 0 supp G i . This is the ℓ = 1 \ell=1 ℓ = 1 , c = r c=r c = r case of Theorem 6.1 , so
H r − r H 1 = D ( ρ ∥ τ ) − log Z , 0 < Z ≤ 1. (30) H_r-rH_1=D(\rho\Vert\tau)-\log Z,\qquad0<Z\leq1. \tag*{(30)} H r − r H 1 = D ( ρ ∥ τ ) − log Z , 0 < Z ≤ 1. (30)
Its saturation and self-consistency conclusions apply. For 1 ≤ ℓ < r 1\leq\ell<r 1 ≤ ℓ < r , let δ ℓ = ℓ H r / r − H ℓ \delta_\ell=\ell H_r/r-H_\ell δ ℓ = ℓ H r / r − H ℓ . Then
D ( ρ ∥ τ ) − log Z ≤ r ( r − 1 ) r − ℓ δ ℓ . (31) D(\rho\Vert\tau)-\log Z\leq\frac{r(r-1)}{r-\ell}\delta_\ell. \tag*{(31)} D ( ρ ∥ τ ) − log Z ≤ r − ℓ r ( r − 1 ) δ ℓ . (31)
∥ ρ − τ ∥ 1 ≤ 2 r ( r − 1 ) r − ℓ δ ℓ . (32) \lVert\rho-\tau\rVert_1\leq\sqrt{\frac{2r(r-1)}{r-\ell}\delta_\ell}. \tag*{(32)} ∥ ρ − τ ∥ 1 ≤ r − ℓ 2 r ( r − 1 ) δ ℓ . (32)
Equality at any proper deletion level is equivalent to Z = 1 Z=1 Z = 1 and ρ = τ \rho=\tau ρ = τ , and every equality state satisfies
( G i − σ ) i ∈ N f ⊗ Herm ( Q ) . (33) (G_i-\sigma)_i\in\mathcal{N}_f\otimes\operatorname{Herm}(Q). \tag*{(33)} ( G i − σ ) i ∈ N f ⊗ Herm ( Q ) . (33)
Proof. Take ℓ = 1 \ell=1 ℓ = 1 and c = r c=r c = r in Theorem 6.1 . Since r D r → 1 ( i ∣ α ) = α i rD_{r\to1}(i\mid\alpha)=\alpha_i r D r → 1 ( i ∣ α ) = α i , its reconstruction is exactly E α E_\alpha E α , and it gives (30) and the self-consistency assertion. The profile chord bound yields
δ ℓ ≥ r − ℓ r ( r − 1 ) ( H r − r H 1 ) . \delta_\ell\geq\frac{r-\ell}{r(r-1)}(H_r-rH_1). δ ℓ ≥ r ( r − 1 ) r − ℓ ( H r − r H 1 ) .
Both terms on the right of (30) are nonnegative. Quantum Pinsker proves the norm estimate, and Theorem 3.1 gives the equality equivalences.
For the null-direction assertion, we record all three quadratic remainders. Let F 2 , i j = F 2 ( e i + e j ) F_{2,ij}=F_2(e_i+e_j) F 2 , ij = F 2 ( e i + e j ) and W = ∇ 2 f ( 1 ) / ( r ( r − 1 ) ) W=\nabla^2f(1)/(r(r-1)) W = ∇ 2 f ( 1 ) / ( r ( r − 1 )) . Let C i j C_{ij} C ij be the support-compressed exponential of log G i + log G j − log σ \log G_i+\log G_j-\log\sigma log G i + log G j − log σ on supp G i ∩ supp G j \operatorname{supp}G_i\cap\operatorname{supp}G_j supp G i ∩ supp G j . Trace compression followed by Lieb’s inequality, with the same regularization as above, gives Tr C i j ≤ Tr [ G i T σ ( G j ) ] \operatorname{Tr}C_{ij}\leq\operatorname{Tr}[G_i\mathcal{T}_\sigma(G_j)] Tr C ij ≤ Tr [ G i T σ ( G j )] . Putting Z i = G i − σ Z_i=G_i-\sigma Z i = G i − σ , direct expansion gives
H 2 − 2 H 1 = ∑ i j W i j D g ( F 2 , i j ∥ C i j ) + ∑ i j W i j ( Tr [ G i T σ ( G j ) ] − Tr C i j ) − ∑ i j W i j Tr [ Z i T σ ( Z j ) ] . (34) \begin{aligned}
H_2-2H_1
&=\sum_{ij}W_{ij}D_g(F_{2,ij}\Vert C_{ij})\\
&\quad+\sum_{ij}W_{ij}\bigl(\operatorname{Tr}[G_i\mathcal{T}_\sigma(G_j)]-\operatorname{Tr}C_{ij}\bigr)\\
&\quad-\sum_{ij}W_{ij}\operatorname{Tr}[Z_i\mathcal{T}_\sigma(Z_j)]. \tag*{(34)}
\end{aligned} H 2 − 2 H 1 = ij ∑ W ij D g ( F 2 , ij ∥ C ij ) + ij ∑ W ij ( Tr [ G i T σ ( G j )] − Tr C ij ) − ij ∑ W ij Tr [ Z i T σ ( Z j )] . (34)
Only pairs with W i j > 0 W_{ij}>0 W ij > 0 are included. The first two sums are nonnegative, and the last term is nonnegative by the Lorentzian signature and ∑ i p i Z i = 0 \sum_i p_iZ_i=0 ∑ i p i Z i = 0 . At equality it follows that
0 ≥ − ∑ i j W i j Tr [ Z i T σ ( Z j ) ] ≥ 0. 0\geq-\sum_{ij}W_{ij}\operatorname{Tr}[Z_i\mathcal{T}_\sigma(Z_j)]\geq0. 0 ≥ − ij ∑ W ij Tr [ Z i T σ ( Z j )] ≥ 0.
Expand the Z i Z_i Z i in a Hermitian basis orthonormal for ⟨ A , B ⟩ σ = Tr [ A T σ ( B ) ] \langle A,B\rangle_\sigma=\operatorname{Tr}[A\mathcal{T}_\sigma(B)] ⟨ A , B ⟩ σ = Tr [ A T σ ( B )] . Every coefficient vector is p p p -centered and has zero quadratic form under W W W . The Lorentzian signature puts it in ker W = N f \ker W=\mathcal{N}_f ker W = N f , proving (33) . □
The common-support convention makes both reconstruction identities exact for singular inputs as well. It also keeps every term in (34) finite: a positive-mass F 2 , i j F_{2,ij} F 2 , ij is supported in both G i G_i G i and G j G_j G j .
Quadratic models of quantum entropy equality
The reconstruction theorem makes an equality density unique once its one-particle matrices are known. We now decide exactly which matrices can occur. Only their two-particle compatibility matters, in every degree and every finite matrix dimension.
Normalize f ( 1 ) = 1 f(1)=1 f ( 1 ) = 1 , remove inactive labels, and suppose its degree is r ≥ 2 r\geq2 r ≥ 2 . Put W = ∇ 2 f ( 1 ) / ( r ( r − 1 ) ) W=\nabla^2f(1)/(r(r-1)) W = ∇ 2 f ( 1 ) / ( r ( r − 1 )) and p = W 1 p=W1 p = W 1 . Write E f \mathcal{E}_f E f for the positive semidefinite densities with Tr E μ F = 1 \operatorname{Tr}E_\mu F=1 Tr E μ F = 1 attaining equality at a proper deletion level. By Theorem 3.1 , these are exactly the densities with H k = k H 1 H_k=kH_1 H k = k H 1 at every level.
Theorem 7.1 (Quadratic extension and the complete equality model). Taking the two-particle conditional density gives a bijection
T 2 : E f ⟶ ∼ E f 2 , f 2 ( z ) = z T W z . T_2:\mathcal{E}_f \overset{\sim}{\longrightarrow} \mathcal{E}_{f_2}, \qquad f_2(z)=z^{\mathsf{T}}Wz. T 2 : E f ⟶ ∼ E f 2 , f 2 ( z ) = z T W z . More explicitly, let G i ⪰ 0 G_i \succeq0 G i ⪰ 0 , put σ = ∑ i p i G i \sigma=\sum_i p_iG_i σ = ∑ i p i G i , and assume Tr σ = 1 \operatorname{Tr}\sigma=1 Tr σ = 1 . On supp σ \operatorname{supp}\sigma supp σ define
B i = σ − 1 / 2 G i σ − 1 / 2 , L i = σ − 1 / 2 G i 1 / 2 . B_i=\sigma^{-1/2}G_i\sigma^{-1/2}, \qquad L_i=\sigma^{-1/2}G_i^{1/2}. B i = σ − 1/2 G i σ − 1/2 , L i = σ − 1/2 G i 1/2 . These are the one-particle matrices of an equality density if and only if
∑ j W i j G j = p i σ for every i , [ B i , L j ] = 0 whenever W i j > 0. (35) \sum_j W_{ij}G_j=p_i\sigma\quad\text{for every }i,\qquad[B_i,L_j]=0 \quad\text{whenever }W_{ij}>0. \tag*{(35)} j ∑ W ij G j = p i σ for every i , [ B i , L j ] = 0 whenever W ij > 0. (35) Supported self-pairs are included. The unique density is
F α = σ 1 / 2 ( ∏ i B i α i ) σ 1 / 2 . (36) F_\alpha=\sigma^{1/2}\left(\prod_i B_i^{\alpha_i}\right)\sigma^{1/2}. \tag*{(36)} F α = σ 1/2 ( i ∏ B i α i ) σ 1/2 . (36) The factors in every supported occupation commute, so the product is unambiguous and positive. For a quadratic input E i j E_{ij} E ij , its extension uses G i = p i − 1 ∑ j W i j E i j G_i=p_i^{-1}\sum_j W_{ij}E_{ij} G i = p i − 1 ∑ j W ij E ij and satisfies H r ( F ) / r = H 2 ( E ) / 2 = H 1 ( G ) H_r(F)/r=H_2(E)/2=H_1(G) H r ( F ) / r = H 2 ( E ) /2 = H 1 ( G ) .
Consequently any two homogeneous Lorentzian laws with the same normalized pair matrix W W W have canonically identified equality families, preserving σ \sigma σ , all G i G_i G i , and entropy per particle. Their degrees may differ. Here a quadratic density is written E i j = E ( e i + e j ) E_{ij}=E(e_i+e_j) E ij = E ( e i + e j ) , so E i j = E j i E_{ij}=E_{ji} E ij = E j i .
The distinction from reconstruction is existence: every quadratic extremizer extends, and (7.1) is a finite test for arbitrary proposed marginals. We prove it by identifying where non-commuting matrices can occur.
Define an equivalence relation by declaring distinct labels i , j i,j i , j equivalent when W i j = 0 W_{ij}=0 W ij = 0 . Denote its classes by C \mathcal{C} C , and set
p C = ∑ i ∈ C p i , q C ( i ) = p i / p C . p_C=\sum_{i\in C}p_i,\qquad q_C(i)=p_i/p_C. p C = i ∈ C ∑ p i , q C ( i ) = p i / p C . The following lemma proves, in particular, that the declared relation is indeed transitive.
Lemma 7.2 (The missing-pair geometry of a Lorentzian law). If W i j = 0 W_{ij}=0 W ij = 0 for active distinct labels i , j i,j i , j , then
W i i = W j j = 0 , W ( e i p i − e j p j ) = 0. W_{ii}=W_{jj}=0,\qquad W\left(\frac{e_i}{p_i}-\frac{e_j}{p_j}\right)=0. W ii = W j j = 0 , W ( p i e i − p j e j ) = 0. Consequently every class containing more than one label is mutually exclusive: a supported occupation contains at most one particle from that entire class. There is a normalized homogeneous Lorentzian polynomial g g g such that
f ( z ) = g ( ( y C ) C ∈ C ) , y C = ∑ i ∈ C q C ( i ) z i . (37) f(z)=g((y_C)_{C\in\mathcal{C}}),\qquad y_C=\sum_{i\in C}q_C(i)z_i. \tag*{(37)} f ( z ) = g (( y C ) C ∈ C ) , y C = i ∈ C ∑ q C ( i ) z i . (37) Its Hessian has strictly positive entries between every two distinct classes. Each class is either exclusive, meaning that g g g has degree at most one in y C y_C y C , or a singleton whose label can occur repeatedly.
Proof. Put h = e i / p i − e j / p j h=e_i/p_i-e_j/p_j h = e i / p i − e j / p j . It lies in p ⊥ p^\perp p ⊥ , where the form W W W is negative semidefinite. Since W i j = 0 W_{ij}=0 W ij = 0 ,
0 ≥ h T W h = W i i p i 2 + W j j p j 2 ≥ 0. 0\ge h^{\mathsf{T}}Wh=\frac{W_{ii}}{p_i^2}+\frac{W_{jj}}{p_j^2}\ge0. 0 ≥ h T W h = p i 2 W ii + p j 2 W j j ≥ 0. Thus both diagonal entries vanish and h h h is null for the restricted form. Polarization makes it W W W -orthogonal to p ⊥ p^\perp p ⊥ . It is also W W W -orthogonal to 1 \mathbf{1} 1 , since W 1 = p W\mathbf{1}=p W 1 = p , so W h = 0 Wh=0 W h = 0 . The rows W i ∙ / p i W_{i\bullet}/p_i W i ∙ / p i and W j ∙ / p j W_{j\bullet}/p_j W j ∙ / p j are therefore equal. This immediately proves transitivity of the missing-pair relation.
Nonnegative coefficients show that a vanishing diagonal prohibits a repetition, while W i j = 0 W_{ij}=0 W ij = 0 prohibits simultaneous occurrence. Lemma 5.1 upgrades each displayed kernel vector to
1 p i ∂ i f = 1 p j ∂ j f . \frac{1}{p_i}\partial_i f=\frac{1}{p_j}\partial_j f. p i 1 ∂ i f = p j 1 ∂ j f .
Thus f f f depends on variables inside each class only through the stated weighted sum. One obtains g g g by setting all variables in a class equal; closure under nonnegative linear substitution makes g g g Lorentzian. The remaining claims follow directly from the definition of the classes. □
Write E ⊆ C \mathcal{E}\subseteq\mathcal{C} E ⊆ C for the exclusive classes. For an occupation α \alpha α , let γ C = ∑ i ∈ C α i \gamma_C=\sum_{i\in C}\alpha_i γ C = ∑ i ∈ C α i . For C ∈ E C\in\mathcal{E} C ∈ E with γ C = 1 \gamma_C=1 γ C = 1 , write x C x_C x C for the selected representative. The reference law is the quotient occupation law ν ( γ ) = [ y γ ] g ( y ) \nu(\gamma)=[y^\gamma]g(y) ν ( γ ) = [ y γ ] g ( y ) followed by independent representative choices q C q_C q C in occupied exclusive classes.
We first derive the row and commutator conditions (7.1) from equality. For the converse we start with arbitrary matrices satisfying those conditions. Their class means need not agree: the relative class densities first determine central blocks, preserved by the faithful mean state. Scalar normalization inside each block then gives ensembles with a common mean, to which the Markov gluing lemma applies. Translation invariance verifies the candidate’s conditional means and entropy profile; Gibbs reconstruction gives uniqueness.
Recovery on supported pairs
Vanishing of the quadratic remainders gives equality in relative-entropy data processing. A faithful invariant state then turns recovery into commutation.
Lemma 7.3 (Petz compatibility on every supported pair). Let f f f be any normalized Lorentzian occupation law of degree r ≥ 2 r\geq2 r ≥ 2 . For a matrix density attaining equality in a proper deletion inequality, compress to the support of σ = E F \sigma=EF σ = E F , write G i = T 1 F ( i ) G_i=T_1F(i) G i = T 1 F ( i ) , and set B i = σ − 1 / 2 G i σ − 1 / 2 B_i=\sigma^{-1/2}G_i\sigma^{-1/2} B i = σ − 1/2 G i σ − 1/2 and L i = σ − 1 / 2 G i 1 / 2 L_i=\sigma^{-1/2}G_i^{1/2} L i = σ − 1/2 G i 1/2 . Let W W W be its ordered two-particle law and p = W 1 p=W1 p = W 1 . Then, whenever W i j > 0 W_{ij}>0 W ij > 0 ,
F 2 ( e i + e j ) = G j 1 / 2 σ − 1 / 2 G i σ − 1 / 2 G j 1 / 2 , (38) F_2(e_i+e_j)=G_j^{1/2}\sigma^{-1/2}G_i\sigma^{-1/2}G_j^{1/2}, \tag*{(38)} F 2 ( e i + e j ) = G j 1/2 σ − 1/2 G i σ − 1/2 G j 1/2 , (38) [ B i , L j ] = 0. (39) [B_i,L_j]=0. \tag*{(39)} [ B i , L j ] = 0. (39) In particular B i B_i B i and B j B_j B j commute on every supported pair. The statement includes i = j i=j i = j whenever repeated labels are supported.
Proof. Equality at a proper retained level makes the entropy profile linear by Theorem 3.1, so H 2 − 2 H 1 = 0 H_2-2H_1=0 H 2 − 2 H 1 = 0 . The three nonnegative terms in (6.7) give
( G i − σ ) i ∈ N f ⊗ Herm ( Q ) (G_i-\sigma)_i\in\mathcal{N}_f\otimes\operatorname{Herm}(Q) ( G i − σ ) i ∈ N f ⊗ Herm ( Q )
and identify F i j : = F 2 ( e i + e j ) F_{ij}:=F_2(e_i+e_j) F ij := F 2 ( e i + e j ) with the support-compressed exponential of log G i + log G j − log σ \log G_i+\log G_j-\log\sigma log G i + log G j − log σ . Thus
Tr F i j log F i j = Tr F i j ( log σ G i + log σ G j − log σ ) . (40) \operatorname{Tr}F_{ij}\log F_{ij}
=\operatorname{Tr}F_{ij}(\log_{\sigma}G_i+\log_{\sigma}G_j-\log\sigma). \tag*{(40)} Tr F ij log F ij = Tr F ij ( log σ G i + log σ G j − log σ ) . (40)
The support of F i j F_{ij} F ij is contained in the supports of both G i G_i G i and G j G_j G j .
Put P i j = W i j / p i P_{ij}=W_{ij}/p_i P ij = W ij / p i . Translation nullity gives ∑ j P i j G j = σ \sum_jP_{ij}G_j=\sigma ∑ j P ij G j = σ , while conditional expectation gives ∑ j P i j F i j = G i \sum_jP_{ij}F_{ij}=G_i ∑ j P ij F ij = G i . For
D g ( A ∥ B ) = Tr [ A ( log A − log B ) − A + B ] , D_g(A\mathbin\Vert B)=\operatorname{Tr}[A(\log A-\log B)-A+B], D g ( A ∥ B ) = Tr [ A ( log A − log B ) − A + B ] ,
expanding (7.6) therefore yields the exact joint-convexity equality
∑ j P i j D g ( F i j ∥ G j ) = D g ( G i ∥ σ ) . (41) \sum_jP_{ij}D_g(F_{ij}\mathbin\Vert G_j)=D_g(G_i\mathbin\Vert\sigma). \tag*{(41)} j ∑ P ij D g ( F ij ∥ G j ) = D g ( G i ∥ σ ) . (41)
If G i = 0 G_i=0 G i = 0 , all F i j F_{ij} F ij on this row vanish and both conclusions are immediate. Otherwise normalize by Tr G i \operatorname{Tr}G_i Tr G i and apply the equality case of relative-entropy data processing [27 Theorem 3] to the channel that forgets the classical index j j j . The Petz recovery map gives (7.4). This remains valid for singular G j G_j G j because the recovered reference mean σ \sigma σ is faithful. Consequently B i B_i B i is fixed by the unital completely positive map
Φ i ( X ) = ∑ j P i j L j X L j ∗ . \Phi_i(X)=\sum_j P_{ij}L_jXL_j^*. Φ i ( X ) = j ∑ P ij L j X L j ∗ .
Indeed, the recovery identity averaged over j j j gives Φ i ( B i ) = B i \Phi_i(B_i)=B_i Φ i ( B i ) = B i ; ∑ j P i j G j = σ \sum_jP_{ij}G_j=\sigma ∑ j P ij G j = σ gives unitality; and
Tr [ σ Φ i ( X ) ] = Tr ( σ X ) \operatorname{Tr}[\sigma\Phi_i(X)]=\operatorname{Tr}(\sigma X) Tr [ σ Φ i ( X )] = Tr ( σ X )
shows that σ \sigma σ is a faithful invariant state. Kadison–Schwarz and invariance imply Φ i ( B i 2 ) − B i 2 = 0 \Phi_i(B_i^2)-B_i^2=0 Φ i ( B i 2 ) − B i 2 = 0 . Equivalently,
∑ j P i j ( B i L j − L j B i ) ( B i L j − L j B i ) ∗ = 0. \sum_j P_{ij}(B_iL_j-L_jB_i)(B_iL_j-L_jB_i)^*=0. j ∑ P ij ( B i L j − L j B i ) ( B i L j − L j B i ) ∗ = 0.
Every term is positive semidefinite, proving (39) . Since B j = L j L j ∗ B_j=L_jL_j^* B j = L j L j ∗ , the two relative densities commute as well.
Pairwise compatibility gives a global quantum Markov state.
Lemma 7.4 (A common Markov extension of the class ensembles). Suppose a family of ensembles { q C ( i ) , G i } i ∈ C \{q_C(i),G_i\}_{i\in C} { q C ( i ) , G i } i ∈ C has common faithful mean σ \sigma σ and satisfies [ B i , L j ] = 0 [B_i,L_j]=0 [ B i , L j ] = 0 whenever i , j i,j i , j belong to distinct classes. For each set J J J of classes, put
F J ( x J ) = σ 1 / 2 ∏ C ∈ J B x C σ 1 / 2 , Ω X J Q = ∑ x J ( ∏ C ∈ J q C ( x C ) ) ∣ x J ⟩ ⟨ x J ∣ ⊗ F J ( x J ) . F_J(x_J)=\sigma^{1/2}\prod_{C\in J}B_{x_C}\sigma^{1/2},\qquad
\Omega_{X_JQ}=\sum_{x_J}\left(\prod_{C\in J}q_C(x_C)\right)|x_J\rangle\langle x_J|\otimes F_J(x_J). F J ( x J ) = σ 1/2 C ∈ J ∏ B x C σ 1/2 , Ω X J Q = x J ∑ ( C ∈ J ∏ q C ( x C ) ) ∣ x J ⟩ ⟨ x J ∣ ⊗ F J ( x J ) .
These are consistent states with marginal σ \sigma σ . For every C ∉ J C\notin J C ∈ / J ,
I ( X C : X J ∣ Q ) Ω = 0. I(X_C:X_J\mid Q)_\Omega=0. I ( X C : X J ∣ Q ) Ω = 0.
Thus the full state satisfies
∑ C S ( X C ∣ Q ) Ω − S ( X C ∣ Q ) Ω = 0 , \sum_C S(X_C\mid Q)_\Omega-S(X_C\mid Q)_\Omega=0, C ∑ S ( X C ∣ Q ) Ω − S ( X C ∣ Q ) Ω = 0 ,
and admits a decomposition
Q = ⨁ z ⨂ C Q C z , Ω X C Q = ⨁ z w z ⨂ C Ω X C Q C z ( z ) . Q=\bigoplus_z\bigotimes_C Q_C^z,\qquad
\Omega_{X_CQ}=\bigoplus_z w_z\bigotimes_C\Omega_{X_CQ_C^z}^{(z)}. Q = z ⨁ C ⨂ Q C z , Ω X C Q = z ⨁ w z C ⨂ Ω X C Q C z ( z ) .
Proof. The commutator condition implies pairwise commutation of relative densities from distinct classes, so every displayed product is positive. The relations ∑ i q C ( i ) B i = I \sum_iq_C(i)B_i=I ∑ i q C ( i ) B i = I show consistency and normalization. For each class define the Petz channel
R C ( X ) = ∑ i ∈ C q C ( i ) ∣ i ⟩ ⟨ i ∣ ⊗ G i 1 / 2 σ − 1 / 2 X σ − 1 / 2 G i 1 / 2 . \mathcal{R}_C(X)=\sum_{i\in C}q_C(i)|i\rangle\langle i|\otimes G_i^{1/2}\sigma^{-1/2}X\sigma^{-1/2}G_i^{1/2}. R C ( X ) = i ∈ C ∑ q C ( i ) ∣ i ⟩ ⟨ i ∣ ⊗ G i 1/2 σ − 1/2 X σ − 1/2 G i 1/2 .
It is trace preserving because ∑ i q C ( i ) G i = σ \sum_iq_C(i)G_i=\sigma ∑ i q C ( i ) G i = σ . If C ∉ J C\notin J C ∈ / J , the product ∏ D ∈ J B x D \prod_{D\in J}B_{x_D} ∏ D ∈ J B x D commutes with L i L_i L i and L i ∗ L_i^* L i ∗ for every i ∈ C i\in C i ∈ C . It follows directly that
( id X J ⊗ R C ) ( Ω X J Q ) = Ω X J ∪ { C } Q , R C ( σ ) = Ω X C Q . (\operatorname{id}_{X_J}\otimes\mathcal{R}_C)(\Omega_{X_JQ})=\Omega_{X_{J\cup\{C\}}Q},\qquad
\mathcal{R}_C(\sigma)=\Omega_{X_CQ}. ( id X J ⊗ R C ) ( Ω X J Q ) = Ω X J ∪ { C } Q , R C ( σ ) = Ω X C Q .
The same channel therefore recovers both the joint state and the reference state in relative-entropy data processing under Tr X C \operatorname{Tr}_{X_C} Tr X C . Equality gives I ( X C : X J ∣ Q ) = 0 I(X_C:X_J\mid Q)=0 I ( X C : X J ∣ Q ) = 0 . The chain rule proves vanishing of the full conditional total correlation. In particular,
TC ( X 1 , … , X m ∣ Q ) = I ( X 1 : X 2 ⋯ X m ∣ Q ) + TC ( X 2 , … , X m ∣ Q ) , \operatorname{TC}(X_1,\ldots,X_m\mid Q)=I(X_1:X_2\cdots X_m\mid Q)+\operatorname{TC}(X_2,\ldots,X_m\mid Q), TC ( X 1 , … , X m ∣ Q ) = I ( X 1 : X 2 ⋯ X m ∣ Q ) + TC ( X 2 , … , X m ∣ Q ) ,
where TC ( X 1 , … , X m ∣ Q ) = ∑ j S ( X j ∣ Q ) − S ( X 1 ⋯ X m ∣ Q ) \operatorname{TC}(X_1,\ldots,X_m\mid Q)=\sum_jS(X_j\mid Q)-S(X_1\cdots X_m\mid Q) TC ( X 1 , … , X m ∣ Q ) = ∑ j S ( X j ∣ Q ) − S ( X 1 ⋯ X m ∣ Q ) .
Apply the Hayden–Jozsa–Petz–Winter equality theorem [27 Theorem 6] to X C 1 X_{C_1} X C 1 , Q Q Q , and all remaining classical registers together. It splits Q Q Q into a direct sum of tensor products, separating X C 1 X_{C_1} X C 1 from the remaining registers inside every block. The residual conditional total correlation is an average of the conditional total correlations in those blocks, and each is nonnegative. It therefore vanishes in every positive-weight block. Repeat this argument for the remaining classes. This yields
Ω X C Q = ⨁ z ( ⨂ C Ω X C Q C ( z ) ) . \Omega_{X_CQ}=\bigoplus_z\left(\bigotimes_C\Omega_{X_CQ_C}^{(z)}\right). Ω X C Q = z ⨁ ( C ⨂ Ω X C Q C ( z ) ) .
with all factors normalized. Dephasing the classical registers preserves the state, so each factor is cq. Write Ω X C Q C ( z ) = ∑ i q C ( i ) ∣ i ⟩ ⟨ i ∣ ⊗ A i C , z \Omega_{X_CQ_C}^{(z)}=\sum_i q_C(i)|i\rangle\langle i|\otimes A_i^{C,z} Ω X C Q C ( z ) = ∑ i q C ( i ) ∣ i ⟩ ⟨ i ∣ ⊗ A i C , z ; this defines positive semidefinite A i C , z A_i^{C,z} A i C , z and gives ∑ i q C ( i ) A i C , z = σ C z \sum_i q_C(i)A_i^{C,z}=\sigma_C^z ∑ i q C ( i ) A i C , z = σ C z .
An elementary invariant-state lemma.
Lemma 7.5 (A faithful invariant state respects the Kraus center). Let
Φ ( X ) = ∑ i t i L i X L i ∗ , t i > 0 , \Phi(X)=\sum_i t_iL_iXL_i^*,\qquad t_i>0, Φ ( X ) = i ∑ t i L i X L i ∗ , t i > 0 , be a unital completely positive map with a faithful invariant state σ \sigma σ , so Tr ( σ Φ ( X ) ) = Tr ( σ X ) \operatorname{Tr}(\sigma\Phi(X))=\operatorname{Tr}(\sigma X) Tr ( σ Φ ( X )) = Tr ( σ X ) . Let A = C ∗ ( L i : i ) \mathcal{A}=C^*(L_i:i) A = C ∗ ( L i : i ) . Then σ \sigma σ commutes with every central element of A \mathcal{A} A .
Proof. Kadison–Schwarz, faithfulness, and the commutator sum used in Lemma 7.3 show that the fixed-point space of Φ \Phi Φ is precisely the commutant A ′ \mathcal{A}' A ′ . Let
E ( X ) = lim N → ∞ 1 N ∑ k = 0 N − 1 Φ k ( X ) E(X)=\lim_{N\to\infty}\frac{1}{N}\sum_{k=0}^{N-1}\Phi^k(X) E ( X ) = N → ∞ lim N 1 k = 0 ∑ N − 1 Φ k ( X )
be the finite-dimensional mean ergodic projection onto that space. It preserves the state σ \sigma σ and is an A ′ \mathcal{A}' A ′ -bimodule map, because each Kraus operator commutes with A ′ \mathcal{A}' A ′ . If P P P is a central projection of A \mathcal{A} A , then P ∈ A ′ P\in\mathcal{A}' P ∈ A ′ and P P P commutes with every value of E E E . Therefore
E ( P X ( I − P ) ) = P E ( X ) ( I − P ) = 0. E(PX(I-P))=PE(X)(I-P)=0. E ( P X ( I − P )) = P E ( X ) ( I − P ) = 0.
State invariance gives Tr ( σ P X ( I − P ) ) = 0 \operatorname{Tr}(\sigma PX(I-P))=0 Tr ( σ P X ( I − P )) = 0 for every X X X , proving that σ \sigma σ has no off-diagonal block between P P P and I − P I-P I − P . Spectral decomposition proves the assertion for the whole center. □
The operator normal form.
Theorem 7.6 (Universal operator structure of Lorentzian entropy equality). Let F α ⪰ 0 F_\alpha\succeq0 F α ⪰ 0 , let σ = E μ F \sigma=\mathbb{E}_\mu F σ = E μ F have trace one, and compress to its support. Equality
H μ ( T ℓ F ) = ℓ r H μ ( F ) for some 0 < ℓ < r H_\mu(T_\ell F)=\frac{\ell}{r}H_\mu(F)\qquad\text{for some }0<\ell<r H μ ( T ℓ F ) = r ℓ H μ ( F ) for some 0 < ℓ < r holds if and only if the following structure exists. There are an orthogonal decomposition
Q = ⨁ z ( Q 0 z ⊗ ⨂ C ∈ C Q C z ) , Q=\bigoplus_z\left(Q_0^z\otimes\bigotimes_{C\in\mathcal{C}}Q_C^z\right), Q = z ⨁ ( Q 0 z ⊗ C ∈ C ⨂ Q C z ) , positive weights w z w_z w z summing to one, faithful density matrices τ 0 z , τ C z \tau_0^z,\tau_C^z τ 0 z , τ C z , scalar vectors a z ∈ [ 0 , ∞ ) C a^z\in[0,\infty)^\mathcal{C} a z ∈ [ 0 , ∞ ) C with
a z − 1 ∈ N g , a^z-1\in\mathcal{N}_g, a z − 1 ∈ N g , and positive semidefinite class matrices A i C , z A_i^{C,z} A i C , z satisfying
∑ i ∈ C q C ( i ) A i C , z = τ C z , \sum_{i\in C}q_C(i)A_i^{C,z}=\tau_C^z, i ∈ C ∑ q C ( i ) A i C , z = τ C z , such that
F α = ⨁ z w z ( a z ) γ ( α ) τ 0 z ⊗ ⨂ C ∈ C { A X C C , z , γ C = 1 , τ C z , γ C = 0. (42) F_\alpha=\bigoplus_z w_z(a^z)^{\gamma(\alpha)}\tau_0^z\otimes\bigotimes_{C\in\mathcal{C}}
\begin{cases}
A_{X_C}^{C,z}, & \gamma_C=1,\\
\tau_C^z, & \gamma_C=0.
\end{cases} \tag*{(42)} F α = z ⨁ w z ( a z ) γ ( α ) τ 0 z ⊗ C ∈ C ⨂ { A X C C , z , τ C z , γ C = 1 , γ C = 0. (42)
The conventions are 0 0 = 1 0^0=1 0 0 = 1 and that an empty tensor product is scalar. The mean is
σ = ⨁ z w z τ 0 z ⊗ ⨂ C ∈ E τ C z . \sigma=\bigoplus_z w_z\tau_0^z\otimes\bigotimes_{C\in\mathcal{E}}\tau_C^z. σ = z ⨁ w z τ 0 z ⊗ C ∈ E ⨂ τ C z .
Every such density has equality at every retained level.
Proof of Theorem 7.6. We prove at the same time the sufficiency of (7.1). Start either with an equality density F F F , or directly with matrices G i G_i G i satisfying that test. In the first case Lemma 7.3 and (6.6) give the test. In both cases put B i = σ − 1 / 2 G i σ − 1 / 2 B_i=\sigma^{-1/2}G_i\sigma^{-1/2} B i = σ − 1/2 G i σ − 1/2 , and L i = σ − 1 / 2 G i 1 / 2 L_i=\sigma^{-1/2}G_i^{1/2} L i = σ − 1/2 G i 1/2 . Set
Γ C = ∑ i ∈ C q C ( i ) G i , B ‾ C = σ − 1 / 2 Γ C σ − 1 / 2 = ∑ i ∈ C q C ( i ) L i L i ∗ . \Gamma_C=\sum_{i\in C}q_C(i)G_i,\qquad\overline{B}_C=\sigma^{-1/2}\Gamma_C\sigma^{-1/2}=\sum_{i\in C}q_C(i)L_iL_i^*. Γ C = i ∈ C ∑ q C ( i ) G i , B C = σ − 1/2 Γ C σ − 1/2 = i ∈ C ∑ q C ( i ) L i L i ∗ .
For distinct classes, every pair is supported, so [ B i , L j ] = 0 [B_i,L_j]=0 [ B i , L j ] = 0 whenever i , j i,j i , j belong to different classes. Hence B ‾ C \overline{B}_C B C commutes with every L j L_j L j outside C C C .
Put P i j = W i j / p i P_{ij}=W_{ij}/p_i P ij = W ij / p i . Let W ‾ \overline{W} W and P ‾ C D = W ‾ C D / p C \overline{P}_{CD}=\overline{W}_{CD}/p_C P C D = W C D / p C be the ordered pair law and conditional kernel of g g g . The splitting identity gives P i j = P ‾ C D q D ( j ) P_{ij}=\overline{P}_{CD}q_D(j) P ij = P C D q D ( j ) for i ∈ C , j ∈ D i\in C,j\in D i ∈ C , j ∈ D . Translation nullity and the common mean imply
∑ D P ‾ C D Γ D = σ , ∑ D p D Γ D = σ . \sum_D\overline{P}_{CD}\Gamma_D=\sigma,\qquad\sum_Dp_D\Gamma_D=\sigma. D ∑ P C D Γ D = σ , D ∑ p D Γ D = σ .
For an exclusive class C C C , one has P ‾ C C = 0 \overline{P}_{CC}=0 P C C = 0 , and therefore
p C Γ C = ∑ D ≠ C ( P ‾ C D − p D ) Γ D . p_C\Gamma_C=\sum_{D\ne C}(\overline{P}_{CD}-p_D)\Gamma_D. p C Γ C = D = C ∑ ( P C D − p D ) Γ D .
After conjugating by σ − 1 / 2 \sigma^{-1/2} σ − 1/2 , the right side commutes with every L i L_i L i inside C C C . Thus B ‾ C \overline{B}_C B C commutes with these operators too. For a repeated singleton class, the supported self-pair in (7.1) gives the same commutation directly. Consequently every B ‾ C \overline{B}_C B C lies in the center of A = C ∗ ( L i : i ) \mathcal{A}=C^*(L_i:i) A = C ∗ ( L i : i ) .
The map Φ ( X ) = ∑ i p i L i X L i ∗ \Phi(X)=\sum_i p_iL_iXL_i^* Φ ( X ) = ∑ i p i L i X L i ∗ is unital and has faithful invariant state σ \sigma σ . Lemma 7.5 therefore makes all the B ‾ C \overline{B}_C B C commute with σ \sigma σ . Take their joint spectral decomposition. On a block z z z , write
B ‾ C = a C z I , σ z = w z τ z , Tr τ z = 1. \overline{B}_C=a_C^z I,\qquad\sigma_z=w_z\tau_z,\qquad\operatorname{Tr}\tau_z=1. B C = a C z I , σ z = w z τ z , Tr τ z = 1.
All G i G_i G i preserve these blocks, and Γ C ∣ z = a C z σ z \Gamma_C|_z=a_C^z\sigma_z Γ C ∣ z = a C z σ z . The averaged translation-null identity descends through (7.3) to
( Γ C − σ ) C ∈ N g ⊗ Herm ( Q ) . (\Gamma_C-\sigma)_C\in\mathcal{N}_g\otimes\operatorname{Herm}(Q). ( Γ C − σ ) C ∈ N g ⊗ Herm ( Q ) .
Restriction to z z z and faithfulness of σ z \sigma_z σ z give a C z − 1 ∈ N g a_C^z-1\in\mathcal{N}_g a C z − 1 ∈ N g .
If a C z > 0 a_C^z>0 a C z > 0 , define the normalized ensemble on block z z z by
H i z = G i ∣ z w z a C z , i ∈ C . H_i^z=\frac{G_i|_z}{w_za_C^z},\qquad i\in C. H i z = w z a C z G i ∣ z , i ∈ C .
Its q C q_C q C -mean is τ z \tau_z τ z , and the cross-class Petz commutators are unchanged after this scalar normalization. If a C z = 0 a_C^z=0 a C z = 0 , positivity forces all G i ∣ z = 0 G_i|_z=0 G i ∣ z = 0 in that class; use the harmless constant ensemble H i z = τ z H_i^z=\tau_z H i z = τ z . Repeated singleton classes also have constant ensembles. Lemma 7.4, applied to the exclusive classes, gives a common quantum Markov extension with marginal τ z \tau_z τ z . Its iterated strong-subadditivity decomposition, refining the index z z z if necessary, has the tensor factors and ensembles in (7.9). A spectator factor Q 0 z Q_0^z Q 0 z allows the case of no exclusive classes.
It remains to prove that the proposed form is an equality density with the original one-particle matrices. This will also prove sufficiency. For clarity, fix one block, suppress z z z , and let R J ( x J ) R_J(x_J) R J ( x J ) denote its common quantum Markov family on a set J J J of exclusive classes, with mean τ \tau τ . The candidate block density is
F ~ α = a γ ( α ) R J ( α ) ( x J ( α ) ) , J ( α ) = { C ∈ E : γ C = 1 } . \widetilde{F}_\alpha=a^{\gamma(\alpha)}R_{J(\alpha)}(x_{J(\alpha)}),\qquad J(\alpha)=\{C\in\mathcal{E}:\gamma_C=1\}. F α = a γ ( α ) R J ( α ) ( x J ( α ) ) , J ( α ) = { C ∈ E : γ C = 1 } .
Translation invariance gives, for every multiindex δ \delta δ ,
δ d g ( a ) = δ d g ( 1 ) . (43) \delta^d g(a)=\delta^d g(1). \tag*{(43)} δ d g ( a ) = δ d g ( 1 ) . (43)
In particular g ( a ) = 1 g(a)=1 g ( a ) = 1 , and the tilted quotient law ν a ( γ ) = ν ( γ ) a γ \nu^{a}(\gamma)=\nu(\gamma)a^{\gamma} ν a ( γ ) = ν ( γ ) a γ is a probability law. Fix a supported retained occupation β \beta β , with quotient counts δ \delta δ . Its posterior quotient law weights γ ≥ δ \gamma\ge\delta γ ≥ δ by the falling factorial ( γ ) δ = ∏ C γ C ! / ( γ C − δ C ) ! (\gamma)_{\delta}=\prod_{C}\gamma_{C}!/(\gamma_{C}-\delta_{C})! ( γ ) δ = ∏ C γ C ! / ( γ C − δ C )! . Consequently
E ν [ a γ ∣ δ ] = a δ ∂ δ g ( a ) ∂ δ g ( 1 ) = a δ . \mathbb{E}_{\nu}[a^{\gamma}\mid\delta]=a^{\delta}\frac{\partial^{\delta}g(a)}{\partial^{\delta}g(1)}=a^{\delta}. E ν [ a γ ∣ δ ] = a δ ∂ δ g ( 1 ) ∂ δ g ( a ) = a δ .
Every unretained representative ensemble averages to its class mean, whether that class is absent or present but unretained. The retained class factors remain fixed. Combining these averages gives
T k F ~ ( β ) = a δ ( β ) R J ( β ) ( x J ( β ) ) . (44) T_{k}\widetilde{F}(\beta)=a^{\delta(\beta)}R_{J(\beta)}(x_{J(\beta)}). \tag*{(44)} T k F ( β ) = a δ ( β ) R J ( β ) ( x J ( β ) ) . (44)
This identity includes zero coordinates of a a a by polynomial continuity. At level zero it gives mean τ \tau τ , and at level one it gives a C H i a_{C}H_{i} a C H i , exactly the original one-particle data after restoring w Z w_{Z} w Z .
Let
e C = ∑ i ∈ C q C ( i ) Tr [ H i ( log s H i − log τ ) ] ( C ∈ E ) . e_{C}=\sum_{i\in C}q_{C}(i)\operatorname{Tr}\left[H_{i}(\log_{s}H_{i}-\log\tau)\right]\qquad(C\in\mathcal{E}). e C = i ∈ C ∑ q C ( i ) Tr [ H i ( log s H i − log τ ) ] ( C ∈ E ) .
For fixed δ \delta δ , the representative average of R J ( β ) R_{J(\beta)} R J ( β ) is τ \tau τ , with trace one. The scalar factor therefore contributes Ent ν k ( a δ ) \operatorname{Ent}_{\nu_{k}}(a^{\delta}) Ent ν k ( a δ ) . Additivity of the common Markov family gives ∑ C ∈ J ( β ) e C \sum_{C\in J(\beta)}e_{C} ∑ C ∈ J ( β ) e C for its matrix part. This part is weighted by ν k ( δ ) a δ \nu_{k}(\delta)a^{\delta} ν k ( δ ) a δ , the retained law of ν a \nu^{a} ν a . Thus
H k ( F ~ ) = Ent ν k ( a δ ) + ∑ C ∈ E ( ν a ) k Pr ( C retained ) e C . (45) H_{k}(\widetilde{F})=\operatorname{Ent}_{\nu_{k}}(a^{\delta})+\sum_{C\in\mathcal{E}\,(\nu^{a})_{k}}\Pr(C\text{ retained})e_{C}. \tag*{(45)} H k ( F ) = Ent ν k ( a δ ) + C ∈ E ( ν a ) k ∑ Pr ( C retained ) e C . (45)
Here the tilted retained law equals the retention of ν a \nu^{a} ν a by (7.10). The scalar equality theorem scales the first term by k / r k/r k / r . Exclusive classes contain at most one particle, so their retention probabilities in the second term also scale by k / r k/r k / r . Thus H k = ( k / r ) H r H_{k}=(k/r)H_{r} H k = ( k / r ) H r in every block and hence in their weighted orthogonal sum.
The candidate is therefore an equality density with the prescribed mean σ \sigma σ and one-particle matrices G i G_{i} G i . This proves sufficiency in (35) . If an equality density F F F was given, Theorem 6.3 makes it unique, so F = F ~ F=\widetilde{F} F = F . The same calculation proves sufficiency of every displayed normal-form parameter. □
Completion of the proof of Theorem 7.1 . Necessity of (35) follows from (33) and Lemma 7.3 ; the normal-form construction just proved its sufficiency. In each block the relative one-particle matrix is the scalar a C a_{C} a C times its class likelihood factor. Their product is exactly (36) .
The pair matrix W W W determines p p p , the exclusive classes, their laws q C q_{C} q C , and the quotient pair matrix W ‾ C D = ∑ i ∈ C , j ∈ D W i j \overline{W}_{CD}=\sum_{i\in C,j\in D}W_{ij} W C D = ∑ i ∈ C , j ∈ D W ij . For every law with this pair matrix, Lemma 5.1 gives N g = ker W ‾ \mathcal{N}_{g}=\ker\overline{W} N g = ker W . Thus the admissible normal-form parameters are the same, and their product formula has the same one-particle means. Uniqueness from Gibbs reconstruction makes this correspondence independent of the chosen block decomposition. Taking the second law to be f 2 f_{2} f 2 proves that the correspondence is T 2 T_{2} T 2 and is bijective. Equality gives H r / r = H 1 H_{r}/r=H_{1} H r / r = H 1 in every law. □
Corollary 7.7 (Matroid equality and parallel choices). Let μ \mu μ be a positive product-field basis law of a loopless matroid of rank r ≥ 2 r\ge2 r ≥ 2 . For every normalized matrix entropy equality density, the normal form in Theorem 7.6 has a Z = 1 a^{\mathcal{Z}}=1 a Z = 1 in every block. If I ( B ) I(B) I ( B ) is the basis of the simplification selected by B B B , then
E μ [ F B ∥ I ] = σ , ρ I Q = μ I ⊗ σ . \mathbb{E}_{\mu}[F_{B}\parallel I]=\sigma,\qquad\rho_{IQ}=\mu_{I}\otimes\sigma. E μ [ F B ∥ I ] = σ , ρ I Q = μ I ⊗ σ . Thus equality retains no information about the simplified basis: all nonconstant factors describe choices within parallel classes. This holds even when the input classical marginal differs from μ \mu μ . If the matroid is simple, every equality density is constant.
Proof. The simplified positive-field basis polynomial has nonsingular Hessian, as used in Theorem 5.6 . Hence N g = 0 \mathcal{N}_g=0 N g = 0 , so every quotient tilt is 1 1 1 . Conditioned on I I I , the representatives are independent with laws q C q_C q C . Averaging each class factor in (7.9) gives τ ~ C \widetilde{\tau}^{C} τ C , and therefore the same matrix σ \sigma σ for every simplified basis. For a simple matroid, each class has one representative and its factor is already constant. □
Corollary 7.8 (Quantum equality is classical when no alternatives are exclusive). If every two distinct active labels can occur together, every matrix entropy equality family is simultaneously diagonalizable. More explicitly, there are a unitary U U U , positive weights s a s_a s a summing to one, and vectors v a ≥ 0 v^a \ge0 v a ≥ 0 with v a − 1 ∈ N f v^a-1\in\mathcal{N}_f v a − 1 ∈ N f such that
F α = U diag ( s a ( v a ) α ) a U ∗ . F_\alpha=U\operatorname{diag}\left(s_a(v^a)^\alpha\right)_aU^*. F α = U diag ( s a ( v a ) α ) a U ∗ . This applies in particular to every multiaffine Lorentzian law whose support matroid is simple, and to every multinomial law.
Proof. Every class is a singleton. Its ensemble is constant, so the tensor part of the universal formula is independent of α \alpha α . Diagonalize the fixed faithful matrix in each orthogonal block and absorb its eigenvalues into the scalar weights. □
Fixed classical marginals remove every quotient tilt.
Corollary 7.9 (Fixed-marginal rigidity and exact attainment). Let F F F be an equality density in Theorem 7.6 , and suppose Tr F α = 1 \operatorname{Tr}F_\alpha=1 Tr F α = 1 for every supported occupation. Then every scalar vector in its universal decomposition satisfies a z = 1 a^z=1 a z = 1 . Thus the quotient occupation is independent of the quantum register, and within each classical block all nonconstant tensor factors belong to exclusive classes.
An informative trace-one equality encoding exists in a finite register if and only if two distinct active labels never co-occur, equivalently W i j = 0 W_{ij}=0 W ij = 0 for some i ≠ j i\ne j i = j . A qubit suffices. If every exclusive class is a singleton, the only equality encoding is constant: F α = σ F_\alpha=\sigma F α = σ for every supported occupation. If, in addition, the active Hessian is singular, the fixed-marginal Holevo contraction coefficient is nevertheless ℓ / r \ell/r ℓ / r , and is approached by nonconstant encodings but never attained by one.
Proof. All conditional densities, including the one- and two-particle densities, have trace one. Average the one-particle identity over representatives inside a class. In the universal decomposition this gives
∑ z w z a C z = 1. \sum_z w_z a_C^z=1. z ∑ w z a C z = 1. For every two distinct classes C , D C,D C , D , their joint pair probability is positive. Average the corresponding two-particle identity over the independent representative laws. The common Markov extension has mean one in trace, so
∑ z w z a C z a D z = 1 ( C ≠ D ) . \sum_z w_z a_C^z a_D^z=1\qquad(C\ne D). z ∑ w z a C z a D z = 1 ( C = D ) . Translation nullity also gives the pointwise identity ∑ C p C a C z = 1 \sum_C p_Ca_C^z=1 ∑ C p C a C z = 1 in every block. Regard z z z as a random index with law w z w_z w z . The preceding pair identities say that all distinct a C , a D a_C,a_D a C , a D have zero covariance. Therefore
0 = Var ( ∑ C p C a C ) = ∑ C p C 2 Var ( a C ) . 0=\operatorname{Var}\left(\sum_Cp_Ca_C\right)=\sum_Cp_C^2\operatorname{Var}(a_C). 0 = Var ( C ∑ p C a C ) = C ∑ p C 2 Var ( a C ) . Since every p C > 0 p_C>0 p C > 0 , all a C a_C a C are constant and hence equal to one. The same conclusion is immediate if there is only one class.
When every exclusive class is a singleton, its ensemble has only one member and is constant. The universal formula with a = 1 a=1 a = 1 then gives F α = σ F_\alpha=\sigma F α = σ .
Conversely, let C C C contain at least two labels. Choose positive definite trace-one qubit states A i A_i A i , not all equal, with mean τ = ∑ i ∈ C q C ( i ) A i \tau=\sum_{i\in C}q_C(i)A_i τ = ∑ i ∈ C q C ( i ) A i . Use A i C A_{i_C} A i C when C C C is occupied and τ \tau τ otherwise. This is the normal form with one block, a = 1 a=1 a = 1 , and only the C C C factor nonconstant, so it attains
equality at every level. The class has positive mass and every q C ( i ) > 0 q_C(i)>0 q C ( i ) > 0 , hence its information is positive.
For sharpness under a singular Hessian, choose a nonzero h ∈ N f h\in\mathcal{N}_f h ∈ N f and set u ( α ) = α ⋅ h u(\alpha)=\alpha\cdot h u ( α ) = α ⋅ h . On a qubit put
F ε ( α ) = 1 2 ( 1 + ε u ( α ) 0 0 1 − ε u ( α ) ) . F_{\varepsilon}(\alpha)=\frac{1}{2}\begin{pmatrix}1+\varepsilon u(\alpha)&0\\0&1-\varepsilon u(\alpha)\end{pmatrix}. F ε ( α ) = 2 1 ( 1 + ε u ( α ) 0 0 1 − ε u ( α ) ) .
For small nonzero ε \varepsilon ε these are positive, have trace one, and have common mean I / 2 I/2 I /2 . The second-order entropy expansion and the exact variance mode u u u make their contraction ratio tend to ℓ / r \ell/r ℓ / r . Their information is nonzero, whereas the preceding classification excludes a nonconstant extremizer. □
Remark 7.10 (The pointwise trace constraint). Theorem 7.1 normalizes Tr E F = 1 \operatorname{Tr}\mathbb{E}F=1 Tr E F = 1 . The extra condition Tr F α = 1 \operatorname{Tr}F_\alpha=1 Tr F α = 1 need not transfer between different supports with the same W W W . For example, double each element of the rank-three Fano matroid and of U 3 , 7 U_{3,7} U 3 , 7 into a parallel pair, choosing representatives uniformly. Their uniform quotient laws have the same ordered pair matrix, W ‾ C D = 1 / 42 \overline{W}_{CD}=1/42 W C D = 1/42 for C ≠ D C\ne D C = D . Use a classical register of seven bits, uniform subject to even parity on one Fano line, and representative likelihoods 21 { register bit = chosen bit } 2\mathbf{1}_{\{\text{register bit}=\text{chosen bit}\}} 2 1 { register bit = chosen bit } . Every Fano basis omits a bit of that line, so its product density has trace one. On the uniform-matroid basis equal to the line, the trace is zero or two according to parity. Both are complete-entropy extremizers under (7.2); only the Fano family is pointwise trace one on its whole support.
Power entropies and orthogonal equality structure
Let f ( z ) = ∑ ∣ α ∣ = r μ ( α ) z α f(z)=\sum_{|\alpha|=r}\mu(\alpha)z^\alpha f ( z ) = ∑ ∣ α ∣ = r μ ( α ) z α be normalized, homogeneous, and Lorentzian, with r ≥ 2 r\ge2 r ≥ 2 . Let T ℓ T_\ell T ℓ be conditional expectation after uniform retention of ℓ \ell ℓ particles. For a positive semidefinite matrix-valued function F F F , put
H p ( F ) = Tr E F p − Tr ( E F ) p , 1 < p ≤ 2. \mathcal{H}_p(F)=\operatorname{Tr}\mathbb{E}F^p-\operatorname{Tr}(\mathbb{E}F)^p,\qquad1<p\le2. H p ( F ) = Tr E F p − Tr ( E F ) p , 1 < p ≤ 2.
For any trace function φ \varphi φ used below, write more generally H φ ( F ) = Tr E φ ( F ) − Tr φ ( E F ) \mathcal{H}_{\varphi}(F)=\operatorname{Tr}\mathbb{E}\varphi(F)-\operatorname{Tr}\varphi(\mathbb{E}F) H φ ( F ) = Tr E φ ( F ) − Tr φ ( E F ) . In particular H log \mathcal{H}_{\log} H l o g denotes φ ( x ) = x log x \varphi(x)=x\log x φ ( x ) = x log x . No trace normalization is imposed. The logarithmic equality results apply after positive scalar rescaling; a zero mean forces F = 0 F=0 F = 0 .
Theorem 8.1 (Complete power-entropy contraction). For every matrix dimension, every 1 < p ≤ 2 1<p\le2 1 < p ≤ 2 , and 0 ≤ ℓ ≤ r 0\le\ell\le r 0 ≤ ℓ ≤ r ,
H p ( T ℓ F ) ≤ ℓ r H p ( F ) . \mathcal{H}_p(T_\ell F)\le\frac{\ell}{r}\mathcal{H}_p(F). H p ( T ℓ F ) ≤ r ℓ H p ( F ) . The complete coefficient of a fixed positive-field law equals ℓ / r \ell/r ℓ / r , for 0 < ℓ < r 0<\ell<r 0 < ℓ < r , if and only if its active Hessian is singular.
Proof. For s > 0 s>0 s > 0 set
h s ( x ) = ( x + s ) log ( x + s ) − s log s − ( 1 + log s ) x . h_s(x)=(x+s)\log(x+s)-s\log s-(1+\log s)x. h s ( x ) = ( x + s ) log ( x + s ) − s log s − ( 1 + log s ) x .
For 1 < p < 2 1<p<2 1 < p < 2 one has the positive integral representation
x p = p ( p − 1 ) sin ( π ( p − 1 ) ) π ∫ 0 ∞ s p − 2 h s ( x ) d s . x^p=\frac{p(p-1)\sin(\pi(p-1))}{\pi}\int_0^\infty s^{p-2}h_s(x)\,ds. x p = π p ( p − 1 ) sin ( π ( p − 1 )) ∫ 0 ∞ s p − 2 h s ( x ) d s .
Indeed, both sides and their first derivatives vanish at zero, while their second derivatives agree by
∫ 0 ∞ s p − 2 x + s d s = π x p − 2 sin ( π ( p − 1 ) ) . \int_0^\infty\frac{s^{p-2}}{x+s}\,ds=\frac{\pi x^{p-2}}{\sin(\pi(p-1))}. ∫ 0 ∞ x + s s p − 2 d s = sin ( π ( p − 1 )) π x p − 2 .
The integral converges at zero and infinity, also for x = 0 x=0 x = 0 . Functional calculus and finite expectation give
H p ( F ) = c p ∫ 0 ∞ s p − 2 H log ( F + s I ) d s , \mathcal{H}_p(F)=c_p\int_0^\infty s^{p-2}\mathcal{H}_{\log}(F+sI)\,ds, H p ( F ) = c p ∫ 0 ∞ s p − 2 H l o g ( F + s I ) d s ,
where c p c_p c p is the positive coefficient above and H log ( A ) = Tr E A log A − Tr ( E A ) log ( E A ) \mathcal{H}_{\log}(A)=\operatorname{Tr}\mathbb{E}A\log A-\operatorname{Tr}(\mathbb{E}A)\log(\mathbb{E}A) H l o g ( A ) = Tr E A log A − Tr ( E A ) log ( E A ) . Since T ℓ ( F + s I ) = T ℓ F + s I T_\ell(F+sI)=T_\ell F+sI T ℓ ( F + s I ) = T ℓ F + s I , integration of complete entropy contraction proves the assertion. The case p = 2 p=2 p = 2 is its variance linearization, or follows by continuity as p ↑ 2 p\uparrow2 p ↑ 2 . If the Hessian is nonsingular, Theorem 3.4 improves entropy contraction uniformly in matrix dimension. Apply it to every shift and integrate. If the Hessian is singular, choose nonzero h ∈ N f h \in\mathcal{N}_f h ∈ N f and g ( α ) = α ⋅ h g(\alpha)=\alpha\cdot h g ( α ) = α ⋅ h . For F ε = 1 + ε g > 0 F_\varepsilon=1+\varepsilon g>0 F ε = 1 + ε g > 0 ,
H p ( F ε ) = p ( p − 1 ) 2 ε 2 E g 2 + O ( ε 3 ) . \mathcal{H}_p(F_\varepsilon)=\frac{p(p-1)}{2}\varepsilon^2\mathbb{E}g^2+O(\varepsilon^3). H p ( F ε ) = 2 p ( p − 1 ) ε 2 E g 2 + O ( ε 3 ) .
The variance identity in Theorem 5.7 shows that the contraction ratio tends to ℓ / r \ell/r ℓ / r . Scalar functions already prove complete sharpness. □
Tensor and orthogonal equality
Take the partition matroid with two classes of three elements and uniform representative laws. Choose positive definite matrices A 1 , A 2 , A 3 A_1,A_2,A_3 A 1 , A 2 , A 3 on C d A \mathbb{C}^{d_A} C d A and B 1 , B 2 , B 3 B_1,B_2,B_3 B 1 , B 2 , B 3 on C d B \mathbb{C}^{d_B} C d B with averages I d A I_{d_A} I d A and I d B I_{d_B} I d B . For instance, in dimension two one may take
A 1 = I + t X , A 2 = I + t Z , A 3 = I − t ( X + Z ) , 0 < t < 1 / 2 , A_1=I+tX,\quad A_2=I+tZ,\quad A_3=I-t(X+Z),\qquad0<t<1/\sqrt{2}, A 1 = I + tX , A 2 = I + tZ , A 3 = I − t ( X + Z ) , 0 < t < 1/ 2 ,
where X , Z X,Z X , Z are the Pauli matrices, and use the same choice for B B B . The matrices in each class do not commute.
For the tensor-product density F i j = A i ⊗ B j F_{ij}=A_i\otimes B_j F ij = A i ⊗ B j , logarithmic entropy satisfies H log ( T 1 F ) = H log ( F ) / 2 \mathcal{H}_{\log}(T_1F)=\mathcal{H}_{\log}(F)/2 H l o g ( T 1 F ) = H l o g ( F ) /2 . For powers put
a p = Tr E A i p − d A , b p = Tr E B j p − d B . a_p=\operatorname{Tr}\mathbb{E}A_i^p-d_A,\qquad b_p=\operatorname{Tr}\mathbb{E}B_j^p-d_B. a p = Tr E A i p − d A , b p = Tr E B j p − d B .
Then
H p ( F ) = d B a p + d A b p + a p b p , H p ( T 1 F ) = 1 2 ( d B a p + d A b p ) , \mathcal{H}_p(F)=d_Ba_p+d_Ab_p+a_pb_p,\qquad\mathcal{H}_p(T_1F)=\frac{1}{2}(d_Ba_p+d_Ab_p), H p ( F ) = d B a p + d A b p + a p b p , H p ( T 1 F ) = 2 1 ( d B a p + d A b p ) ,
so the power deficit is exactly a p b p / 2 > 0 a_pb_p/2>0 a p b p /2 > 0 for nonconstant ensembles.
For the orthogonal direct-sum density F ~ i j = A i ⊕ B j \widetilde{F}_{ij}=A_i\oplus B_j F ij = A i ⊕ B j , one has
H p ( F ~ ) = a p + b p , H p ( T 1 F ~ ) = 1 2 ( a p + b p ) \mathcal{H}_p(\widetilde{F})=a_p+b_p,\qquad\mathcal{H}_p(T_1\widetilde{F})=\frac{1}{2}(a_p+b_p) H p ( F ) = a p + b p , H p ( T 1 F ) = 2 1 ( a p + b p )
for every 1 < p ≤ 2 1<p\le2 1 < p ≤ 2 , and the logarithmic equality holds as well. The tensor product preserves logarithmic equality; the orthogonal sum preserves every trace profile. The next theorem identifies all power extremizers and explains this distinction.
Orthogonal structure of all power extremizers
Put σ = E μ F \sigma=\mathbb{E}_{\mu}F σ = E μ F , G i = ( T 1 F ) ( i ) G_i=(T_1F)(i) G i = ( T 1 F ) ( i ) and F 2 = T 2 F F_2=T_2F F 2 = T 2 F , and restrict all operators to supp σ \operatorname{supp}\sigma supp σ .
Theorem 8.2 (Universal power equality). Let f f f be any normalized homogeneous Lorentzian law of degree r ≥ 2 r\ge2 r ≥ 2 , and let E \mathcal{E} E be its exclusive classes from Lemma 7.2. For 1 < p < 2 1<p<2 1 < p < 2 and 0 < ℓ < r 0<\ell<r 0 < ℓ < r , a matrix density F F F attains equality in
H p ( T ℓ F ) ≤ ℓ r H p ( F ) \mathcal{H}_p(T_\ell F)\le\frac{\ell}{r}\mathcal{H}_p(F) H p ( T ℓ F ) ≤ r ℓ H p ( F ) if and only if there is an orthogonal decomposition into σ \sigma σ -reducing subspaces
Q = Q 0 ⊕ ⨁ C ∈ E Q C , σ = σ 0 ⊕ ⨁ C ∈ E σ C , Q=Q_0\oplus\bigoplus_{C\in\mathcal{E}}Q_C,\qquad\sigma=\sigma_0\oplus\bigoplus_{C\in\mathcal{E}}\sigma_C, Q = Q 0 ⊕ C ∈ E ⨁ Q C , σ = σ 0 ⊕ C ∈ E ⨁ σ C , and centered Hermitian ensembles D i C D_i^C D i C satisfying
∑ i ∈ C q C ( i ) D i C = 0 , σ C + D i C ⪰ 0 , \sum_{i\in C}q_C(i)D_i^C=0,\qquad\sigma_C+D_i^C\succeq0, i ∈ C ∑ q C ( i ) D i C = 0 , σ C + D i C ⪰ 0 , such that
F α = σ 0 ⊕ ⨁ C ∈ E ( σ C + 1 { γ C ( α ) = 1 } D α C ) . (46) F_\alpha=\sigma_0\oplus\bigoplus_{C\in\mathcal{E}}\left(\sigma_C+\mathbf{1}_{\{\gamma_C(\alpha)=1\}}D_\alpha^C\right). \tag*{(46)} F α = σ 0 ⊕ C ∈ E ⨁ ( σ C + 1 { γ C ( α ) = 1 } D α C ) . (46) Zero-dimensional factors are permitted. These densities have exact linear profiles for every trace function φ \varphi φ for which the quantities are finite.
In particular, if every exclusive class is a singleton, there are no nonconstant matrix power-entropy extremizers. This conclusion holds in every finite matrix dimension, even when the sharp complete coefficient equals ℓ / r \ell/r ℓ / r because the Hessian is singular.
Proof. Power equality makes F + s I F+sI F + s I a logarithmic equality density for every s > 0 s>0 s > 0 . Write D i = G i − σ D_i=G_i-\sigma D i = G i − σ , S s = σ + s I S_s=\sigma+sI S s = σ + s I , and B i ( s ) = I + S s − 1 / 2 D i S s − 1 / 2 B_i(s)=I+S_s^{-1/2}D_iS_s^{-1/2} B i ( s ) = I + S s − 1/2 D i S s − 1/2 . For every supported ordered pair, the Petz identity and its commutator in Lemma 7.3 give the exact formula
F 2 , i j + s I = S s 1 / 2 B i ( s ) B j ( s ) S s 1 / 2 = S s + D i + D j + D i S s − 1 D j . \begin{aligned}
F_{2,ij}+sI=S_s^{1/2}B_i(s)B_j(s)S_s^{1/2} \\
&=S_s+D_i+D_j+D_iS_s^{-1}D_j.
\end{aligned} F 2 , ij + s I = S s 1/2 B i ( s ) B j ( s ) S s 1/2 = S s + D i + D j + D i S s − 1 D j .
Letting s → ∞ s\to\infty s → ∞ yields F 2 , i j = σ + D i + D j F_{2,ij}=\sigma+D_i+D_j F 2 , ij = σ + D i + D j , so
D i ( σ + s I ) − 1 D j = 0 (47) D_i(\sigma+sI)^{-1}D_j=0 \tag*{(47)} D i ( σ + s I ) − 1 D j = 0 (47)
for every s > 0 s>0 s > 0 whenever W i j > 0 W_{ij}>0 W ij > 0 . If a label can repeat, its self-pair is supported and this identity forces D i = 0 D_i=0 D i = 0 .
For each exclusive class form the cyclic subspace
Q C = span { σ a D i v : i ∈ C , a ≥ 0 , v ∈ Q } . Q_C=\operatorname{span}\{\sigma^aD_iv:i\in C,\ a\geq0,\ v\in Q\}. Q C = span { σ a D i v : i ∈ C , a ≥ 0 , v ∈ Q } .
Distinct classes have supported cross pairs. Expanding (8.2) at infinity and using
⟨ σ a D i v , σ b D j w ⟩ = ⟨ v , D i σ a + b D j w ⟩ = 0 \langle\sigma^aD_iv,\sigma^bD_jw\rangle=\langle v,D_i\sigma^{a+b}D_jw\rangle=0 ⟨ σ a D i v , σ b D j w ⟩ = ⟨ v , D i σ a + b D j w ⟩ = 0
shows that the Q C Q_C Q C are pairwise orthogonal and reduce σ \sigma σ , and that every D i D_i D i in class C C C is supported on Q C Q_C Q C . The identity
0 = ∑ i p i D i = ∑ C ∈ E p C ∑ i ∈ C q C ( i ) D i 0=\sum_i p_iD_i=\sum_{C\in\mathcal{E}}p_C\sum_{i\in C}q_C(i)D_i 0 = i ∑ p i D i = C ∈ E ∑ p C i ∈ C ∑ q C ( i ) D i
is an orthogonal block identity, hence every class mean vanishes. Positivity of G i = σ + D i G_i=\sigma+D_i G i = σ + D i gives the required block positivity.
Construct the candidate F ~ \widetilde F F by (46) . Its class-centered form and the reference splitting (37) give mean σ \sigma σ and the same one-particle matrices G i G_i G i as the original density. Conditioning on a retained configuration keeps precisely the perturbations in its occupied exclusive classes; unrevealed representatives average to zero. Orthogonal-sum functional calculus therefore gives
H φ ( T k F ~ ) = ∑ C ∈ E μ k Pr ( C retained ) { ∑ i ∈ C q C ( i ) Tr φ ( σ C + D i C ) − Tr φ ( σ C ) } . \mathcal{H}_{\varphi}(T_k\widetilde F)
=\sum_{C\in\mathcal{E}^{\mu_k}}\Pr(C\text{ retained})
\left\{\sum_{i\in C}q_C(i)\operatorname{Tr}\varphi(\sigma_C+D_i^C)-\operatorname{Tr}\varphi(\sigma_C)\right\}. H φ ( T k F ) = C ∈ E μ k ∑ Pr ( C retained ) { i ∈ C ∑ q C ( i ) Tr φ ( σ C + D i C ) − Tr φ ( σ C ) } .
Exclusive classes occur at most once, so each probability scales by k / r k/r k / r . This proves exact linearity for every trace function.
The original F F F is a logarithmic equality density by letting s ↓ 0 s\downarrow0 s ↓ 0 in the shifted equalities. The candidate is also such a density and has the same mean and one-particle data. Exact Gibbs reconstruction therefore gives F = F ~ F=\widetilde F F = F . Conversely, the displayed profile formula proves power equality for every candidate. If an exclusive class is a singleton, its centered perturbation is zero, proving the final assertion. □
An exact fourth-order criterion
Let f f f be a normalized homogeneous polynomial with nonnegative coefficients, of degree r ≥ 2 r\geq2 r ≥ 2 , with occupation law μ \mu μ , and put
W = ∇ 2 f ( 1 ) r ( r − 1 ) , p = W 1 = 1 r ∇ f ( 1 ) . W=\frac{\nabla^2f(1)}{r(r-1)},\qquad p=W1=\frac{1}{r}\nabla f(1). W = r ( r − 1 ) ∇ 2 f ( 1 ) , p = W 1 = r 1 ∇ f ( 1 ) .
Let D i D_i D i be Hermitian d × d d\times d d × d matrices satisfying the matrix-valued translation identity
∑ i D i ∂ i f = 0 (48) \sum_iD_i\partial_i f=0 \tag*{(48)} i ∑ D i ∂ i f = 0 (48)
Thus every matrix-entry vector of ( D i ) i (D_i)_i ( D i ) i is a polynomial translation direction. Define
H α = ∑ i α i D i , F ε = I + ε H . H_\alpha=\sum_i\alpha_iD_i,\qquad F_\varepsilon=I+\varepsilon H. H α = i ∑ α i D i , F ε = I + ε H .
Then E H = 0 \mathbb{E}H=0 E H = 0 , and F ε F_\varepsilon F ε is positive definite for all sufficiently small real ε \varepsilon ε .
Theorem 8.3 (The fourth-order obstruction is sufficient as well as necessary). Set
Q f ( D ) = ∑ i , j W i j ( 2 Tr D i 2 D j 2 + Tr D i D j D i D j ) (49) Q_f(D)=\sum_{i,j}W_{ij}\left(2\operatorname{Tr}D_i^2D_j^2+\operatorname{Tr}D_iD_jD_iD_j\right) \tag*{(49)} Q f ( D ) = i , j ∑ W ij ( 2 Tr D i 2 D j 2 + Tr D i D j D i D j ) (49) = ∑ i , j W i j ( ∥ D i D j ∥ H S 2 + 1 2 ∥ D i D j + D j D i ∥ H S 2 ) ≥ 0. =\sum_{i,j}W_{ij}\left(\lVert D_iD_j\rVert_{\mathrm{HS}}^2+\frac{1}{2}\lVert D_iD_j+D_jD_i\rVert_{\mathrm{HS}}^2\right)\ge0. = i , j ∑ W ij ( ∥ D i D j ∥ HS 2 + 2 1 ∥ D i D j + D j D i ∥ HS 2 ) ≥ 0. For 1 < p < 2 1<p<2 1 < p < 2 and 0 < ℓ < r 0<\ell<r 0 < ℓ < r , the power-entropy deficit has expansion
ℓ r H p ( F ε ) − H p ( T ℓ F ε ) = p ( p − 1 ) ( 2 − p ) ( 3 − p ) 24 ℓ ( r − ℓ ) Q f ( D ) ε 4 + O ( ε 5 ) . (50) \frac{\ell}{r}\mathcal{H}_p(F_\varepsilon)-\mathcal{H}_p(T_\ell F_\varepsilon)=\frac{p(p-1)(2-p)(3-p)}{24}\ell(r-\ell)Q_f(D)\varepsilon^4+O(\varepsilon^5). \tag*{(50)} r ℓ H p ( F ε ) − H p ( T ℓ F ε ) = 24 p ( p − 1 ) ( 2 − p ) ( 3 − p ) ℓ ( r − ℓ ) Q f ( D ) ε 4 + O ( ε 5 ) . (50) For logarithmic entropy the corresponding expansion is
ℓ r H log ( F ε ) − H log ( T ℓ F ε ) = ℓ ( r − ℓ ) 12 Q f ( D ) ε 4 + O ( ε 5 ) . (51) \frac{\ell}{r}\mathcal{H}_{\log}(F_\varepsilon)-\mathcal{H}_{\log}(T_\ell F_\varepsilon)=\frac{\ell(r-\ell)}{12}Q_f(D)\varepsilon^4+O(\varepsilon^5). \tag*{(51)} r ℓ H l o g ( F ε ) − H l o g ( T ℓ F ε ) = 12 ℓ ( r − ℓ ) Q f ( D ) ε 4 + O ( ε 5 ) . (51) Here the mean is I I I ; trace-one normalization is not imposed.
The following are equivalent:
Q f ( D ) = 0 Q_f(D)=0 Q f ( D ) = 0 .
D i D j = 0 D_iD_j=0 D i D j = 0 whenever W i j > 0 W_{ij}>0 W ij > 0 , including supported self-pairs.
For every admissible ε \varepsilon ε and every trace function φ \varphi φ for which the quantities are finite,
H φ ( T k F ε ) = k r H φ ( F ε ) ( 0 ≤ k ≤ r ) . \mathcal{H}_{\varphi}(T_kF_\varepsilon)=\frac{k}{r}\mathcal{H}_{\varphi}(F_\varepsilon)\qquad(0\le k\le r). H φ ( T k F ε ) = r k H φ ( F ε ) ( 0 ≤ k ≤ r ) . The affine family F ε F_\varepsilon F ε attains equality for one fixed power 1 < p < 2 1<p<2 1 < p < 2 and one fixed proper retained level for all sufficiently small ε \varepsilon ε .
In (3), admissibility means F ε ( α ) ⪰ 0 F_\varepsilon(\alpha)\succeq0 F ε ( α ) ⪰ 0 for every supported α \alpha α . Thus no higher-order obstruction remains after the quartic obstruction vanishes.
Proof. Realize an occupation α \alpha α by an exchangeable ordered list of particles ( X 1 , … , X r ) (X_1,\ldots,X_r) ( X 1 , … , X r ) , uniformly ordered conditionally on α \alpha α . The law of X 1 X_1 X 1 is p p p and that of ( X 1 , X 2 ) (X_1,X_2) ( X 1 , X 2 ) is W W W . For any list of previously specified distinct particle positions, let β \beta β be their occupation count. Differentiating (8.3) by ∂ β \partial^\beta ∂ β shows that the conditional mean of the matrix at a new particle position vanishes:
E [ D X a ∣ the specified other particle labels ] = 0. (52) \mathbb{E}[D_{X_a}\mid\text{ the specified other particle labels}]=0. \tag*{(52)} E [ D X a ∣ the specified other particle labels ] = 0. (52) Indeed, its numerator is ∑ i D i ∂ β + e i f ( 1 ) = 0 \sum_iD_i\partial^{\beta+e_i}f(1)=0 ∑ i D i ∂ β + e i f ( 1 ) = 0 . The normalizing denominator is positive whenever the conditioning event has positive probability. This also gives
T k H ( β ) = ∑ i β i D i . T_kH(\beta)=\sum_i\beta_iD_i. T k H ( β ) = i ∑ β i D i . Expand Tr ( ∑ a = 1 r D X a ) j \operatorname{Tr}(\sum_{a=1}^rD_{X_a})^j Tr ( ∑ a = 1 r D X a ) j by partitions of the j j j positions according to which particle they use. Every pattern with a singleton block vanishes by (8.7). This argument is valid without commutativity: after conditioning, the singleton matrix occurs linearly between fixed matrix factors. Consequently
E Tr H 2 = r ∑ i p i Tr D i 2 , \mathbb{E}\operatorname{Tr}H^2=r\sum_i p_i\operatorname{Tr}D_i^2, E Tr H 2 = r i ∑ p i Tr D i 2 , E Tr H 3 = r ∑ i p i Tr D i 3 , \mathbb{E}\operatorname{Tr}H^3=r\sum_i p_i\operatorname{Tr}D_i^3, E Tr H 3 = r i ∑ p i Tr D i 3 , E Tr H 4 = r ∑ i p i Tr D i 4 + r ( r − 1 ) Q f ( D ) . (53) \mathbb{E}\operatorname{Tr}H^4=r\sum_i p_i\operatorname{Tr}D_i^4+r(r-1)Q_f(D). \tag*{(53)} E Tr H 4 = r i ∑ p i Tr D i 4 + r ( r − 1 ) Q f ( D ) . (53) At order four the only remaining patterns are a single block of size four and the three pair partitions. Two pair partitions have trace Tr D i 2 D j 2 \operatorname{Tr}D_i^2D_j^2 Tr D i 2 D j 2 , while the alternating partition has Tr D i D j D i D j \operatorname{Tr}D_iD_jD_iD_j Tr D i D j D i D j , yielding the stated formula.
Uniform retention keeps the same one- and two-particle laws, replacing r r r by k k k in (8.8). Thus the second- and third-order deficit moments vanish, while
ℓ r E Tr H 4 − E Tr ( T ℓ H ) 4 = ℓ ( r − ℓ ) Q f ( D ) . \frac{\ell}{r}\mathbb{E}\operatorname{Tr}H^4-\mathbb{E}\operatorname{Tr}(T_\ell H)^4=\ell(r-\ell)Q_f(D). r ℓ E Tr H 4 − E Tr ( T ℓ H ) 4 = ℓ ( r − ℓ ) Q f ( D ) .
Taylor expansion around the scalar matrix I I I now proves (50) and (51) : the fourth coefficients of ( 1 + x ) p (1+x)^p ( 1 + x ) p and ( 1 + x ) log ( 1 + x ) (1+x)\log(1+x) ( 1 + x ) log ( 1 + x ) are respectively p ( p − 1 ) ( p − 2 ) ( p − 3 ) / 24 p(p-1)(p-2)(p-3)/24 p ( p − 1 ) ( p − 2 ) ( p − 3 ) /24 and 1 / 12 1/12 1/12 .
For Hermitian A A A , B B B , cyclicity gives
∥ A B ∥ H S 2 = Tr A 2 B 2 , 1 2 ∥ A B + B A ∥ H S 2 = Tr A 2 B 2 + Tr A B A B . \lVert AB\rVert_{\mathrm{HS}}^2=\operatorname{Tr} A^2B^2,\qquad\frac{1}{2}\lVert AB+BA\rVert_{\mathrm{HS}}^2=\operatorname{Tr} A^2B^2+\operatorname{Tr} ABAB. ∥ A B ∥ HS 2 = Tr A 2 B 2 , 2 1 ∥ A B + B A ∥ HS 2 = Tr A 2 B 2 + Tr A B A B .
This proves the positive sum-of-squares formula and the equivalence of (1) and (2).
Assume (2). If a supported occupation repeats label i i i , its self-pair has positive mass, and D i 2 = 0 D_i^2=0 D i 2 = 0 forces D i = 0 D_i=0 D i = 0 . Distinct nonzero matrices appearing in the same supported occupation have orthogonal ranges, since D i D j = 0 D_iD_j=0 D i D j = 0 and the matrices are Hermitian. Therefore functional calculus splits on those ranges and their orthogonal complement. Writing
e i φ ( ε ) = Tr φ ( I + ε D i ) − Tr φ ( I ) , e_i^\varphi(\varepsilon)=\operatorname{Tr}\varphi(I+\varepsilon D_i)-\operatorname{Tr}\varphi(I), e i φ ( ε ) = Tr φ ( I + ε D i ) − Tr φ ( I ) ,
one has the exact identity on every supported occupation
Tr φ ( I + ε H α ) − Tr φ ( I ) = ∑ i α i e i φ ( ε ) . \operatorname{Tr}\varphi(I+\varepsilon H_\alpha)-\operatorname{Tr}\varphi(I)=\sum_i\alpha_i e_i^\varphi(\varepsilon). Tr φ ( I + ε H α ) − Tr φ ( I ) = i ∑ α i e i φ ( ε ) .
The same identity holds for every retained occupation. Since E β i = ( k / r ) E α i \mathbb{E}\beta_i=(k/r)\mathbb{E}\alpha_i E β i = ( k / r ) E α i and the retained mean is again I I I , averaging proves (3). It immediately implies (4), and the positive fourth coefficient in (50) shows that (4) implies (1). □
Remark 8.4. (Generality of the moment argument). No Lorentzian hypothesis is needed for the displayed moment identities or exact-profile criterion. For Lorentzian laws, these translations are precisely the Hessian null directions, and the surrounding contraction theorem holds for every positive semidefinite input.
Corollary 8.5. (Scalar and trace-one qubit forms). For a scalar h ∈ N f h\in\mathcal{N}_f h ∈ N f , put u ( α ) = α ⋅ h u(\alpha)=\alpha\cdot h u ( α ) = α ⋅ h and
Q f ( h ) = ( h ⊙ h ) T W ( h ⊙ h ) . Q_f(h)=(h\odot h)^{\mathsf{T}}W(h\odot h). Q f ( h ) = ( h ⊙ h ) T W ( h ⊙ h ) .
Then Q f ( h ) = 3 Q f ( h ) Q_f(h)=3Q_f(h) Q f ( h ) = 3 Q f ( h ) , and
ℓ r H p ( 1 + ε u ) − H p ( 1 + ε T ℓ u ) = p ( p − 1 ) ( 2 − p ) ( 3 − p ) 8 ℓ ( r − ℓ ) Q f ( h ) ε 4 + O ( ε 5 ) . \frac{\ell}{r}\mathcal{H}_p(1+\varepsilon u)-\mathcal{H}_p(1+\varepsilon T_\ell u)=\frac{p(p-1)(2-p)(3-p)}{8}\ell(r-\ell)Q_f(h)\varepsilon^4+O(\varepsilon^5). r ℓ H p ( 1 + ε u ) − H p ( 1 + ε T ℓ u ) = 8 p ( p − 1 ) ( 2 − p ) ( 3 − p ) ℓ ( r − ℓ ) Q f ( h ) ε 4 + O ( ε 5 ) .
For the trace-one qubit encoding
R ε ( α ) = 1 2 diag ( 1 + ε u ( α ) , 1 − ε u ( α ) ) , R_\varepsilon(\alpha)=\frac{1}{2}\operatorname{diag}(1+\varepsilon u(\alpha),1-\varepsilon u(\alpha)), R ε ( α ) = 2 1 diag ( 1 + ε u ( α ) , 1 − ε u ( α )) ,
one has
ℓ r H log ( R ε ) − H log ( T ℓ R ε ) = 1 4 ℓ ( r − ℓ ) Q f ( h ) ε 4 + O ( ε 6 ) , \frac{\ell}{r}\mathcal{H}_{\log}(R_\varepsilon)-\mathcal{H}_{\log}(T_\ell R_\varepsilon)=\frac{1}{4}\ell(r-\ell)Q_f(h)\varepsilon^4+O(\varepsilon^6), r ℓ H l o g ( R ε ) − H l o g ( T ℓ R ε ) = 4 1 ℓ ( r − ℓ ) Q f ( h ) ε 4 + O ( ε 6 ) ,
ℓ r H p ( R ε ) − H p ( T ℓ R ε ) = 2 1 − p p ( p − 1 ) ( 2 − p ) ( 3 − p ) 8 ℓ ( r − ℓ ) Q f ( h ) ε 4 + O ( ε 6 ) . \frac{\ell}{r}\mathcal{H}_p(R_\varepsilon)-\mathcal{H}_p(T_\ell R_\varepsilon)=2^{1-p}\frac{p(p-1)(2-p)(3-p)}{8}\ell(r-\ell)Q_f(h)\varepsilon^4+O(\varepsilon^6). r ℓ H p ( R ε ) − H p ( T ℓ R ε ) = 2 1 − p 8 p ( p − 1 ) ( 2 − p ) ( 3 − p ) ℓ ( r − ℓ ) Q f ( h ) ε 4 + O ( ε 6 ) .
The remainders are even because replacing ε \varepsilon ε by − ε -\varepsilon − ε exchanges the two qubit eigenvalues.
Proof. The scalar identity is immediate from the definition of Q f Q_f Q f . For the qubit, use D i = h i diag ( 1 , − 1 ) D_i=h_i\operatorname{diag}(1,-1) D i = h i diag ( 1 , − 1 ) , so Q f ( D ) = 6 Q f ( h ) Q_f(D)=6Q_f(h) Q f ( D ) = 6 Q f ( h ) , and then divide the density by two. Logarithmic entropy is positively homogeneous, whereas power entropy is homogeneous of degree p p p . □
Polynomial modes and deletion under label splitting
The linear extremal modes belong to a whole family of exact polynomial eigenfunctions. For independently split labels, this family sits in an explicit decomposition of every deletion–refill operator.
Let f ( z ) = ∑ ∣ α ∣ = r μ ( α ) z α f(z)=\sum_{|\alpha|=r}\mu(\alpha)z^\alpha f ( z ) = ∑ ∣ α ∣ = r μ ( α ) z α be a normalized occupation law of degree r ≥ 1 r\ge1 r ≥ 1 . Write T ℓ F = E [ F ∣ β ] T_\ell F=\mathbb{E}[F\mid\beta] T ℓ F = E [ F ∣ β ] for conditional expectation after uniform retention of ℓ \ell ℓ particles, and K ℓ = T ℓ ∗ T ℓ K_\ell=T_\ell^*T_\ell K ℓ = T ℓ ∗ T ℓ . The K ℓ K_\ell K ℓ are self-adjoint operators on the reference L 2 L^2 L 2 space.
The scalar equality densities already contain exact eigenfunctions of all orders. The mechanism uses only homogeneity and polynomial translation invariance; Lorentzian identifies these translations from the Hessian and identifies the generating densities as entropy extremizers.
Theorem 9.1 (Polynomial eigenfunctions from translations). Let f ( z ) = ∑ ∣ α ∣ = r μ ( α ) z α f(z)=\sum_{|\alpha|=r}\mu(\alpha)z^\alpha f ( z ) = ∑ ∣ α ∣ = r μ ( α ) z α , r ≥ 1 r\ge1 r ≥ 1 , have nonnegative coefficients and f ( 1 ) = 1 f(\mathbf{1})=1 f ( 1 ) = 1 , and set N f = { h : D h f = 0 } \mathcal{N}_f=\{h:D_hf=0\} N f = { h : D h f = 0 } . For h ∈ N f h\in\mathcal{N}_f h ∈ N f , define
Φ h ( t ; α ) = ∏ i ( 1 + t h i ) α i = ∑ j = 0 r t j P j , h ( α ) , E j = span { P j , h : h ∈ N f } . \Phi_h(t;\alpha)=\prod_i(1+th_i)^{\alpha_i}=\sum_{j=0}^{r}t^jP_{j,h}(\alpha),\qquad\mathcal{E}_j=\operatorname{span}\{P_{j,h}:h\in\mathcal{N}_f\}. Φ h ( t ; α ) = i ∏ ( 1 + t h i ) α i = j = 0 ∑ r t j P j , h ( α ) , E j = span { P j , h : h ∈ N f } . For every 0 ≤ ℓ ≤ r 0\le\ell\le r 0 ≤ ℓ ≤ r and 0 ≤ j ≤ r 0\le j\le r 0 ≤ j ≤ r ,
K ℓ ∣ E j = ( ℓ j ) ( r j ) I , ( ℓ j ) = 0 ( j > ℓ ) . K_\ell|_{\mathcal{E}_j}=\frac{\binom{\ell}{j}}{\binom{r}{j}}I,\qquad\binom{\ell}{j}=0\quad(j>\ell). K ℓ ∣ E j = ( j r ) ( j ℓ ) I , ( j ℓ ) = 0 ( j > ℓ ) . The spaces E j \mathcal{E}_j E j are pairwise orthogonal. More precisely, with u ⊙ v = ( u i v i ) i u\odot v=(u_iv_i)_i u ⊙ v = ( u i v i ) i ,
⟨ P j , u , P k , v ⟩ L 2 ( μ ) = 1 j = k [ z j ] f ( 1 + z ( u ⊙ v ) ) . \langle P_{j,u},P_{k,v}\rangle_{L^2(\mu)}=\mathbf{1}_{j=k}[z^j]f(1+z(u\odot v)). ⟨ P j , u , P k , v ⟩ L 2 ( μ ) = 1 j = k [ z j ] f ( 1 + z ( u ⊙ v )) . Consequently the one-particle replacement semigroup acts on E j \mathcal{E}_j E j by e − j t / r e^{-jt/r} e − j t / r . When f f f is Lorentzian, ⨁ j E j \bigoplus_j\mathcal{E}_j ⨁ j E j is exactly the linear span of all strictly positive scalar equality densities. Proof. Write a = 1 + t h a=1+th a = 1 + t h . Translation invariance passes to every partial derivative, so the conditional-mean identity gives
T ℓ Φ h ( t ; ⋅ ) ( β ) = a β ∂ β f ( a ) ∂ β f ( 1 ) = Φ h ( t ; β ) . T_\ell\Phi_h(t;\cdot)(\beta)=a^\beta\frac{\partial^\beta f(a)}{\partial^\beta f(\mathbf{1})}=\Phi_h(t;\beta). T ℓ Φ h ( t ; ⋅ ) ( β ) = a β ∂ β f ( 1 ) ∂ β f ( a ) = Φ h ( t ; β ) . It suffices to prove this for small real t t t , when a > 0 a>0 a > 0 ; the identity then holds polynomially. Each degree-j j j term in Φ h ( t ; α ) \Phi_h(t;\alpha) Φ h ( t ; α ) is a product of the labels of j j j distinct particle positions. A uniformly retained ℓ \ell ℓ -subset contains those positions with probability ( ℓ j ) / ( r j ) \binom{\ell}{j}/\binom{r}{j} ( j ℓ ) / ( j r ) . Thus
T ℓ ∗ P j , h ( β ) = ( ℓ j ) ( r j ) P j , h ( α ) , T_\ell^*P_{j,h}(\beta)=\frac{\binom{\ell}{j}}{\binom{r}{j}}P_{j,h}(\alpha), T ℓ ∗ P j , h ( β ) = ( j r ) ( j ℓ ) P j , h ( α ) , which proves the eigenvalue formula. The Gram identity follows from
E μ [ Φ u ( s ; α ) Φ v ( t ; α ) ] = f ( 1 + s u + t v + s t ( u ⊙ v ) ) = f ( 1 + s t ( u ⊙ v ) ) . \mathbb{E}_\mu[\Phi_u(s;\alpha)\Phi_v(t;\alpha)]=f(1+su+tv+st(u\odot v))=f(1+st(u\odot v)). E μ [ Φ u ( s ; α ) Φ v ( t ; α )] = f ( 1 + s u + t v + s t ( u ⊙ v )) = f ( 1 + s t ( u ⊙ v )) . Compare coefficients of s j t k s^jt^k s j t k . For ℓ = r − 1 \ell=r-1 ℓ = r − 1 , the eigenvalue is 1 − j / r 1-j/r 1 − j / r , proving the semigroup assertion. Finally, Theorem 5.2 identifies the strictly positive extremizers with Φ h ( 1 ; ⋅ ) \Phi_h(1;\cdot) Φ h ( 1 ; ⋅ ) whenever 1 + h > 0 1+h>0 1 + h > 0 . Evaluating at r + 1 r+1 r + 1 sufficiently small distinct values of t t t and inverting the Vandermonde matrix recovers each P j , h P_{j,h} P j , h . Conversely each positive extremizer is the sum of its homogeneous coefficients. In rank one, the same equality classification follows directly from ∑ i μ ( e i ) h i = 0 \sum_i\mu(e_i)h_i=0 ∑ i μ ( e i ) h i = 0 .
Remark 9.2 (Classical specialization). For f ( z ) = ( ∑ i p i z i ) r f(z)=(\sum_i p_iz_i)^r f ( z ) = ( ∑ i p i z i ) r , the translation space is p ⊥ p^\perp p ⊥ and the Gram kernel is ( 1 + s t ∑ i p i u i v i ) r (1+st\sum_i p_iu_iv_i)^r ( 1 + s t ∑ i p i u i v i ) r . The spaces above give the classical multivariate Krawtchouk decomposition of a multinomial law [19 ] . In general some E j \mathcal{E}_j E j can vanish, and these spaces need not exhaust the full eigenspaces of a single K ℓ K_\ell K ℓ . Their construction is canonical and simultaneous over all retained levels.
The full decomposition under multinomial splitting
The exact spectral decomposition has a probabilistic form that includes repeated labels. It requires only independent splitting of each coarse label. The quotient occupation law may be arbitrary.
Let
g ( y ) = ∑ ∣ s ∣ = r ν ( s ) y s , g ( 1 ) = 1 , (54) g(y)=\sum_{|s|=r}\nu(s)y^s,\qquad g(\mathbf{1})=1, \tag*{(54)} g ( y ) = ∣ s ∣ = r ∑ ν ( s ) y s , g ( 1 ) = 1 , (54)
be homogeneous with nonnegative coefficients. For each coarse label C C C , choose a finite set C C \mathcal{C}_C C C , with these sets pairwise disjoint, and a positive probability law q C q_C q C on that set. Define
f ( z ) = g ( ( ∑ i ∈ C C q C ( i ) z i ) C ) . f(z)=g\left(\left(\sum_{i\in\mathcal{C}_C}q_C(i)z_i\right)_C\right). f ( z ) = g ( i ∈ C C ∑ q C ( i ) z i ) C .
The law of f f f first samples the coarse occupation s s s from ν \nu ν , then independently assigns colors with law q C q_C q C to the s C s_C s C particles of class C C C . Write K ℓ f K_\ell^f K ℓ f for uniform retention of ℓ \ell ℓ particles followed by conditional refill.
For a multiindex a = ( a C ) C a=(a_C)_C a = ( a C ) C with ∂ a g ≠ 0 \partial^a g\ne0 ∂ a g = 0 , put
d = ∣ a ∣ , g a ( y ) = ∂ a g ( y ) ∂ a g ( 1 ) , V C = L 0 2 ( C C , q C ) , V a = ⨂ C Sym a C V C . d=|a|,\qquad g_a(y)=\frac{\partial^a g(y)}{\partial^a g(1)},\qquad V_C=L^2_0(\mathcal{C}_C,q_C),\qquad V_a=\bigotimes_C\operatorname{Sym}^{a_C}V_C. d = ∣ a ∣ , g a ( y ) = ∂ a g ( 1 ) ∂ a g ( y ) , V C = L 0 2 ( C C , q C ) , V a = C ⨂ Sym a C V C .
The polynomial g a g_a g a is a normalized occupation law of degree r − d r-d r − d . Symmetric powers here are the symmetric tensor subspaces with their product Hilbert inner products. The index a C a_C a C counts centered color slots in class C C C . A conditional statistic in this sector can survive only if all those slots are retained.
Theorem 9.3 (Exact derivative-law decomposition). There is an orthogonal decomposition
L 2 ( μ f ) ≅ ⨁ a > 0 ∂ a g ≠ 0 L 2 ( μ g a ) ⊗ V a L^2(\mu_f)\cong\bigoplus_{\substack{a>0\\ \partial^a g\ne0}}L^2(\mu_{g_a})\otimes V_a L 2 ( μ f ) ≅ a > 0 ∂ a g = 0 ⨁ L 2 ( μ g a ) ⊗ V a under which, simultaneously for every 0 ≤ ℓ ≤ r 0\le\ell\le r 0 ≤ ℓ ≤ r ,
K ℓ f ≅ { ( ℓ d ) / ( r d ) K ℓ − d g a ⊗ I V a , d ≤ ℓ , 0 , d > ℓ . K_\ell^f\cong
\begin{cases}
\binom{\ell}{d}\bigg/\binom{r}{d}\,K_{\ell-d}^{g_a}\otimes I_{V_a},&d\le\ell,\\
0,&d>\ell.
\end{cases} K ℓ f ≅ ⎩ ⎨ ⎧ ( d ℓ ) / ( d r ) K ℓ − d g a ⊗ I V a , 0 , d ≤ ℓ , d > ℓ . For a constant polynomial g a g_a g a , its rank-zero operator is the identity on the one-point probability space. Zero-dimensional V a V_a V a may be omitted. The multiplicity of the derivative-law block is
dim V a = ∏ C dim Sym a C V C . \dim V_a=\prod_C\dim\operatorname{Sym}^{a_C}V_C. dim V a = C ∏ dim Sym a C V C . If every color class has at least two elements, this is
∏ C ( ∣ C C ∣ + a C − 2 a C ) . \prod_C\binom{|\mathcal{C}_C|+a_C-2}{a_C}. C ∏ ( a C ∣ C C ∣ + a C − 2 ) . A singleton class contributes zero when a C > 0 a_C>0 a C > 0 , and one when a C = 0 a_C=0 a C = 0 .
Proof. Fix a coarse occupation s s s . The space of functions of the color counts in class C C C identifies with the symmetric functions in L 2 ( q C ) ⊗ s C L^2(q_C)^{\otimes s_C} L 2 ( q C ) ⊗ s C . The decomposition L 2 ( q C ) = R 1 ⊕ V C L^2(q_C)=\mathbb{R}1\oplus V_C L 2 ( q C ) = R 1 ⊕ V C gives the orthogonal sum
Sym s C L 2 ( q C ) ≅ ⨁ a C = 0 s C Sym a C V C . \operatorname{Sym}^{s_C}L^2(q_C)\cong\bigoplus_{a_C=0}^{s_C}\operatorname{Sym}^{a_C}V_C. Sym s C L 2 ( q C ) ≅ a C = 0 ⨁ s C Sym a C V C .
We make the normalization explicit. For a symmetric tensor v ∈ Sym a C V C v\in\operatorname{Sym}^{a_C}V_C v ∈ Sym a C V C , define the symmetric statistic
U a C , v ( x 1 , … , x s C ) = ∑ J ⊆ [ s C ] ∣ J ∣ = a C v ( x J ) . U_{a_C,v}(x_1,\ldots,x_{s_C})=\sum_{\substack{J\subseteq[s_C]\\ |J|=a_C}}v(x_J). U a C , v ( x 1 , … , x s C ) = J ⊆ [ s C ] ∣ J ∣ = a C ∑ v ( x J ) .
For two distinct subsets J , J ′ J,J' J , J ′ , a coordinate in their symmetric difference is centered in exactly one factor; integrating it gives zero. Hence
∥ U a C , v ∥ L 2 ( q C ⊗ s C ) 2 = ( s C a C ) ∥ v ∥ 2 . \lVert U_{a_C,v}\rVert^2_{L^2(q_C^{\otimes s_C})}=\binom{s_C}{a_C}\lVert v\rVert^2. ∥ U a C , v ∥ L 2 ( q C ⊗ s C ) 2 = ( a C s C ) ∥ v ∥ 2 .
Take tensor products across classes, and extend linearly to v ∈ V a v\in V_a v ∈ V a ; write the resulting statistic as U a , v U_{a,v} U a , v .
For a coefficient function ψ ( s − a ) \psi(s-a) ψ ( s − a ) , the squared norm of ψ ( s − a ) U a , v \psi(s-a)U_{a,v} ψ ( s − a ) U a , v is
∑ s ≥ a ν ( s ) ∏ C ( s C a C ) ∣ ψ ( s − a ) ∣ 2 ∥ v ∥ 2 . \sum_{s\geq a}\nu(s)\prod_C\binom{s_C}{a_C}|\psi(s-a)|^2\lVert v\rVert^2. s ≥ a ∑ ν ( s ) C ∏ ( a C s C ) ∣ ψ ( s − a ) ∣ 2 ∥ v ∥ 2 .
The normalization factor is
c a = ∑ s ν ( s ) ∏ C ( s C a C ) = ∂ a g ( 1 ) ∏ C a C ! . c_a=\sum_s\nu(s)\prod_C\binom{s_C}{a_C}=\frac{\partial^a g(1)}{\prod_C a_C!}. c a = s ∑ ν ( s ) C ∏ ( a C s C ) = ∏ C a C ! ∂ a g ( 1 ) .
After division by c a c_a c a , the coefficient law is precisely μ g a \mu_{g_a} μ g a , because the coefficient of y s − a y^{s-a} y s − a in ∂ a g \partial^a g ∂ a g is ν ( s ) ∏ C ( s C ) a C \nu(s)\prod_C(s_C)_{a_C} ν ( s ) ∏ C ( s C ) a C , where ( u ) v = u ( u − 1 ) ⋯ ( u − v + 1 ) (u)_v=u(u-1)\cdots(u-v+1) ( u ) v = u ( u − 1 ) ⋯ ( u − v + 1 ) . Thus
J a ( ψ ⊗ v ) = c a − 1 / 2 ψ ( s − a ) U a , v J_a(\psi\otimes v)=c_a^{-1/2}\psi(s-a)U_{a,v} J a ( ψ ⊗ v ) = c a − 1/2 ψ ( s − a ) U a , v
is isometric. The fixed-s s s tensor decompositions show that the images of all J a J_a J a are orthogonal and exhaustive.
Now condition on a retained detailed color occupation, with coarse counts b C b_C b C and total size ℓ \ell ℓ . If b C < a C b_C<a_C b C < a C for some C C C , the conditional statistic is zero. Otherwise b − a b-a b − a is a retained occupation for the derivative law. Every term of U a , v U_{a,v} U a , v that uses an unretained centered color has conditional mean zero. The surviving terms are exactly U a , v U_{a,v} U a , v evaluated on the retained colors. The conditional distribution of the coefficient occupation s − a s-a s − a is the posterior for g a g_a g a after retention of b − a b-a b − a . Indeed, the two posterior weights agree up to a factor independent of s s s , by
( s ) a ( s − a b − a ) = ( b ) a ( s b ) . (s)_a\binom{s-a}{b-a}=(b)_a\binom{s}{b}. ( s ) a ( b − a s − a ) = ( b ) a ( b s ) .
with products taken coordinatewise.
The corresponding normalization at level ℓ \ell ℓ is
c a , ℓ = E ∏ C ( b C a C ) = ( ℓ d ) ( r d ) c a . c_{a,\ell}=\mathbb{E}\prod_C\binom{b_C}{a_C}=\frac{\binom{\ell}{d}}{\binom{r}{d}}c_a. c a , ℓ = E C ∏ ( a C b C ) = ( d r ) ( d ℓ ) c a .
This follows by counting a fixed set of d d d marked particle positions: uniform retention keeps it with probability ( ℓ d ) / ( r d ) \binom{\ell}{d}/\binom{r}{d} ( d ℓ ) / ( d r ) . Therefore, under the isometries at the input and output levels, conditional expectation on the a a a -sector is
( ℓ d ) ( r d ) T ℓ − d g a ⊗ I . \sqrt{\frac{\binom{\ell}{d}}{\binom{r}{d}}}\,T^{g_a}_{\ell-d}\otimes I. ( d r ) ( d ℓ ) T ℓ − d g a ⊗ I .
Taking its adjoint followed by itself gives the claimed K ℓ K_\ell K ℓ block. If ℓ < d \ell<d ℓ < d , no such marked set survives and the block is zero. □
Corollary 9.4 (Isospectrality under color redistribution). Fix g g g and the cardinality of each color class. Varying the positive laws q C q_C q C gives one unitary equivalence intertwining all the operators K ℓ f K_\ell^f K ℓ f simultaneously.
Proof. The derivative laws g a g_a g a are unchanged. Choose an isometry between the old and new centered spaces V C V_C V C , and take its symmetric and tensor powers in the decomposition. □
For one coarse class, g ( y ) = y r g(y)=y^r g ( y ) = y r , every derivative law is a point mass. The decomposition becomes the usual multivariate Krawtchouk decomposition of a multinomial law [19 ] . For a multiaffine g g g with matroid support, each nonzero multiindex a a a is the indicator of an independent set in the support matroid, and differentiation becomes contraction. This gives the next section.
Contraction minors of a parallel extension
Let M M M be a loopless rank-r r r matroid with positive fields λ i \lambda_i λ i , and let S S S be its simplification. Its elements are the parallel classes C C C . Put
w C = ∑ i ∈ C λ i , q C ( i ) = λ i / w C , V C = L 0 2 ( C , q C ) . w_C=\sum_{i\in C}\lambda_i,\qquad q_C(i)=\lambda_i/w_C,\qquad V_C=L^2_0(C,q_C). w C = i ∈ C ∑ λ i , q C ( i ) = λ i / w C , V C = L 0 2 ( C , q C ) .
The simplified basis I I I has law proportional to ∏ C ∈ I w C \prod_{C\in I}w_C ∏ C ∈ I w C ; given I I I , its representatives are independent with laws q C q_C q C . For A ∈ I ( S ) A\in\mathcal{I}(S) A ∈ I ( S ) , let π A \pi^A π A be the basis law of S / A S/A S / A with fields w C w_C w C , equivalently the simplified law conditioned to contain A A A .
Corollary 9.5 (All retained-level walks through contraction minors). One orthogonal decomposition, independent of ℓ \ell ℓ , identifies
L 2 ( μ M ) ≅ ⨁ A ∈ I ( S ) L 2 ( π A ) ⊗ ⨂ C ∈ A V C . L^2(\mu_M)\cong\bigoplus_{A\in\mathcal{I}(S)}L^2(\pi^A)\otimes\bigotimes_{C\in A}V_C. L 2 ( μ M ) ≅ A ∈ I ( S ) ⨁ L 2 ( π A ) ⊗ C ∈ A ⨂ V C . On the summand indexed by A A A , with j = ∣ A ∣ ≤ ℓ j=|A|\leq\ell j = ∣ A ∣ ≤ ℓ ,
K ℓ M ≅ ( ℓ j ) / ( r j ) K ℓ − j S / A ⊗ I . K_\ell^M\cong\binom{\ell}{j}\bigg/\binom{r}{j}\,K_{\ell-j}^{S/A}\otimes I. K ℓ M ≅ ( j ℓ ) / ( j r ) K ℓ − j S / A ⊗ I . The summand is killed when j > ℓ j>\ell j > ℓ . A rank-zero contracted walk is the identity on a one-point space. Redistributing positive fields within parallel classes while preserving every w C w_C w C gives one unitary equivalence intertwining all retained-level walks. In particular, every strictly positive scalar entropy equality tilt preserves their entire spectra simultaneously.
Proof. In Theorem 9.3, multiaffinity makes a a a the indicator of an independent set A A A , and the derivative law is π A \pi^A π A . This gives the decomposition. The simplification and all its conditional laws depend only on w C w_C w C ; the representative spaces retain dimension ∣ C ∣ − 1 |C|-1 ∣ C ∣ − 1 under redistribution. Apply Corollary 9.4. Theorem 5.6 identifies positive scalar equality tilts with exactly these redistributions.
The constants on the contracted spaces give the polynomial modes of degree j j j , with dimension
m j = ∑ A ∈ I ( S ) ∣ A ∣ = j ∏ C ∈ A ( ∣ C ∣ − 1 ) . m_j=\sum_{\substack{A\in\mathcal{I}(S)\\ |A|=j}}\prod_{C\in A}(|C|-1). m j = A ∈ I ( S ) ∣ A ∣ = j ∑ C ∈ A ∏ ( ∣ C ∣ − 1 ) .
Deleting one representative from each parallel class gives a matroid on N ′ = ∑ C ( ∣ C ∣ − 1 ) N'=\sum_C(|C|-1) N ′ = ∑ C ( ∣ C ∣ − 1 ) elements with independent-set numbers m j m_j m j . Thus ( m j ) j = 0 N ′ (m_j)_{j=0}^{N'} ( m j ) j = 0 N ′ is ultra log-concave by Mason’s theorem [6 , 13 ] : ( m j / ( N ′ j ) ) j = 0 N ′ (m_j/\binom{N'}{j})_{j=0}^{N'} ( m j / ( j N ′ ) ) j = 0 N ′ is log-concave. These are the dimensions of the canonical mode spaces, not necessarily the full multiplicities of a given eigenvalue: other contraction blocks can contribute that eigenvalue.
The operator decomposition also applies to every multiaffine Lorentzian coefficient law. Lemma 7.2 gives independent representative choices after merging parallel classes; the contraction laws are then normalized derivatives of the quotient polynomial, without a product-field assumption.
For such coefficient laws, quotient translations can give additional polynomial modes beyond the within-class spaces counted by m j m_j m j .
Exchange edges, isoperimetry, and rigidity
Conditioning on a retained face leaves a clique of possible completions. Factor entropy inside each such clique, then count how often each pair of configurations occurs. This converts deletion contraction into a complete inequality on individual exchange edges. Equality remembers the clique sizes and identifies exactly the binary factors of the basis graph.
Let μ \mu μ be positive on a family B \mathcal{B} B of r r r -element sets. For an unordered pair e = { B , B ′ } e=\{B,B'\} e = { B , B ′ } , let K e K_e K e replace the two matrix values by their μ \mu μ -weighted mean and leave every other value fixed. Define its entropy loss by
J e ( F ) = H μ ( F ) − H μ ( K e F ) . J_e(F)=\mathcal{H}_\mu(F)-\mathcal{H}_\mu(K_eF). J e ( F ) = H μ ( F ) − H μ ( K e F ) .
Lemma 10.1 (Complete-graph factorization). For every probability law ν \nu ν on a finite set and every positive semidefinite matrix function F F F ,
H ν ( F ) ≤ ∑ x < y H ( ν ( x ) , ν ( y ) ; F ( x ) , F ( y ) ) . \mathcal{H}_{\nu}(F) \le\sum_{x<y} H\bigl(\nu(x),\nu(y);F(x),F(y)\bigr). H ν ( F ) ≤ x < y ∑ H ( ν ( x ) , ν ( y ) ; F ( x ) , F ( y ) ) . where
H ( a , b ; A , B ) = Tr [ a A log A + b B log B − ( a A + b B ) log a A + b B a + b ] . H(a,b;A,B)=\operatorname{Tr}\left[aA\log A+bB\log B-(aA+bB)\log\frac{aA+bB}{a+b}\right]. H ( a , b ; A , B ) = Tr [ a A log A + b B log B − ( a A + b B ) log a + b a A + b B ] .
Proof. The two-point loss is jointly convex in ( A , B ) (A,B) ( A , B ) , homogeneous in ( a , b ) (a,b) ( a , b ) , and nondecreasing in each mass. To see the last assertion, write M = ( a A + b B ) / ( a + b ) M=(aA+bB)/(a+b) M = ( a A + b B ) / ( a + b ) and D g ( A ∥ M ) = Tr [ A ( log A − log M ) − A + M ] D_g(A\mathbin{\|}M)=\operatorname{Tr}[A(\log A-\log M)-A+M] D g ( A ∥ M ) = Tr [ A ( log A − log M ) − A + M ] . Then H = a D g ( A ∥ M ) + b D g ( B ∥ M ) H=aD_g(A\mathbin{\|}M)+bD_g(B\mathbin{\|}M) H = a D g ( A ∥ M ) + b D g ( B ∥ M ) and ∂ a H = D g ( A ∥ M ) ≥ 0 \partial_aH=D_g(A\mathbin{\|}M)\ge0 ∂ a H = D g ( A ∥ M ) ≥ 0 . Joint convexity follows from relative entropy; regularization covers singular matrices.
Induct on the support size. Split off x x x of mass a a a , and put b = 1 − a b=1-a b = 1 − a . The entropy chain rule separates the remaining conditional entropy from the loss between F ( x ) F(x) F ( x ) and the remaining mean. Convexity bounds this latter loss by
∑ y ≠ x ν ( y ) b H ( a , b ; F ( x ) , F ( y ) ) = ∑ y ≠ x H ( a ν ( y ) / b , ν ( y ) ; F ( x ) , F ( y ) ) . \sum_{y\ne x}\frac{\nu(y)}{b}H\bigl(a,b;F(x),F(y)\bigr)
=\sum_{y\ne x}H\bigl(a\nu(y)/b,\nu(y);F(x),F(y)\bigr). y = x ∑ b ν ( y ) H ( a , b ; F ( x ) , F ( y ) ) = y = x ∑ H ( a ν ( y ) / b , ν ( y ) ; F ( x ) , F ( y ) ) .
Since ν ( y ) ≤ b \nu(y)\le b ν ( y ) ≤ b , mass monotonicity bounds these terms by the required cross pairs. Homogeneity and induction treat the remaining conditional entropy. □
Theorem 10.2 (Intersection-weighted edge factorization). Suppose uniform deletion to ℓ \ell ℓ elements, 0 ≤ ℓ < r 0\le\ell<r 0 ≤ ℓ < r , satisfies H μ ℓ ( F ℓ ) ≤ η H μ ( F ) \mathcal{H}_{\mu_\ell}(F_\ell)\le\eta\mathcal{H}_{\mu}(F) H μ ℓ ( F ℓ ) ≤ η H μ ( F ) completely, with η < 1 \eta<1 η < 1 . Then
H μ ( F ) ≤ 1 ( r ℓ ) ( 1 − η ) ∑ B < B ′ ( ∣ B ∩ B ′ ∣ ℓ ) J { B , B ′ } ( F ) . (55) \mathcal{H}_{\mu}(F)\le\frac{1}{\binom{r}{\ell}(1-\eta)}\sum_{B<B'}\binom{|B\cap B'|}{\ell}J_{\{B,B'\}}(F). \tag*{(55)} H μ ( F ) ≤ ( ℓ r ) ( 1 − η ) 1 B < B ′ ∑ ( ℓ ∣ B ∩ B ′ ∣ ) J { B , B ′ } ( F ) . (55) For every homogeneous multiaffine Lorentzian law, in particular,
H μ ( F ) ≤ ∑ e ∈ E J e ( F ) , E = { { B , B ′ } : ∣ B ∩ B ′ ∣ = r − 1 } . (56) \mathcal{H}_{\mu}(F)\le\sum_{e\in E}J_e(F),\qquad E=\{\{B,B'\}:|B\cap B'|=r-1\}. \tag*{(56)} H μ ( F ) ≤ e ∈ E ∑ J e ( F ) , E = {{ B , B ′ } : ∣ B ∩ B ′ ∣ = r − 1 } . (56) Under the gap hypothesis (3.3), the right side can be multiplied by
C κ = r + ( r − 1 ) κ r ( 1 + κ ) . (57) C_\kappa=\frac{r+(r-1)\kappa}{r(1+\kappa)}. \tag*{(57)} C κ = r ( 1 + κ ) r + ( r − 1 ) κ . (57)
Proof. The exact chain rule gives
H μ ( F ) − H μ ℓ ( F ℓ ) = ∑ I μ ℓ ( I ) H μ ( ⋅ ∣ I ) ( F ) . \mathcal{H}_{\mu}(F)-\mathcal{H}_{\mu_\ell}(F_\ell)=\sum_I\mu_\ell(I)\mathcal{H}_{\mu(\cdot\mid I)}(F). H μ ( F ) − H μ ℓ ( F ℓ ) = I ∑ μ ℓ ( I ) H μ ( ⋅ ∣ I ) ( F ) .
Apply Lemma 10.1 to each completion law. For I ⊆ B I\subseteq B I ⊆ B , Bayes’ formula gives μ ℓ ( I ) μ ( B ∣ I ) = μ ( B ) / ( r ℓ ) \mu_\ell(I)\mu(B\mid I)=\mu(B)/\binom{r}{\ell} μ ℓ ( I ) μ ( B ∣ I ) = μ ( B ) / ( ℓ r ) . Homogeneity therefore makes every shared I I I contribute J { B , B ′ } ( F ) / ( r ℓ ) J_{\{B,B'\}}(F)/\binom{r}{\ell} J { B , B ′ } ( F ) / ( ℓ r ) . There are ( ∣ B ∩ B ′ ∣ ℓ ) \binom{|B\cap B'|}{\ell} ( ℓ ∣ B ∩ B ′ ∣ ) such faces, proving (10.1). For Lorentzian laws take ℓ = r − 1 \ell=r-1 ℓ = r − 1 and η = ( r − 1 ) / r \eta=(r-1)/r η = ( r − 1 ) / r ; for r = 1 r=1 r = 1 the output is the mean and η = 0 \eta=0 η = 0 . The intersection weight is exactly one on exchange edges. Using η = a r − 1 / a r \eta=a_{r-1}/a_r η = a r − 1 / a r from Theorem 3.4 instead produces C κ C_\kappa C κ . □
Corollary 10.3 (Edge heat flow and matrix concentration). Let L E = ∑ e ∈ E ( I − K e ) L_E=\sum_{e\in E}(I-K_e) L E = ∑ e ∈ E ( I − K e ) , so each edge updates at rate one. For every Lorentzian basis law,
H μ ( e − t L E F ) ≤ e − t H μ ( F ) . \mathcal{H}_{\mu}(e^{-tL_E}F)\le e^{-t}\mathcal{H}_{\mu}(F). H μ ( e − t L E F ) ≤ e − t H μ ( F ) . The complete modified logarithmic Sobolev constant is at least one, and at least C κ − 1 C_\kappa^{-1} C κ − 1 under (3.3). Theorem 11.1 applies with this constant and the actual Γ L E \Gamma_{L_E} Γ L E . The exponent one is optimal uniformly over positive product fields.
Proof. Convexity gives J e ( F ) ≤ E μ Tr [ ( ( I − K e ) F ) log F ] J_e(F)\le\mathbb{E}_{\mu}\operatorname{Tr}[((I-K_e)F)\log F] J e ( F ) ≤ E μ Tr [(( I − K e ) F ) log F ] . Sum the edge factorization and integrate entropy production along the semigroup. For sharpness use the rank-one two-base law μ = ( ε , 1 − ε ) \mu=(\varepsilon,1-\varepsilon) μ = ( ε , 1 − ε ) . Starting from ( a , 1 − a ) (a,1-a) ( a , 1 − a ) , 0 < a < 1 0<a<1 0 < a < 1 , the first coordinate at time t t t is ε + ( a − ε ) e − t \varepsilon+(a-\varepsilon)e^{-t} ε + ( a − ε ) e − t . As ε ↓ 0 \varepsilon\downarrow0 ε ↓ 0 , the output-to-input relative-entropy ratio tends to e − t e^{-t} e − t . □
For the uniform law, the unit-edge graph Laplacian is L G = 2 L E L_G = 2L_E L G = 2 L E . When E ≠ ∅ E \ne\varnothing E = ∅ , the discrete kernel ∣ E ∣ − 1 ∑ e K e |E|^{-1}\sum_e K_e ∣ E ∣ − 1 ∑ e K e has complete contraction at most 1 − 1 / ∣ E ∣ 1 - 1/|E| 1 − 1/∣ E ∣ : convexity bounds its output entropy by the average output entropy of the edge updates. These clocks differ from one deletion–replacement update; we keep their normalizations explicit.
Complete entropy decay for the unit-edge clock
For the uniform law, the sharp matrix midpoint estimate below converts edge factorization into complete entropy production, in the modified logarithmic Sobolev framework of [21 ] . No normalization of E Tr F \mathbb{E}\operatorname{Tr} F E Tr F is required.
Lemma 10.4 (Complete midpoint entropy bound). For positive definite matrices A A A , B B B , put M = ( A + B ) / 2 M = (A+B)/2 M = ( A + B ) /2 . Then
J ( A , B ) : = 1 2 Tr ( A log A + B log B ) − Tr ( M log M ) ≤ 1 8 Tr [ ( A − B ) ( log A − log B ) ] . (58) J(A,B) := \frac{1}{2}\operatorname{Tr}(A\log A + B\log B) - \operatorname{Tr}(M\log M) \le\frac{1}{8}\operatorname{Tr}[(A-B)(\log A-\log B)]. \tag*{(58)} J ( A , B ) := 2 1 Tr ( A log A + B log B ) − Tr ( M log M ) ≤ 8 1 Tr [( A − B ) ( log A − log B )] . (58) The constant 1 / 8 1/8 1/8 is sharp. Semidefinite inputs follow by regularization, with the usual extended-value interpretation of entropy production.
Proof. Set D = ( A − B ) / 2 D = (A-B)/2 D = ( A − B ) /2 and
φ ( t ) = 1 2 Tr [ ( M + t D ) log ( M + t D ) + ( M − t D ) log ( M − t D ) ] . \varphi(t) = \frac{1}{2}\operatorname{Tr}[(M+tD)\log(M+tD) + (M-tD)\log(M-tD)]. φ ( t ) = 2 1 Tr [( M + t D ) log ( M + t D ) + ( M − t D ) log ( M − t D )] . For s ≥ 0 s \ge0 s ≥ 0 let C s = ( M + s I ) − 1 / 2 D ( M + s I ) − 1 / 2 C_s = (M+sI)^{-1/2}D(M+sI)^{-1/2} C s = ( M + s I ) − 1/2 D ( M + s I ) − 1/2 . Since M ± D > 0 M \pm D > 0 M ± D > 0 , all eigenvalues of C s C_s C s lie in ( − 1 , 1 ) (-1,1) ( − 1 , 1 ) . The resolvent formula for the Fréchet derivative of the logarithm and cyclicity of trace give
φ ′ ′ ( t ) = ∫ 0 ∞ 1 2 Tr [ C s 2 ( ( I + t C s ) − 2 + ( I − t C s ) − 2 ) ] d s . \varphi''(t) = \int_0^\infty\frac{1}{2}\operatorname{Tr}\left[C_s^2\left((I+tC_s)^{-2}+(I-tC_s)^{-2}\right)\right]\,ds. φ ′′ ( t ) = ∫ 0 ∞ 2 1 Tr [ C s 2 ( ( I + t C s ) − 2 + ( I − t C s ) − 2 ) ] d s . Diagonalize C s C_s C s inside this integrand. Its even power-series coefficients are nonnegative. Since φ ′ ( 0 ) = 0 \varphi'(0)=0 φ ′ ( 0 ) = 0 , integrating twice and using monotone convergence yields the exact expansion
φ ( t ) − φ ( 0 ) = ∑ m ≥ 1 t 2 m 2 m ∫ 0 ∞ Tr ( C s 2 m ) d s , 0 ≤ t ≤ 1. (59) \varphi(t)-\varphi(0) = \sum_{m\ge1}\frac{t^{2m}}{2m}\int_0^\infty\operatorname{Tr}(C_s^{2m})\,ds,\qquad0\le t\le1. \tag*{(59)} φ ( t ) − φ ( 0 ) = m ≥ 1 ∑ 2 m t 2 m ∫ 0 ∞ Tr ( C s 2 m ) d s , 0 ≤ t ≤ 1. (59) The coefficients are finite: C s = O ( s − 1 ) C_s = O(s^{-1}) C s = O ( s − 1 ) at infinity, while positive definiteness controls s s s near zero. Nonnegativity of the coefficients gives φ ( 1 ) − φ ( 0 ) ≤ φ ′ ( 1 ) / 2 \varphi(1)-\varphi(0) \le\varphi'(1)/2 φ ( 1 ) − φ ( 0 ) ≤ φ ′ ( 1 ) /2 . Direct differentiation gives
φ ′ ( 1 ) = 1 4 Tr [ ( A − B ) ( log A − log B ) ] , \varphi'(1) = \frac{1}{4}\operatorname{Tr}[(A-B)(\log A-\log B)], φ ′ ( 1 ) = 4 1 Tr [( A − B ) ( log A − log B )] , which proves (58) . Taking scalar A = 1 + ε A = 1+\varepsilon A = 1 + ε , B = 1 − ε B = 1-\varepsilon B = 1 − ε and sending ε \varepsilon ε to zero shows sharpness. Adding ε I \varepsilon I ε I and passing to the limit handles the boundary. □
Theorem 10.5 (Complete entropy decay and the sharp universal rate). Let G G G be a finite graph with uniform law μ \mu μ . For each edge e e e let K e K_e K e average its two matrix values and leave all other values fixed. Assume, for some C > 0 C>0 C > 0 , the complete edge entropy factorization
H μ ( F ) ≤ C ∑ e ∈ E ( G ) [ H μ ( F ) − H μ ( K e F ) ] . (60) \mathcal{H}_{\mu}(F) \le C\sum_{e\in E(G)}[\mathcal{H}_{\mu}(F)-\mathcal{H}_{\mu}(K_eF)]. \tag*{(60)} H μ ( F ) ≤ C e ∈ E ( G ) ∑ [ H μ ( F ) − H μ ( K e F )] . (60) Then the unit-edge Laplacian satisfies
E μ Tr [ ( L G F ) log F ] ≥ 4 C H μ ( F ) , (61) \mathbb{E}_{\mu}\operatorname{Tr}[(L_GF)\log F] \ge\frac{4}{C}\mathcal{H}_{\mu}(F), \tag*{(61)} E μ Tr [( L G F ) log F ] ≥ C 4 H μ ( F ) , (61) H μ ( e − t L G F ) ≤ e − 4 t / C H μ ( F ) ( t ≥ 0 ) . (62) \mathcal{H}_{\mu}(e^{-tL_G}F) \le e^{-4t/C}\mathcal{H}_{\mu}(F)\qquad(t\ge0). \tag*{(62)} H μ ( e − t L G F ) ≤ e − 4 t / C H μ ( F ) ( t ≥ 0 ) . (62) Both statements hold in every finite matrix dimension. In particular, Theorem 10.2 gives C = 1 C=1 C = 1 for every uniform matroid basis graph. The resulting rate four is optimal uniformly over this class, already on a binary edge. For rank r ≥ 2 r\ge2 r ≥ 2 under (15) , one may take C = C κ C=C_{\kappa} C = C κ , giving the stronger rate
4 C κ = 4 r ( 1 + κ ) r + ( r − 1 ) κ . (63) \frac{4}{C_{\kappa}}=\frac{4r(1+\kappa)}{r+(r-1)\kappa}. \tag*{(63)} C κ 4 = r + ( r − 1 ) κ 4 r ( 1 + κ ) . (63)
Every simple matroid of rank at least two admits κ > 0 \kappa> 0 κ > 0 by [38 ] , so this guaranteed rate is strictly greater than four.
Proof. If N N N is the number of vertices, the loss on e = { x , y } e = \{x,y\} e = { x , y } is J e ( F ) = ( 2 / N ) J ( F ( x ) , F ( y ) ) J_e(F) = (2/N)J(F(x),F(y)) J e ( F ) = ( 2/ N ) J ( F ( x ) , F ( y )) . Lemma 10.4 therefore gives
∑ e J e ( F ) ≤ 1 4 N ∑ { x , y } ∈ E ( G ) Tr [ ( F ( x ) − F ( y ) ) ( log F ( x ) − log F ( y ) ) ] = 1 4 E μ Tr [ ( L G F ) log F ] . \sum_e J_e(F) \le\frac{1}{4N}\sum_{\{x,y\}\in E(G)} \operatorname{Tr}[(F(x)-F(y))(\log F(x)-\log F(y))]
= \frac{1}{4}\mathbb{E}_{\mu}\operatorname{Tr}[(L_G F)\log F]. e ∑ J e ( F ) ≤ 4 N 1 { x , y } ∈ E ( G ) ∑ Tr [( F ( x ) − F ( y )) ( log F ( x ) − log F ( y ))] = 4 1 E μ Tr [( L G F ) log F ] .
Combine this with (60) . The derivative of entropy along e − t L G e^{-tL_G} e − t L G is the negative of this production, so integration proves (62) . Regularization gives the semidefinite case.
On a binary edge take F = 1 + ε h F = 1 + \varepsilon h F = 1 + ε h , with E h = 0 \mathbb{E}h = 0 E h = 0 and L G h = 2 h L_Gh = 2h L G h = 2 h . Entropy is quadratic to leading order, so production divided by entropy tends to four; for each fixed t > 0 t > 0 t > 0 , the output-to-input entropy ratio tends to e − 4 t e^{-4t} e − 4 t .
Small sets and scalar hypercontractivity
The scalar consequences below follow from scalar deletion [17 ] and the clique lemma. Montenegro–Son established the small-set profile for balanced matroids [37 ] , and Anari–Liu–Oveis Gharan–Vinzant proved expansion at least one for all matroid basis graphs [5 ] . The complete factorization retains matrix-valued inputs throughout the same argument.
Corollary 10.6 (Small sets and scalar hypercontractivity). Let G G G be the basis graph of a matroid with N N N bases and uniform law μ \mu μ . Every vertex set U U U , with a = min ( ∣ U ∣ , N − ∣ U ∣ ) > 0 a = \min(|U|,N-|U|) > 0 a = min ( ∣ U ∣ , N − ∣ U ∣ ) > 0 , satisfies
∣ ∂ U ∣ ≥ a log 2 ( N / a ) . (64) |\partial U| \ge a\log_2(N/a). \tag*{(64)} ∣ ∂ U ∣ ≥ a log 2 ( N / a ) . (64) For ( L G g ) ( B ) = ∑ B ′ ∼ B ( g ( B ) − g ( B ′ ) ) (L_Gg)(B) = \sum_{B'\sim B}(g(B)-g(B')) ( L G g ) ( B ) = ∑ B ′ ∼ B ( g ( B ) − g ( B ′ )) ,
Ent μ ( g 2 ) ≤ ⟨ g , L G g ⟩ μ = 1 N ∑ { B , B ′ } ∈ E ( g ( B ) − g ( B ′ ) ) 2 . (65) \operatorname{Ent}_{\mu}(g^2) \le\langle g,L_Gg\rangle_{\mu} = \frac{1}{N}\sum_{\{B,B'\}\in E}(g(B)-g(B'))^2. \tag*{(65)} Ent μ ( g 2 ) ≤ ⟨ g , L G g ⟩ μ = N 1 { B , B ′ } ∈ E ∑ ( g ( B ) − g ( B ′ ) ) 2 . (65) Thus e − t L G e^{-tL_G} e − t L G contracts L p ( μ ) L^p(\mu) L p ( μ ) to L q ( μ ) L^q(\mu) L q ( μ ) whenever 1 < p ≤ q 1 < p \le q 1 < p ≤ q and q − 1 ≤ ( p − 1 ) e 4 t q-1 \le(p-1)e^{4t} q − 1 ≤ ( p − 1 ) e 4 t . Both constants are sharp uniformly over matroids. Under (15) , the right sides of (64) and (65) improve by the factors C κ − 1 C_\kappa^{-1} C κ − 1 and C κ C_\kappa C κ , respectively, and hypercontractivity holds when q − 1 ≤ ( p − 1 ) e 4 t / C κ q-1 \le(p-1)e^{4t/C_\kappa} q − 1 ≤ ( p − 1 ) e 4 t / C κ .
Proof. Apply (56) to the indicator of the smaller side. Its entropy is ( a / N ) log ( N / a ) (a/N)\log(N/a) ( a / N ) log ( N / a ) ; each crossing edge contributes ( log 2 ) / N (\log2)/N ( log 2 ) / N . For g 2 g^2 g 2 , use the two-point inequality
u 2 log u 2 + v 2 log v 2 − ( u 2 + v 2 ) log u 2 + v 2 2 ≤ ( u − v ) 2 . u^2\log u^2+v^2\log v^2-(u^2+v^2)\log\frac{u^2+v^2}{2}\le(u-v)^2. u 2 log u 2 + v 2 log v 2 − ( u 2 + v 2 ) log 2 u 2 + v 2 ≤ ( u − v ) 2 .
This inequality is strict unless u = v u=v u = v . For a proof, absolute values reduce to u , v ≥ 0 u,v\ge0 u , v ≥ 0 , and homogeneity reduces a nonzero pair to u 2 + v 2 = 2 u^2+v^2=2 u 2 + v 2 = 2 . Write u 2 = 1 + s u^2=1+s u 2 = 1 + s , v 2 = 1 − s v^2=1-s v 2 = 1 − s , with 0 ≤ s ≤ 1 0\le s\le1 0 ≤ s ≤ 1 . The difference between the right and left sides is
R ( s ) = 2 − 2 1 − s 2 − ( 1 + s ) log ( 1 + s ) − ( 1 − s ) log ( 1 − s ) . R(s)=2-2\sqrt{1-s^2}-(1+s)\log(1+s)-(1-s)\log(1-s). R ( s ) = 2 − 2 1 − s 2 − ( 1 + s ) log ( 1 + s ) − ( 1 − s ) log ( 1 − s ) .
It satisfies R ( 0 ) = R ′ ( 0 ) = 0 R(0)=R'(0)=0 R ( 0 ) = R ′ ( 0 ) = 0 and R ′ ′ ( s ) = 2 ( 1 − s 2 ) − 3 / 2 − 2 ( 1 − s 2 ) − 1 > 0 R''(s)=2(1-s^2)^{-3/2}-2(1-s^2)^{-1}>0 R ′′ ( s ) = 2 ( 1 − s 2 ) − 3/2 − 2 ( 1 − s 2 ) − 1 > 0 for 0 < s < 1 0<s<1 0 < s < 1 . Continuity treats s = 1 s=1 s = 1 , proving the inequality and its strictness.
For hypercontractivity, differentiate log ∥ e − t L G g ∥ q ( t ) \log\|e^{-tL_G}g\|_{q(t)} log ∥ e − t L G g ∥ q ( t ) for positive g g g . The Gross inequality and the edge bound
( u q − 1 − v q − 1 ) ( u − v ) ≥ 4 ( q − 1 ) q 2 ( u q / 2 − v q / 2 ) 2 (u^{q-1}-v^{q-1})(u-v)\ge\frac{4(q-1)}{q^2}(u^{q/2}-v^{q/2})^2 ( u q − 1 − v q − 1 ) ( u − v ) ≥ q 2 4 ( q − 1 ) ( u q /2 − v q /2 ) 2
make this derivative nonpositive when q ′ ≤ 4 ( q − 1 ) q'\le4(q-1) q ′ ≤ 4 ( q − 1 ) . Approximation and positivity of the semigroup treat arbitrary inputs. Retaining C κ C_\kappa C κ in the same edge factorization gives all three improvements, with q ′ ≤ 4 ( q − 1 ) / C κ q'\le4(q-1)/C_\kappa q ′ ≤ 4 ( q − 1 ) / C κ . Direct sums of U 1 , 2 U_{1,2} U 1 , 2 have cube basis graphs: subcubes attain (64) , and perturbations of a constant along one binary factor approach equality in (65) .
A forbidden spectral interval for matroid basis graphs
We continue with the uniform basis law and unit-edge Laplacian L G L_G L G . Write E G ( g , g ) = ⟨ g , L G g ⟩ μ \mathcal{E}_G(g,g)=\langle g,L_Gg\rangle_{\mu} E G ( g , g ) = ⟨ g , L G g ⟩ μ ; by (65) ,
E G ( g , g ) = 1 N ∑ { B , B ′ } ∈ E ( G ) ( g ( B ) − g ( B ′ ) ) 2 . (66) \mathcal{E}_G(g,g)=\frac{1}{N}\sum_{\{B,B'\}\in E(G)}\left(g(B)-g(B')\right)^2. \tag*{(66)} E G ( g , g ) = N 1 { B , B ′ } ∈ E ( G ) ∑ ( g ( B ) − g ( B ′ ) ) 2 . (66)
Loops do not affect the basis graph, and coloops contribute no varying coordinate.
Theorem 10.7 (Universal spectral separation). Let M M M be a finite matroid with exactly k k k direct summands isomorphic to U 1 , 2 U_{1,2} U 1 , 2 . Then
Spec ( L G ) ⊆ { 0 , 2 } ∪ [ 3 , ∞ ) . (67) \operatorname{Spec}(L_G)\subseteq\{0,2\}\cup[3,\infty). \tag*{(67)} Spec ( L G ) ⊆ { 0 , 2 } ∪ [ 3 , ∞ ) . (67) The eigenvalue-two space has dimension k k k and is
ker ( L G − 2 I ) = span { 1 { a ∈ B } − 1 { b ∈ B } : { a , b } is a U 1 , 2 direct summand } . (68) \ker(L_G-2I)=\operatorname{span}\left\{\mathbf{1}_{\{a\in B\}}-\mathbf{1}_{\{b\in B\}}:\{a,b\}\text{ is a }U_{1,2}\text{ direct summand}\right\}. \tag*{(68)} ker ( L G − 2 I ) = span { 1 { a ∈ B } − 1 { b ∈ B } : { a , b } is a U 1 , 2 direct summand } . (68) In particular, if M M M has no U 1 , 2 U_{1,2} U 1 , 2 direct summand and has at least two bases, then λ g a p ( L G ) ≥ 3 \lambda_{\mathrm{gap}}(L_G)\geq3 λ gap ( L G ) ≥ 3 . The threshold three is sharp: the basis graphs of U 1 , 3 U_{1,3} U 1 , 3 and U 2 , 3 U_{2,3} U 2 , 3 are triangles.
The proof uses the following weighted statement. Its rates must be kept explicit: they are the rates produced by collapsing parallel classes in a uniform basis graph.
Lemma 10.8 (Weighted upper-star factorization). Let M M M be a loopless matroid with at least two bases, and let m i > 0 m_i>0 m i > 0 . On its bases put
π ( B ) = 1 Z ∏ i ∈ B m i , ( L m g ) ( B ) = ∑ i ∈ B , j ∉ B B − i + j ∈ B ( M ) m j ( g ( B ) − g ( B − i + j ) ) . \pi(B)=\frac{1}{Z}\prod_{i\in B}m_i,\qquad(L_mg)(B)=\sum_{\substack{i\in B,\ j\notin B\\ B-i+j\in\mathcal{B}(M)}}m_j\left(g(B)-g(B-i+j)\right). π ( B ) = Z 1 i ∈ B ∏ m i , ( L m g ) ( B ) = i ∈ B , j ∈ / B B − i + j ∈ B ( M ) ∑ m j ( g ( B ) − g ( B − i + j ) ) . If g ( M ) g(M) g ( M ) denotes the girth of M M M , then
E m ( g , g ) : = ⟨ g , L m g ⟩ π ≥ g ( M ) min i m i Var π ( g ) . (69) \mathcal{E}_m(g,g):=\langle g,L_mg\rangle_\pi\geq g(M)\min_i m_i\operatorname{Var}_\pi(g). \tag*{(69)} E m ( g , g ) := ⟨ g , L m g ⟩ π ≥ g ( M ) i min m i Var π ( g ) . (69) In particular, a simple matroid with all m i ≥ 1 m_i\geq1 m i ≥ 1 has gap at least three for this weighted generator.
Proof. Write r = rank ( M ) r=\operatorname{rank}(M) r = rank ( M ) and s = ∣ E ( M ) ∣ − r > 0 s=|E(M)|-r>0 s = ∣ E ( M ) ∣ − r > 0 . Under complementation, the law π \pi π is the positive product-field basis law of the dual matroid, with fields 1 / m i 1/m_i 1/ m i . Delete one particle from a dual basis, and let I ∗ I^* I ∗ be the retained dual face of size s − 1 s-1 s − 1 . Its complement U U U has size r + 1 r+1 r + 1 . The possible primal bases inside U U U are B i = U ∖ { i } B_i=U\setminus\{i\} B i = U ∖ { i } , where i i i ranges over the unique circuit C ⊆ U C\subseteq U C ⊆ U . Define
W U = ∑ B ⊂ U π ( B ) , S C = ∑ i ∈ C 1 m i , π U ( i ) = 1 / m i S C . W_U=\sum_{B\subset U}\pi(B),\qquad S_C=\sum_{i\in C}\frac{1}{m_i},\qquad\pi_U(i)=\frac{1/m_i}{S_C}. W U = B ⊂ U ∑ π ( B ) , S C = i ∈ C ∑ m i 1 , π U ( i ) = S C 1/ m i .
The retained-face probability is W U / s W_U/s W U / s . The scalar Lorentzian deletion inequality, or its variance linearization, gives
Var ( E [ g ∣ I ∗ ] ) ≤ s − 1 s Var π ( g ) . \operatorname{Var}\left(\mathbb{E}[g\mid I^*]\right)\leq\frac{s-1}{s}\operatorname{Var}_\pi(g). Var ( E [ g ∣ I ∗ ] ) ≤ s s − 1 Var π ( g ) .
The variance chain rule therefore yields
Var π ( g ) ≤ ∑ U W U Var π U ( g ( B i ) ) . (70) \operatorname{Var}_\pi(g)\leq\sum_U W_U\operatorname{Var}_{\pi_U}\left(g(B_i)\right). \tag*{(70)} Var π ( g ) ≤ U ∑ W U Var π U ( g ( B i ) ) . (70)
Thus the dual rank cancels: the weights in (70) are W U W_U W U , rather than the retained-face probabilities W U / s W_U/s W U / s .
On the clique indexed by C C C , take transition rates q ( i , j ) = m i q(i,j)=m_i q ( i , j ) = m i for i ≠ j i\ne j i = j . They are reversible for π U \pi_U π U , with conductance 1 / S C 1/S_C 1/ S C on every unordered pair. With c = ∣ C ∣ c=|C| c = ∣ C ∣ and D m = diag ( m i ) i ∈ C D_m=\operatorname{diag}(m_i)_{i\in C} D m = diag ( m i ) i ∈ C , the positive generator is
L C = D m ( c I − 11 T ) . L_C=D_m(cI-\mathbf{1}\mathbf{1}^{\mathsf{T}}). L C = D m ( c I − 1 1 T ) .
Its nonzero eigenvalues agree with those of the symmetric matrix
( c I − 11 T ) 1 / 2 D m ( c I − 11 T ) 1 / 2 . (cI-\mathbf{1}\mathbf{1}^{\mathsf{T}})^{1/2}D_m(cI-\mathbf{1}\mathbf{1}^{\mathsf{T}})^{1/2}. ( c I − 1 1 T ) 1/2 D m ( c I − 1 1 T ) 1/2 .
Since D m ⪰ ( min i ∈ C m i ) I D_m \succeq(\min_{i\in C} m_i)I D m ⪰ ( min i ∈ C m i ) I , each such eigenvalue is at least c min i ∈ C m i ≥ g ( M ) min i m i c\min_{i\in C}m_i \ge g(M)\min_i m_i c min i ∈ C m i ≥ g ( M ) min i m i . Consequently
Var π U ( g ) ≤ E C ( g , g ) g ( M ) min i m i . \operatorname{Var}_{\pi_U}(g) \le\frac{\mathcal{E}_C(g,g)}{g(M)\min_i m_i}. Var π U ( g ) ≤ g ( M ) min i m i E C ( g , g ) .
The local energies sum exactly to the global energy. Indeed,
W U S C = ∏ i ∈ U m i Z = π ( U ∖ { i } ) m i . (71) \frac{W_U}{S_C}=\frac{\prod_{i\in U}m_i}{Z}=\pi(U\setminus\{i\})m_i. \tag*{(71)} S C W U = Z ∏ i ∈ U m i = π ( U ∖ { i }) m i . (71)
This is the conductance of the global edge U ∖ { i } ↔ U ∖ { j } U\setminus\{i\}\leftrightarrow U\setminus\{j\} U ∖ { i } ↔ U ∖ { j } , and every exchange edge has a unique union U U U . Hence ∑ U W U E C ( g , g ) = E m ( g , g ) \sum_U W_U\mathcal{E}_C(g,g)=\mathcal{E}_m(g,g) ∑ U W U E C ( g , g ) = E m ( g , g ) . Substitution in (70) proves the claim. □
Corollary 10.9 (Weighted girth and cogirth bounds). Under the hypotheses of Lemma 10.8 ,
λ g a p ( L m ) ≥ max { g ( M ) min i m i , min C ∗ a cocircuit ∑ i ∈ C ∗ m i } . (72) \lambda_{\mathrm{gap}}(L_m)\ge\max\left\{g(M)\min_i m_i,\min_{C^*\text{ a cocircuit}}\sum_{i\in C^*}m_i\right\}. \tag*{(72)} λ gap ( L m ) ≥ max { g ( M ) i min m i , C ∗ a cocircuit min i ∈ C ∗ ∑ m i } . (72) In particular, for the unit-edge basis graph, λ g a p ( L G ) ≥ max { g ( M ) , g ( M ∗ ) } \lambda_{\mathrm{gap}}(L_G)\ge\max\{g(M),g(M^*)\} λ gap ( L G ) ≥ max { g ( M ) , g ( M ∗ )} .
Proof. The first bound is Lemma 10.8 . For the second, use primal deletion to a face I I I of size r − 1 r-1 r − 1 . Its completion set is the cocircuit C ∗ = E ( M ) ∖ cl ( I ) C^*=E(M)\setminus\operatorname{cl}(I) C ∗ = E ( M ) ∖ cl ( I ) . Writing T = ∑ i ∈ C ∗ m i T=\sum_{i\in C^*}m_i T = ∑ i ∈ C ∗ m i , the conditional law on completions is m i / T m_i/T m i / T . The local generator with rates q ( i , j ) = m j q(i,j)=m_j q ( i , j ) = m j is T ( I − E m / T ) T(I-\mathcal{E}_{m/T}) T ( I − E m / T ) , so its gap is exactly T T T when the clique has at least two vertices. A singleton completion has zero conditional variance. The same variance factorization and the identity
W I m i T m j = ( ∏ a ∈ I m a ) m i m j Z = π ( I ∪ { i } ) m j W_I\frac{m_i}{T}m_j=\frac{\left(\prod_{a\in I}m_a\right)m_i m_j}{Z}=\pi(I\cup\{i\})m_j W I T m i m j = Z ( ∏ a ∈ I m a ) m i m j = π ( I ∪ { i }) m j show that the clique energies again sum exactly to E m \mathcal{E}_m E m . This proves the second bound. Set every m i = 1 m_i=1 m i = 1 for the last assertion. □
Lemma 10.10 (Centered sectors on parallel classes). If M M M has no U 1 , 2 U_{1,2} U 1 , 2 direct summand, its unit-edge basis graph has gap at least three whenever it has more than one vertex.
Proof. Apply Corollary 9.5 at uniform original fields. The simplification S S S has fields m C = ∣ C ∣ m_C=|C| m C = ∣ C ∣ , and V C = L C 2 ( C , u n i f ) V_C=L_C^2(C,\mathrm{unif}) V C = L C 2 ( C , unif ) . Write I ( B ) I(B) I ( B ) for the simplified basis and x C ( B ) x_C(B) x C ( B ) for its representative in C C C . With c A = Pr μ ( A ⊆ I ( B ) ) − 1 / 2 c_A=\Pr_\mu(A\subseteq I(B))^{-1/2} c A = Pr μ ( A ⊆ I ( B ) ) − 1/2 , an elementary tensor is realized by
u ( B ) = { c A h ( I ( B ) ) ∏ C ∈ A v C ( x C ( B ) ) , A ⊆ I ( B ) , 0 , A ⊈ I ( B ) , ∑ x ∈ C v C ( x ) = 0. u(B)=
\begin{cases}
c_A h(I(B))\prod_{C\in A}v_C(x_C(B)), & A\subseteq I(B),\\
0, & A\nsubseteq I(B),
\end{cases}
\qquad\sum_{x\in C}v_C(x)=0. u ( B ) = { c A h ( I ( B )) ∏ C ∈ A v C ( x C ( B )) , 0 , A ⊆ I ( B ) , A ⊈ I ( B ) , x ∈ C ∑ v C ( x ) = 0. The coefficient law on I ⊇ A I\supseteq A I ⊇ A is proportional to ∏ C ∈ I m C \prod_{C\in I}m_C ∏ C ∈ I m C . A nonzero contrast requires m C ≥ 2 m_C\ge2 m C ≥ 2 . We compute the generator on this function. For an internal swap in a required class C ∈ A C\in A C ∈ A ,
∑ y ∈ C y ≠ x C ( v C ( x C ) − v C ( y ) ) = m C v C ( x C ) . \sum_{\substack{y\in C\\y\ne x_C}}\left(v_C(x_C)-v_C(y)\right)=m_Cv_C(x_C). y ∈ C y = x C ∑ ( v C ( x C ) − v C ( y ) ) = m C v C ( x C ) . Thus internal swaps contribute ∑ C ∈ A m C \sum_{C\in A}m_C ∑ C ∈ A m C ; swaps in other classes leave u u u unchanged. An exchange I − C + D I-C+D I − C + D retaining A A A preserves every contrast factor. Summing over the m D m_D m D possible inserted representatives gives
( L ( A ) h ) ( I ) = ∑ C ∈ I ∖ A , D ∉ I I − C + D ∈ B ( S ) m D ( h ( I ) − h ( I − C + D ) ) . (L^{(A)}h)(I)=\sum_{\substack{C\in I\setminus A,\ D\notin I\\I-C+D\in\mathcal{B}(S)}}m_D\left(h(I)-h(I-C+D)\right). ( L ( A ) h ) ( I ) = C ∈ I ∖ A , D ∈ / I I − C + D ∈ B ( S ) ∑ m D ( h ( I ) − h ( I − C + D ) ) . If C ∈ A C\in A C ∈ A is removed, the new value is zero. These exchanges therefore give the multiplication operator
V A ( I ) = ∑ C ∈ A ∑ D ∉ I I − C + D ∈ B ( S ) m D . V_A(I)=\sum_{C\in A}\ \sum_{\substack{D\notin I\\I-C+D\in\mathcal{B}(S)}}m_D. V A ( I ) = C ∈ A ∑ D ∈ / I I − C + D ∈ B ( S ) ∑ m D .
Finally take A ⊈ I ( B ) A \not\subseteq I(B) A ⊆ I ( B ) . A neighbor can contain A A A only by inserting the unique missing class C C C . All its representatives occur with unit rate, so their contribution sums to zero by ∑ x ∈ C v C ( x ) = 0 \sum_{x\in C} v_C(x)=0 ∑ x ∈ C v C ( x ) = 0 . This proves invariance, including at the boundary, and gives
L G ∣ H A ≅ [ ( ∑ C ∈ A m C ) I + L ( A ) + V A ] ⊗ I . (73) L_G|_{\mathcal{H}_A} \cong\left[\left(\sum_{C\in A}m_C\right)I+L^{(A)}+V_A\right]\otimes I. \tag*{(73)} L G ∣ H A ≅ [ ( C ∈ A ∑ m C ) I + L ( A ) + V A ] ⊗ I . (73)
Both L ( A ) L^{(A)} L ( A ) and V A V_A V A are nonnegative. For A = ∅ A=\varnothing A = ∅ , Lemma 10.8 gives gap at least three on the simplification; if it has one basis, this sector consists of constants. Every other sector has lower bound at least three unless A = { C } A=\{C\} A = { C } and m C = 2 m_C=2 m C = 2 . This is the only possible sector below three. Here C C C is not a coloop of S S S , since M M M has no U 1 , 2 U_{1,2} U 1 , 2 direct summand. Exchange with a basis omitting C C C gives an exit from C C C at every simplified basis containing it. Thus V A ≥ 1 V_A\geq1 V A ≥ 1 , and the lower bound is again 2 + 1 = 3 2+1=3 2 + 1 = 3 .
Proof of Theorem 10.7 . Split off the binary direct factors:
M = U 1 , 2 ⊕ k ⊕ M 0 . M=U_{1,2}^{\oplus k}\oplus M_0. M = U 1 , 2 ⊕ k ⊕ M 0 .
The basis graph is the Cartesian product of a k k k -cube and the basis graph of M 0 M_0 M 0 , and the unit-edge Laplacians add. The cube has eigenvalues 2 j 2j 2 j , 0 ≤ j ≤ k 0\leq j\leq k 0 ≤ j ≤ k , with the usual degree-j j j binary characters. By Lemma 10.10 , the remaining factor is either a single vertex or has all its nonconstant eigenvalues at least three. The product spectrum consists of their sums. Its only positive eigenvalue below three is therefore two, attained precisely by one binary contrast times constants on the other factors. Triangles show the sharpness of three.
Corollary 10.11 (Feder–Mihail equality; Alaniz Pintos’s Problem 8.4). Let M M M have at least two bases and let G G G be its basis graph. For 0 < p ≤ 1 / Δ ( G ) 0<p\leq1/\Delta(G) 0 < p ≤ 1/Δ ( G ) , put P F M = I − p L G P_{\mathrm{FM}}=I-pL_G P FM = I − p L G . Then
gap ( P F M ) = 2 p ⟺ M has a U 1 , 2 direct summand . \operatorname{gap}(P_{\mathrm{FM}})=2p\quad\Longleftrightarrow\quad M\text{ has a }U_{1,2}\text{ direct summand}. gap ( P FM ) = 2 p ⟺ M has a U 1 , 2 direct summand .
The entire eigenspace at 1 − 2 p 1-2p 1 − 2 p is spanned by the contrasts on these binary summands. If there is no such summand, then gap ( P F M ) ≥ 3 p \operatorname{gap}(P_{\mathrm{FM}})\geq3p gap ( P FM ) ≥ 3 p . Thus the extremal matroids asked for in [2 Problem 8.4] are exactly those with a binary direct component.
Proof. Since I − P F M = p L G I-P_{\mathrm{FM}}=pL_G I − P FM = p L G , the positive gap is p λ g a p ( L G ) p\lambda_{\mathrm{gap}}(L_G) p λ gap ( L G ) and the eigenspaces are unchanged. Apply Theorem 10.7 . This includes rank one: its basis graph is K N K_N K N , with nonzero Laplacian eigenvalue N N N , so equality requires N = 2 N=2 N = 2 .
Corollary 10.12 (Quantitative rigidity, heat flow, and expansion). Let Π b i n \Pi_{\mathrm{bin}} Π bin be the orthogonal projection onto the binary-coordinate contrasts. For every real function g g g ,
E G ( g , g ) − 2 Var μ ( g ) ≥ ∥ g − E μ g − Π b i n g ∥ 2 2 . (74) \mathcal{E}_G(g,g)-2\operatorname{Var}_{\mu}(g)\geq\lVert g-\mathbb{E}_{\mu}g-\Pi_{\mathrm{bin}}g\rVert_2^2. \tag*{(74)} E G ( g , g ) − 2 Var μ ( g ) ≥ ∥ g − E μ g − Π bin g ∥ 2 2 . (74)
The coefficient one is sharp. For mean-zero g g g and t ≥ 0 t\geq0 t ≥ 0 ,
∥ e − t L G g − e − 2 t Π b i n g ∥ 2 ≤ e − 3 t ∥ g − Π b i n g ∥ 2 . (75) \lVert e^{-tL_G}g-e^{-2t}\Pi_{\mathrm{bin}}g\rVert_2\leq e^{-3t}\lVert g-\Pi_{\mathrm{bin}}g\rVert_2. \tag*{(75)} ∥ e − t L G g − e − 2 t Π bin g ∥ 2 ≤ e − 3 t ∥ g − Π bin g ∥ 2 . (75)
If there is no binary direct factor, then every vertex set U U U with 0 < a = ∣ U ∣ ≤ N / 2 0<a=|U|\leq N/2 0 < a = ∣ U ∣ ≤ N /2 satisfies
∣ ∂ U ∣ ≥ 3 a ( 1 − a / N ) ≥ 3 2 a . (76) |\partial U|\geq3a(1-a/N)\geq\frac{3}{2}a. \tag*{(76)} ∣ ∂ U ∣ ≥ 3 a ( 1 − a / N ) ≥ 2 3 a . (76)
Together with (64) , this gives ∣ ∂ U ∣ ≤ a max { log 2 ( N / a ) , 3 ( 1 − a / N ) } |\partial U|\leq a\max\{\log_2(N/a),3(1-a/N)\} ∣ ∂ U ∣ ≤ a max { log 2 ( N / a ) , 3 ( 1 − a / N )} .
Proof. Apply the spectral theorem to (67) . On the orthogonal complement of constants and binary contrasts, L G − 2 I ⪰ I L_G-2I\succeq I L G − 2 I ⪰ I , which proves (74) ; an eigenfunction on a triangle shows sharpness. The same decomposition proves the heat-flow estimate. For the expansion bound apply the gap-three inequality to 1 U \mathbf{1}_U 1 U : Var μ ( 1 U ) = ( a / N ) ( 1 − a / N ) \operatorname{Var}_{\mu}(\mathbf{1}_U)=(a/N)(1-a/N) Var μ ( 1 U ) = ( a / N ) ( 1 − a / N ) and E G ( 1 U , 1 U ) = ∣ ∂ U ∣ / N \mathcal{E}_G(\mathbf{1}_U,\mathbf{1}_U)=|\partial U|/N E G ( 1 U , 1 U ) = ∣ ∂ U ∣/ N .
Remark 10.13. The spectral threshold three is optimal. The expansion conclusion applies, in particular, to every connected matroid with more than two elements.
Corollary 10.14 (Spanning-tree exchange gaps). Let H H H be a finite connected loopless graph with more than one spanning tree. For m e > 0 m_e>0 m e > 0 , give its trees law proportional to ∏ e ∈ T m e \prod_{e\in T}m_e ∏ e ∈ T m e and use the generator
( L e x , μ ) ( T ) = ∑ e ∈ T , f ∉ T T − e + f a tree m f ( u ( T ) − u ( T − e + f ) ) . (L_{\mathrm{ex},\mu})(T)=\sum_{\substack{e\in T,\ f\notin T\\ T-e+f\text{ a tree}}}m_f\bigl(u(T)-u(T-e+f)\bigr). ( L ex , μ ) ( T ) = e ∈ T , f ∈ / T T − e + f a tree ∑ m f ( u ( T ) − u ( T − e + f ) ) . Then
λ g a p ( L e x , m ) ≥ max { girth ( H ) min e m e , min C a bond ∑ e ∈ C m e } . \lambda_{\mathrm{gap}}(L_{\mathrm{ex},m})\ge\max\left\{\operatorname{girth}(H)\min_{e}m_e,\ \min_{C\text{ a bond}}\sum_{e\in C}m_e\right\}. λ gap ( L ex , m ) ≥ max { girth ( H ) e min m e , C a bond min e ∈ C ∑ m e } . For unit rates this is at least the maximum of the girth and edge connectivity of H H H .
Proof. Apply Corollary 10.9 to the graphic matroid: its circuits are cycles and its cocircuits are bonds. □
Theorem 10.15 (All edge-isoperimetric equality sets). Separate all binary components as M = U 1 , 2 ⊕ b ⊕ M 0 M=U_{1,2}^{\oplus b}\oplus M_0 M = U 1 , 2 ⊕ b ⊕ M 0 . For 0 < ∣ U ∣ ≤ N / 2 0<|U|\le N/2 0 < ∣ U ∣ ≤ N /2 , equality in (64) holds if and only if
U = C × B ( M 0 ) , U=C\times B(M_0), U = C × B ( M 0 ) , where C C C is a subcube of { 0 , 1 } b \{0,1\}^b { 0 , 1 } b . Equivalently, U U U fixes one choice in some binary components and leaves every other component unrestricted. In particular ∣ U ∣ / N = 2 − k |U|/N=2^{-k} ∣ U ∣/ N = 2 − k for an integer k ≥ 1 k\ge1 k ≥ 1 . Every nontrivial cut in a connected matroid with more than two active elements satisfies strict inequality.
Proof. For rank one, G = K N G=K_N G = K N and the claim follows directly from ∣ ∂ U ∣ = ∣ U ∣ ( N − ∣ U ∣ ) |\partial U|=|U|(N-|U|) ∣ ∂ U ∣ = ∣ U ∣ ( N − ∣ U ∣ ) ; equality requires N = 2 N=2 N = 2 , ∣ U ∣ = 1 |U|=1 ∣ U ∣ = 1 . Assume r ≥ 2 r\ge2 r ≥ 2 . Equality in the edge bound requires equality in deletion and in each conditional clique bound, since every discarded term is nonnegative. Apply Theorem 5.6 to ( N / ∣ U ∣ ) 1 U (N/|U|)\mathbf{1}_U ( N /∣ U ∣ ) 1 U . Its product-tilt form shows that U U U consists of the bases avoiding selected forbidden labels; every parallel class contains an allowed label.
If a completion clique has c c c vertices, with 0 < u < c 0<u<c 0 < u < c inside U U U , equality in its indicator factorization requires
log 2 ( c / u ) = c − u . \log_2(c/u)=c-u. log 2 ( c / u ) = c − u .
Writing k = c − u ≥ 1 k=c-u\ge1 k = c − u ≥ 1 gives log 2 ( 1 + k / u ) ≤ log 2 ( 1 + k ) ≤ k \log_2(1+k/u)\le\log_2(1+k)\le k log 2 ( 1 + k / u ) ≤ log 2 ( 1 + k ) ≤ k , with equality only when u = k = 1 u=k=1 u = k = 1 . Thus every mixed completion clique has exactly two vertices.
A parallel class containing both allowed and forbidden labels must therefore have size two: extend that class to a simplified basis and choose allowed labels in its other classes; the resulting facet has a mixed completion clique containing the whole class. If its simplified element is not a coloop, choose an exchange removing that class from a simplified basis, and choose allowed representatives in all the other classes. The resulting facet has a mixed completion clique containing the two labels of the cut class and an allowed label from another class. This contradicts its required size. Every cut class is consequently a binary direct summand, giving the asserted cylinder. Conversely, fixing k k k binary choices gives N 2 − k N2^{-k} N 2 − k bases, each with k k k outgoing edges, and attains the bound. □
This recovers the classical subcube equality cases on the discrete cube [20 ] and identifies their exact occurrence inside every matroid basis graph.
Corollary 10.16 (Sharpness of the Gross constant). For a matroid with at least two bases, the optimal constant C G C_G C G in Ent μ ( g 2 ) ≤ C G ⟨ g , L G g ⟩ μ \operatorname{Ent}_{\mu}(g^2)\le C_G\langle g,L_Gg\rangle_{\mu} Ent μ ( g 2 ) ≤ C G ⟨ g , L G g ⟩ μ equals one if and only if M M M has a U 1 , 2 U_{1,2} U 1 , 2 direct summand.
Proof. A perturbation of a constant by a binary-factor contrast gives ratios tending to one. Conversely, the two-point inequality in the proof of (65) is strict on every edge with different endpoint values. The basis graph is connected, so the ratio is strictly below one for every nonconstant g g g . Normalize ∥ g ∥ 2 = 1 \|g\|_2=1 ∥ g ∥ 2 = 1 . Near either constant function write g = a 1 + h g=a\mathbf{1}+h g = a 1 + h , where E μ h = 0 \mathbb{E}_{\mu}h=0 E μ h = 0 . In this finite-dimensional space, Taylor expansion gives, uniformly in the direction of h h h ,
Ent μ ( g 2 ) = 2 ∥ h ∥ 2 2 + o ( ∥ h ∥ 2 2 ) , ⟨ g , L G g ⟩ μ = ⟨ h , L G h ⟩ μ . \operatorname{Ent}_{\mu}(g^2)=2\|h\|_2^2+o(\|h\|_2^2),\qquad\langle g,L_Gg\rangle_{\mu}=\langle h,L_Gh\rangle_{\mu}. Ent μ ( g 2 ) = 2∥ h ∥ 2 2 + o ( ∥ h ∥ 2 2 ) , ⟨ g , L G g ⟩ μ = ⟨ h , L G h ⟩ μ .
Thus the limiting upper bound for the ratio is 2 / λ g a p ( L G ) 2/\lambda_{\mathrm{gap}}(L_G) 2/ λ gap ( L G ) . If there is no binary summand, Theorem 10.7 makes this strictly less than one. The rest of the normalized sphere is compact and has a continuous ratio strictly below one, completing the proof. □
Matrix concentration for arbitrary nonlinear observables
We now turn entropy decay into concentration for an arbitrary Hermitian observable. The proof has two stages. First, follow a matrix density along the reversible heat flow and bound the change in its expectation of the observable. Then optimize over densities in the Gibbs variational formula. The matrix mean survives both stages, keeping the variance matrix inside the final exponential and allowing an arbitrary Hermitian offset.
Let μ \mu μ be a positive probability measure on a finite set Ω \Omega Ω , and let
( L A ) ( x ) = ∑ y q ( x , y ) ( A ( x ) − A ( y ) ) , μ ( x ) q ( x , y ) = μ ( y ) q ( y , x ) , q ( x , y ) ≥ 0 ( x ≠ y ) , (LA)(x)=\sum_y q(x,y)(A(x)-A(y)),\qquad\mu(x)q(x,y)=\mu(y)q(y,x),\qquad q(x,y)\ge0\quad(x\ne y), ( L A ) ( x ) = y ∑ q ( x , y ) ( A ( x ) − A ( y )) , μ ( x ) q ( x , y ) = μ ( y ) q ( y , x ) , q ( x , y ) ≥ 0 ( x = y ) ,
be an irreducible reversible Markov generator, with P t = e − t L P_t=e^{-tL} P t = e − t L . For a Hermitian matrix function A A A , its matrix carré du champ is
Γ L ( A ) ( x ) = 1 2 ∑ y q ( x , y ) ( A ( x ) − A ( y ) ) 2 . \Gamma_L(A)(x)=\frac{1}{2}\sum_y q(x,y)(A(x)-A(y))^2. Γ L ( A ) ( x ) = 2 1 y ∑ q ( x , y ) ( A ( x ) − A ( y ) ) 2 .
We use the entropy-production convention
α H μ ( G ) ≤ I L ( G ) : = E μ Tr [ ( L G ) log G ] , G > 0. (77) \alpha H_{\mu}(G)\le I_L(G):=\mathbb{E}_{\mu}\operatorname{Tr}[(LG)\log G],\qquad G>0. \tag*{(77)} α H μ ( G ) ≤ I L ( G ) := E μ Tr [( L G ) log G ] , G > 0. (77)
When (11.1) holds in every matrix dimension with the same α > 0 \alpha>0 α > 0 , it is a complete modified logarithmic Sobolev inequality.
Transport and the matrix Laplace comparison
For a density with matrix mean M M M , variance is measured by Tr ( M V ) \operatorname{Tr}(MV) Tr ( M V ) . This state-dependent quantity is the link between transport and the matrix Laplace transform.
Theorem 11.1 (Matrix transport and Laplace comparison). Suppose (11.1) holds in matrix dimension d d d . Let A A A take values in the Hermitian d × d d\times d d × d matrices, put A ‾ = E μ A \overline{A}=\mathbb{E}_{\mu}A A = E μ A , and suppose a fixed positive semidefinite matrix V V V satisfies
Γ L ( A ) ( x ) ⪯ V ( x ∈ Ω ) . (78) \Gamma_L(A)(x)\preceq V\qquad(x\in\Omega). \tag*{(78)} Γ L ( A ) ( x ) ⪯ V ( x ∈ Ω ) . (78) For every G ⪰ 0 G\succeq0 G ⪰ 0 with Tr E μ G = 1 \operatorname{Tr}\mathbb{E}_{\mu}G=1 Tr E μ G = 1 , writing M = E μ G M=\mathbb{E}_{\mu}G M = E μ G ,
∣ E μ Tr [ ( A − A ‾ ) G ] ∣ ≤ 2 Tr ( M V ) α H μ ( G ) . (79) \left|\mathbb{E}_{\mu}\operatorname{Tr}[(A-\overline{A})G]\right|\le2\sqrt{\frac{\operatorname{Tr}(MV)}{\alpha}H_{\mu}(G)}. \tag*{(79)} E μ Tr [( A − A ) G ] ≤ 2 α Tr ( M V ) H μ ( G ) . (79) For every Hermitian K K K and every real θ \theta θ ,
E μ Tr e K + θ A ≤ Tr exp ( K + θ A ‾ + θ 2 α V ) . (80) \mathbb{E}_{\mu}\operatorname{Tr}e^{K+\theta A}\le\operatorname{Tr}\exp\left(K+\theta\overline{A}+\frac{\theta^2}{\alpha}V\right). \tag*{(80)} E μ Tr e K + θ A ≤ Tr exp ( K + θ A + α θ 2 V ) . (80) Both conclusions hold uniformly in d d d under the complete inequality.
Proof. Subtract A ‾ \overline{A} A first.
The edge estimate. Entropy production pairs a difference of densities with a difference of logarithms. Their logarithmic mean converts this pairing into a Hilbert–space Cauchy–Schwarz inequality. For positive definite matrices B B B , C C C , define the positive self-adjoint operator on the Hilbert–Schmidt space
J B , C ( X ) = ∫ 0 1 B s X C 1 − s d s . \mathcal{J}_{B,C}(X)=\int_0^1 B^sXC^{1-s}\,ds. J B , C ( X ) = ∫ 0 1 B s X C 1 − s d s .
The operators of left multiplication by B B B and right multiplication by C C C commute and are positive. Differentiating B s C 1 − s B^{s}C^{1-s} B s C 1 − s gives
B − C = J B , C ( log B − log C ) . (81) B-C=\mathcal{J}_{B,C}(\log B-\log C). \tag*{(81)} B − C = J B , C ( log B − log C ) . (81)
For Hermitian X X X , the weighted arithmetic–geometric mean inequality, applied in eigenbases of B B B and C C C , gives
Tr [ X J B , C ( X ) ] ≤ 1 2 Tr [ X 2 ( B + C ) ] . (82) \operatorname{Tr}[X\mathcal{J}_{B,C}(X)]\leq\frac{1}{2}\operatorname{Tr}[X^{2}(B+C)]. \tag*{(82)} Tr [ X J B , C ( X )] ≤ 2 1 Tr [ X 2 ( B + C )] . (82)
Indeed, before integration the coefficient of the squared absolute value of the ( i , j ) (i,j) ( i , j ) matrix entry of X X X is b i s c j 1 − s b_i^{s}c_j^{1-s} b i s c j 1 − s , which is at most s b i + ( 1 − s ) c j sb_i+(1-s)c_j s b i + ( 1 − s ) c j .
Assume first that every G ( x ) G(x) G ( x ) is positive definite. Reversibility, (81) , and Cauchy–Schwarz in the direct sum of the Hilbert–Schmidt spaces with weights μ ( x ) q ( x , y ) / 2 \mu(x)q(x,y)/2 μ ( x ) q ( x , y ) /2 yield
∣ E μ Tr [ A L G ] ∣ 2 ≤ I L ( G ) 1 2 ∑ x , y μ ( x ) q ( x , y ) Tr [ ( A ( x ) − A ( y ) ) J G ( x ) , G ( y ) ( A ( x ) − A ( y ) ) ] ≤ I L ( G ) E μ Tr [ G Γ L ( A ) ] ≤ I L ( G ) Tr ( M V ) . (83) \begin{aligned}
\left|\mathbb{E}_{\mu}\operatorname{Tr}[A\,LG]\right|^{2}\leq\mathcal{I}_{L}(G)\frac{1}{2}\sum_{x,y}\mu(x)q(x,y)\operatorname{Tr}\left[(A(x)-A(y))\mathcal{J}_{G(x),G(y)}(A(x)-A(y))\right] \\
&\leq\mathcal{I}_{L}(G)\mathbb{E}_{\mu}\operatorname{Tr}[G\Gamma_{L}(A)]\leq\mathcal{I}_{L}(G)\operatorname{Tr}(MV). \tag*{(83)}
\end{aligned} ∣ E μ Tr [ A L G ] ∣ 2 ≤ I L ( G ) 2 1 x , y ∑ μ ( x ) q ( x , y ) Tr [ ( A ( x ) − A ( y )) J G ( x ) , G ( y ) ( A ( x ) − A ( y )) ] ≤ I L ( G ) E μ Tr [ G Γ L ( A )] ≤ I L ( G ) Tr ( M V ) . (83)
The second inequality is (82) ; the two endpoint terms agree after exchanging x x x and y y y .
Integration along the heat flow. Set G t = P t G G_t=P_tG G t = P t G , h ( t ) = H μ ( G t ) h(t)=\mathcal{H}_{\mu}(G_t) h ( t ) = H μ ( G t ) , and i ( t ) = I L ( G t ) i(t)=\mathcal{I}_{L}(G_t) i ( t ) = I L ( G t ) . Its matrix mean remains M M M , h ′ ( t ) = − i ( t ) h'(t)=-i(t) h ′ ( t ) = − i ( t ) , and i ( t ) ≥ α h ( t ) i(t)\geq\alpha h(t) i ( t ) ≥ α h ( t ) . Moreover G t → M G_t\to M G t → M and h ( t ) → 0 h(t)\to0 h ( t ) → 0 . Integrating (83) along the semigroup gives
∣ E μ Tr [ A G ] ∣ ≤ Tr ( M V ) ∫ 0 ∞ i ( t ) d t ≤ Tr ( M V ) α ∫ 0 ∞ i ( t ) h ( t ) d t = 2 Tr ( M V ) α h ( 0 ) . \left|\mathbb{E}_{\mu}\operatorname{Tr}[AG]\right|\leq\sqrt{\operatorname{Tr}(MV)}\int_{0}^{\infty}\sqrt{i(t)}\,dt\leq\sqrt{\frac{\operatorname{Tr}(MV)}{\alpha}}\int_{0}^{\infty}\frac{i(t)}{\sqrt{h(t)}}\,dt=2\sqrt{\frac{\operatorname{Tr}(MV)}{\alpha}}h(0). ∣ E μ Tr [ A G ] ∣ ≤ Tr ( M V ) ∫ 0 ∞ i ( t ) d t ≤ α Tr ( M V ) ∫ 0 ∞ h ( t ) i ( t ) d t = 2 α Tr ( M V ) h ( 0 ) .
The integrals stop if h h h reaches zero. If h ( 0 ) = 0 h(0)=0 h ( 0 ) = 0 , the density is constant and the assertion is immediate. Replacing G G G by ( G + ε I ) / ( 1 + ε d ) (G+\varepsilon I)/(1+\varepsilon d) ( G + ε I ) / ( 1 + ε d ) and passing to the limit proves (79) for all G ⪰ 0 G\succeq0 G ⪰ 0 .
Optimization over the matrix mean. Young’s inequality applied to (79) gives
θ E μ Tr [ A G ] − H μ ( G ) ≤ θ 2 α Tr ( M V ) . \theta\mathbb{E}_{\mu}\operatorname{Tr}[AG]-\mathcal{H}_{\mu}(G)\leq\frac{\theta^{2}}{\alpha}\operatorname{Tr}(MV). θ E μ Tr [ A G ] − H μ ( G ) ≤ α θ 2 Tr ( M V ) .
The Gibbs variational formula on the finite direct sum of matrix algebras, with the weights μ ( x ) \mu(x) μ ( x ) , reads
log E μ Tr e K + θ A = sup G ⪰ 0 Tr E μ G = 1 { Tr ( M K ) + θ E μ Tr [ A G ] − H μ ( G ) − Tr ( M log M ) } ≤ sup M ⪰ 0 Tr M = 1 { Tr [ M ( K + θ 2 V / α ) ] − Tr ( M log M ) } = log Tr e K + θ 2 V / α . \begin{aligned}
\log\mathbb{E}_{\mu}\operatorname{Tr}e^{K+\theta A}
=\sup_{\substack{G\succeq0\\ \operatorname{Tr}\mathbb{E}_{\mu}G=1}}\{\operatorname{Tr}(MK)+\theta\mathbb{E}_{\mu}\operatorname{Tr}[AG]-\mathcal{H}_{\mu}(G)-\operatorname{Tr}(M\log M)\} \\
&\leq\sup_{\substack{M\succeq0\\ \operatorname{Tr}M=1}}\{\operatorname{Tr}[M(K+\theta^{2}V/\alpha)]-\operatorname{Tr}(M\log M)\}
=\log\operatorname{Tr}e^{K+\theta^{2}V/\alpha}.
\end{aligned} log E μ Tr e K + θ A = G ⪰ 0 Tr E μ G = 1 sup { Tr ( M K ) + θ E μ Tr [ A G ] − H μ ( G ) − Tr ( M log M )} ≤ M ⪰ 0 Tr M = 1 sup { Tr [ M ( K + θ 2 V / α )] − Tr ( M log M )} = log Tr e K + θ 2 V / α .
This proves the centered version of (80) . Replacing K K K by K + θ A ‾ K+\theta\overline{A} K + θ A proves the stated formula.
Huang–Tropp [28 ] obtain nonlinear matrix subgaussian concentration from the Bakry–Émery curvature criterion. Gao–Junge–LaRacuente [21, Theorems 6.7 and 6.10] prove noncommutative transport from modified logarithmic Sobolev inequalities and its equivalence to conditional L p L_p L p concentration. See also Gao–Junge–LaRacuente–Li [22, Section 5.4] and De Palma–Rouzé [18 ] . The proof above follows the same entropy-decay argument on a finite reversible chain. It keeps Tr ( M V ) \operatorname{Tr}(MV) Tr ( M V ) in the transport estimate, where the Lipschitz seminorm in the cited results uses ∥ V ∥ \lVert V\rVert ∥ V ∥ . Optimizing the matrix mean M M M then retains both V V V and the arbitrary offset K K K in the trace comparison.
Corollary 11.2 (Subgaussian matrix concentration). Under Theorem 11.1 , put v = ∥ V ∥ v = \lVert V\rVert v = ∥ V ∥ . If v > 0 v > 0 v > 0 , then for every t > 0 t > 0 t > 0 ,
Pr { λ max ( A − A ‾ ) ≥ t } ≤ inf θ > 0 e − θ t Tr e θ 2 V / α ≤ d exp ( − α t 2 4 v ) (84) \Pr\{\lambda_{\max}(A-\overline{A}) \ge t\} \le\inf_{\theta>0} e^{-\theta t}\operatorname{Tr}e^{\theta^{2}V/\alpha} \le d\exp\left(-\frac{\alpha t^{2}}{4v}\right) \tag*{(84)} Pr { λ m a x ( A − A ) ≥ t } ≤ θ > 0 inf e − θ t Tr e θ 2 V / α ≤ d exp ( − 4 v α t 2 ) (84) Pr { ∥ A − A ‾ ∥ ≥ t } ≤ 2 d exp ( − α t 2 4 v ) . (85) \Pr\{\lVert A-\overline{A}\rVert\ge t\} \le2d\exp\left(-\frac{\alpha t^{2}}{4v}\right). \tag*{(85)} Pr {∥ A − A ∥ ≥ t } ≤ 2 d exp ( − 4 v α t 2 ) . (85) The one-sided bound has the intrinsic-dimension refinement
Pr { λ max ( A − A ‾ ) ≥ t } ≤ e Tr V ∥ V ∥ exp ( − α t 2 4 v ) . (86) \Pr\{\lambda_{\max}(A-\overline{A}) \ge t\} \le e^{\frac{\operatorname{Tr}V}{\lVert V\rVert}}\exp\left(-\frac{\alpha t^{2}}{4v}\right). \tag*{(86)} Pr { λ m a x ( A − A ) ≥ t } ≤ e ∥ V ∥ Tr V exp ( − 4 v α t 2 ) . (86) If V = 0 V = 0 V = 0 , the observable is constant on the reference support.
Proof. Take K = 0 K = 0 K = 0 in (80) for A − A ‾ A-\overline{A} A − A , apply the trace Laplace bound [39 ] , and optimize at θ = α t / ( 2 v ) \theta= \alpha t/(2v) θ = α t / ( 2 v ) . Apply the same argument to − A -A − A for the opposite tail. If V = 0 V = 0 V = 0 , every edge difference vanishes.
For (86) , let ψ ( u ) = e u − u − 1 \psi(u) = e^{u}-u-1 ψ ( u ) = e u − u − 1 and X = A − A ‾ X = A-\overline{A} X = A − A . Since E X = 0 \mathbb{E}X = 0 E X = 0 ,
E Tr ψ ( θ X ) ≤ Tr ( e θ 2 V / α − I ) ≤ Tr V v ( e θ 2 v / α − 1 ) . \mathbb{E}\operatorname{Tr}\psi(\theta X) \le\operatorname{Tr}\left(e^{\theta^{2}V/\alpha}-I\right) \le\frac{\operatorname{Tr}V}{v}\left(e^{\theta^{2}v/\alpha}-1\right). E Tr ψ ( θ X ) ≤ Tr ( e θ 2 V / α − I ) ≤ v Tr V ( e θ 2 v / α − 1 ) .
On the upper-tail event the left-hand random variable is at least ψ ( θ t ) \psi(\theta t) ψ ( θ t ) . Put E = α t 2 / ( 4 v ) E=\alpha t^{2}/(4v) E = α t 2 / ( 4 v ) and use the same θ \theta θ , so that θ t = 2 E \theta t=2E θ t = 2 E and θ 2 v / α = E \theta^{2}v/\alpha=E θ 2 v / α = E . For E ≥ 1 E\ge1 E ≥ 1 , ( e E − 1 ) / ( e 2 E − 2 E − 1 ) ≤ e 1 − E (e^{E}-1)/(e^{2E}-2E-1)\le e^{1-E} ( e E − 1 ) / ( e 2 E − 2 E − 1 ) ≤ e 1 − E . For E < 1 E<1 E < 1 the claimed bound exceeds one because Tr V / v ≥ 1 \operatorname{Tr}V/v\ge1 Tr V / v ≥ 1 . □
Affine variance and Bernstein bounds
An affine variance bound allows the local fluctuations to grow with the observable itself. Along the heat flow, its square root has a bounded rate of change. Integrating that bound and completing one square gives the Bernstein denominator, while retaining the full negative-parameter range.
Theorem 11.3 (An affine carré-du-champ bound). Suppose (77) holds in dimension d d d . Let A A A be a centered Hermitian observable and suppose
Γ L ( A ) ( x ) ⪯ a A ( x ) + B ( x ∈ Ω ) , a ≥ 0 , B ⪰ 0 , (87) \Gamma_L(A)(x) \preceq aA(x)+B \qquad(x\in\Omega), \qquad a\ge0,\quad B\succeq0, \tag*{(87)} Γ L ( A ) ( x ) ⪯ a A ( x ) + B ( x ∈ Ω ) , a ≥ 0 , B ⪰ 0 , (87) where a a a and B B B are fixed. For every Hermitian K K K ,
E μ Tr e K + θ A ≤ Tr exp ( K + θ 2 α − a θ B ) ( θ < α / a ) . (88) \mathbb{E}_{\mu}\operatorname{Tr}e^{K+\theta A}\le\operatorname{Tr}\exp\left(K+\frac{\theta^{2}}{\alpha-a\theta}B\right) \qquad(\theta<\alpha/a). \tag*{(88)} E μ Tr e K + θ A ≤ Tr exp ( K + α − a θ θ 2 B ) ( θ < α / a ) . (88) When a = 0 a=0 a = 0 , the parameter range is all of R \mathbb{R} R . When a > 0 a>0 a > 0 , every normalized matrix density G G G satisfies
∣ a E μ Tr ( A G ) + Tr ( M B ) − Tr ( M B ) ∣ ≤ a H μ ( G ) / α , M = E μ G . (89) \left|\sqrt{a\mathbb{E}_{\mu}\operatorname{Tr}(AG)+\operatorname{Tr}(MB)}-\sqrt{\operatorname{Tr}(MB)}\right|\le a\sqrt{\mathcal{H}_{\mu}(G)/\alpha}, \qquad M=\mathbb{E}_{\mu}G. \tag*{(89)} a E μ Tr ( A G ) + Tr ( M B ) − Tr ( M B ) ≤ a H μ ( G ) / α , M = E μ G . (89) Thus the comparison includes all negative θ \theta θ as well.
Proof. The case a = 0 a=0 a = 0 is Theorem 11.1 . Suppose a > 0 a>0 a > 0 and first take G > 0 G>0 G > 0 . Set G t = P t G G_t=P_tG G t = P t G , f ( t ) = E μ Tr ( A G t ) f(t)=\mathbb{E}_{\mu}\operatorname{Tr}(AG_t) f ( t ) = E μ Tr ( A G t ) , h ( t ) = H μ ( G t ) h(t)=\mathcal{H}_{\mu}(G_t) h ( t ) = H μ ( G t ) , i ( t ) = I L ( G t ) i(t)=\mathcal{I}_L(G_t) i ( t ) = I L ( G t ) , and b = Tr ( M B ) b=\operatorname{Tr}(MB) b = Tr ( M B ) . The Cauchy–Schwarz step in (83) and (87) give
∣ f ′ ( t ) ∣ 2 ≤ i ( t ) E μ Tr [ G t Γ L ( A ) ] ≤ i ( t ) [ a f ( t ) + b ] . |f'(t)|^{2}\le i(t)\mathbb{E}_{\mu}\operatorname{Tr}[G_t\Gamma_L(A)]\le i(t)[af(t)+b]. ∣ f ′ ( t ) ∣ 2 ≤ i ( t ) E μ Tr [ G t Γ L ( A )] ≤ i ( t ) [ a f ( t ) + b ] .
The quantity a f ( t ) + b af(t)+b a f ( t ) + b is nonnegative, since a A + B ⪰ Γ L ( A ) ⪰ 0 aA+B\succeq\Gamma_L(A)\succeq0 a A + B ⪰ Γ L ( A ) ⪰ 0 . For ε > 0 \varepsilon>0 ε > 0 , let g ε ( t ) = a f ( t ) + b + ε g_{\varepsilon}(t)=\sqrt{af(t)+b+\varepsilon} g ε ( t ) = a f ( t ) + b + ε . Then ∣ g ε ′ ( t ) ∣ ≤ ( a / 2 ) i ( t ) |g_{\varepsilon}'(t)|\le(a/2)\sqrt{i(t)} ∣ g ε ′ ( t ) ∣ ≤ ( a /2 ) i ( t ) . Since f ( t ) → 0 f(t)\to0 f ( t ) → 0 , integration as in Theorem 11.1 yields
∣ a f ( 0 ) + b + ε − b + ε ∣ ≤ a 2 ∫ 0 ∞ i ( t ) d t ≤ a h ( 0 ) / α . \left|\sqrt{af(0)+b+\varepsilon}-\sqrt{b+\varepsilon}\right|\le\frac{a}{2}\int_{0}^{\infty}\sqrt{i(t)}\,dt\le a\sqrt{h(0)/\alpha}. a f ( 0 ) + b + ε − b + ε ≤ 2 a ∫ 0 ∞ i ( t ) d t ≤ a h ( 0 ) / α .
Letting ε ↓ 0 \varepsilon\downarrow0 ε ↓ 0 and then regularizing G G G proves (89) , including b = 0 b=0 b = 0 .
For brevity put f = E μ Tr ( A G ) f=\mathbb{E}_{\mu}\operatorname{Tr}(AG) f = E μ Tr ( A G ) and h = H μ ( G ) h=\mathcal{H}_{\mu}(G) h = H μ ( G ) . Writing u = a f + b u=\sqrt{af+b} u = a f + b , the transport bound gives h ≥ ( α / a 2 ) ( u − b ) 2 h\geq(\alpha/a^{2})(u-\sqrt{b})^{2} h ≥ ( α / a 2 ) ( u − b ) 2 . For every θ < α / a \theta<\alpha/a θ < α / a , completing the square gives
θ f − h ≤ θ a ( u 2 − b ) − α a 2 ( u − b ) 2 = θ 2 b α − a θ − α − a θ a 2 ( u − α b α − a θ ) 2 ≤ θ 2 α − a θ Tr ( M B ) . \begin{aligned}
\theta f-h\leq\frac{\theta}{a}(u^{2}-b)-\frac{\alpha}{a^{2}}(u-\sqrt{b})^{2} \\
&=\frac{\theta^{2}b}{\alpha-a\theta}-\frac{\alpha-a\theta}{a^{2}}\left(u-\frac{\alpha\sqrt{b}}{\alpha-a\theta}\right)^{2}\leq\frac{\theta^{2}}{\alpha-a\theta}\operatorname{Tr}(MB).
\end{aligned} θ f − h ≤ a θ ( u 2 − b ) − a 2 α ( u − b ) 2 = α − a θ θ 2 b − a 2 α − a θ ( u − α − a θ α b ) 2 ≤ α − a θ θ 2 Tr ( M B ) .
The Gibbs variational formula, followed by optimization over M M M , now proves (11.12). □
Affine conditional variance bounds also appear in the matrix Stein-pair inequalities of Mackey–Jordan–Chen–Farrell–Tropp [36 Theorem 4.1] . Theorem 11.3 uses complete entropy production for an arbitrary nonlinear observable and retains a non-scalar matrix B B B and every Hermitian offset K K K .
Corollary 11.4 (Nonlinear matrix Bernstein bounds). Under Theorem 11.3, put v = ∥ B ∥ v=\lVert B\rVert v = ∥ B ∥ . If v > 0 v>0 v > 0 , then for t > 0 t>0 t > 0 ,
Pr { λ max ( A ) ≥ t } ≤ inf 0 < θ < α / a e − θ t Tr e θ 2 B / ( α − a θ ) ≤ d exp ( − α t 2 4 v + 2 a t ) . (90) \Pr\{\lambda_{\max}(A)\geq t\}\leq\inf_{0<\theta<\alpha/a}e^{-\theta t}\operatorname{Tr}e^{\theta^{2}B/(\alpha-a\theta)}\leq d\exp\left(-\frac{\alpha t^{2}}{4v+2at}\right). \tag*{(90)} Pr { λ m a x ( A ) ≥ t } ≤ 0 < θ < α / a inf e − θ t Tr e θ 2 B / ( α − a θ ) ≤ d exp ( − 4 v + 2 a t α t 2 ) . (90)
Pr { λ min ( A ) ≤ − t } ≤ d exp ( − α t 2 4 v ) . (91) \Pr\{\lambda_{\min}(A)\leq-t\}\leq d\exp\left(-\frac{\alpha t^{2}}{4v}\right). \tag*{(91)} Pr { λ m i n ( A ) ≤ − t } ≤ d exp ( − 4 v α t 2 ) . (91)
For a > 0 a>0 a > 0 , the optimized scalar exponent in the upper tail is
Pr { λ max ( A ) ≥ t } ≤ d exp [ − α a 2 ( v + a t − v ) 2 ] . (92) \Pr\{\lambda_{\max}(A)\geq t\}\leq d\exp\left[-\frac{\alpha}{a^{2}}\left(\sqrt{v+at}-\sqrt{v}\right)^{2}\right]. \tag*{(92)} Pr { λ m a x ( A ) ≥ t } ≤ d exp [ − a 2 α ( v + a t − v ) 2 ] . (92)
For a > 0 a>0 a > 0 and 0 < t < v / a 0<t<v/a 0 < t < v / a , the lower-tail exponent improves to − α ( v − v − a t ) 2 / a 2 -\alpha(\sqrt{v}-\sqrt{v-at})^{2}/a^{2} − α ( v − v − a t ) 2 / a 2 ; for t ≥ v / a t\geq v/a t ≥ v / a the lower event is impossible. The elementary upper bound (11.14) also holds with prefactor e Tr ( B ) / ∥ B ∥ e\operatorname{Tr}(B)/\lVert B\rVert e Tr ( B ) / ∥ B ∥ . If B = 0 B=0 B = 0 , then A = 0 A=0 A = 0 on the reference support.
Proof. Apply the trace Laplace bound to (11.12). The positive choice θ = α t / ( 2 v + a t ) \theta=\alpha t/(2v+at) θ = α t / ( 2 v + a t ) gives (11.14); direct optimization for a > 0 a>0 a > 0 gives θ = ( α / a ) ( 1 − v / ( v + a t ) ) \theta=(\alpha/a)(1-\sqrt{v/(v+at)}) θ = ( α / a ) ( 1 − v / ( v + a t ) ) and (11.16). For negative θ \theta θ , θ 2 / ( α − a θ ) ≤ θ 2 / α \theta^{2}/(\alpha-a\theta)\leq\theta^{2}/\alpha θ 2 / ( α − a θ ) ≤ θ 2 / α , which gives (11.15). Direct optimization gives the stronger stated lower exponent. If a > 0 a>0 a > 0 , then a A + B ⪰ 0 aA+B\succeq0 a A + B ⪰ 0 gives A ⪰ − B / a ⪰ − ( v / a ) I A\succeq-B/a\succeq-(v/a)I A ⪰ − B / a ⪰ − ( v / a ) I . Equality on a unit vector would force every adjacent difference of A A A to annihilate that vector, since Γ L ( A ) ⪯ a A + B \Gamma_{L}(A)\preceq aA+B Γ L ( A ) ⪯ a A + B . The same value would propagate through the irreducible support, contradicting E A = 0 \mathbb{E}A=0 E A = 0 . Hence λ min ( A ) > − v / a \lambda_{\min}(A)>-v/a λ m i n ( A ) > − v / a .
For the intrinsic-dimension bound, apply ψ ( u ) = e u − u − 1 \psi(u)=e^{u}-u-1 ψ ( u ) = e u − u − 1 and E A = 0 \mathbb{E}A=0 E A = 0 to obtain
E Tr ψ ( θ A ) ≤ Tr B v ( e θ 2 v / ( α − a θ ) − 1 ) . \mathbb{E}\operatorname{Tr}\psi(\theta A)\leq\frac{\operatorname{Tr}B}{v}\left(e^{\theta^{2}v/(\alpha-a\theta)}-1\right). E Tr ψ ( θ A ) ≤ v Tr B ( e θ 2 v / ( α − a θ ) − 1 ) .
At the positive parameter above, put E = α t 2 / ( 4 v + 2 a t ) E=\alpha t^{2}/(4v+2at) E = α t 2 / ( 4 v + 2 a t ) . Then θ t = 2 E \theta t=2E θ t = 2 E and θ 2 v / ( α − a θ ) = E \theta^{2}v/(\alpha-a\theta)=E θ 2 v / ( α − a θ ) = E . The same elementary estimate as in Corollary 11.2 gives the claimed prefactor. Finally, if B = 0 B=0 B = 0 , taking expectations of (11.11) gives E Γ L ( A ) ⪯ a E A = 0 \mathbb{E}\Gamma_{L}(A)\preceq a\mathbb{E}A=0 E Γ L ( A ) ⪯ a E A = 0 , so every edge difference vanishes and the centered observable is zero. □
Natural matroid basis laws and Lorentzian occupations.
Corollary 11.5 (All nonlinear Lorentzian observables). Let μ \mu μ be a degree-r r r Lorentzian occupation law, and let K 0 K_{0} K 0 be deletion of one particle followed by conditional replacement. Set L = I − K 0 L=I-K_{0} L = I − K 0 . For every Hermitian observable A A A satisfying Γ L ( A ) ⪯ V \Gamma_{L}(A)\preceq V Γ L ( A ) ⪯ V , every Hermitian offset K K K , and every real θ \theta θ ,
E μ Tr e K + θ A ≤ Tr exp ( K + θ E μ A + r θ 2 V ) . (93) \mathbb{E}_{\mu}\operatorname{Tr}e^{K+\theta A}\leq\operatorname{Tr}\exp(K+\theta\mathbb{E}_{\mu}A+r\theta^{2}V). \tag*{(93)} E μ Tr e K + θ A ≤ Tr exp ( K + θ E μ A + r θ 2 V ) . (93)
If every allowed one-particle replacement changes A A A in operator norm by at most a > 0 a>0 a > 0 , then
Pr { λ max ( A − E A ) ≥ t } ≤ d exp ( − t 2 2 r a 2 ) . (94) \Pr\{\lambda_{\max}(A-\mathbb{E}A)\geq t\}\leq d\exp\left(-\frac{t^{2}}{2ra^{2}}\right). \tag*{(94)} Pr { λ m a x ( A − E A ) ≥ t } ≤ d exp ( − 2 r a 2 t 2 ) . (94)
Both statements hold, in particular, for arbitrary Hermitian functions on the bases of every matroid at every positive product field. Under the affine assumptions of Theorem 11.3 , a centered observable with Γ L ( A ) ⪯ a A + B \Gamma_L(A) \preceq aA+B Γ L ( A ) ⪯ a A + B satisfies
E μ Tr e K + θ A ≤ Tr exp ( K + r θ 2 1 − r a θ B ) ( θ < 1 / ( r a ) ) . (95) \mathbb{E}_{\mu}\operatorname{Tr}e^{K+\theta A}\leq\operatorname{Tr}\exp\left(K+\frac{r\theta^2}{1-ra\theta}B\right)\qquad(\theta<1/(ra)). \tag*{(95)} E μ Tr e K + θ A ≤ Tr exp ( K + 1 − r a θ r θ 2 B ) ( θ < 1/ ( r a )) . (95)
For a = 0 a=0 a = 0 , every real θ \theta θ is allowed. Under (3.3), replace r r r in these conclusions by [ r + ( r − 1 ) κ ] / ( 1 + κ ) [r+(r-1)\kappa]/(1+\kappa) [ r + ( r − 1 ) κ ] / ( 1 + κ ) .
Proof. Corollary 2.2 gives (11.1) with α = 1 / r \alpha=1/r α = 1/ r . Theorem 3.4 gives the improved constant. The replacement Lipschitz assumption implies Γ L ( A ) ⪯ a 2 I / 2 \Gamma_L(A)\preceq a^2I/2 Γ L ( A ) ⪯ a 2 I /2 , since K 0 K_0 K 0 is stochastic. Apply Theorem 11.1 and Corollary 11.2 . The affine comparison follows from Theorem 11.3 . □
This includes the nonlinear observable class discussed in Kathuria [29 Sections 4–5] : every Hermitian function with bounded exchange differences is allowed. The bound is subgaussian throughout its range and applies to arbitrary matroids. Exponential matrix concentration for Lorentzian measures was already obtained from the scalar-to-matrix Poincaré principle of Garg–Kathuria–Srivastava [23 ] .
Corollary 11.6 (Rank and corank in each component). Let μ \mu μ be a positive product-field basis law of a matroid M M M . For each connected component C C C , put s C = min ( rk M ( C ) , ∣ C ∣ − rk M ( C ) ) s_C=\min(\operatorname{rk}_M(C),|C|-\operatorname{rk}_M(C)) s C = min ( rk M ( C ) , ∣ C ∣ − rk M ( C )) . Suppose that every basis exchange inside C C C changes a Hermitian observable A A A in operator norm by at most a C a_C a C . Then
Pr { λ max ( A − E A ) ≥ t } ≤ d exp ( − t 2 2 ∑ C s C a C 2 ) (96) \Pr\{\lambda_{\max}(A-\mathbb{E}A)\geq t\}\leq d\exp\left(-\frac{t^2}{2\sum_Cs_Ca_C^2}\right) \tag*{(96)} Pr { λ m a x ( A − E A ) ≥ t } ≤ d exp ( − 2 ∑ C s C a C 2 t 2 ) (96) when the denominator is positive. If it is zero, A A A is constant on the reference support.
Proof. In each component of rank larger than its corank, replace the basis by its complement and the field by its reciprocal. This pushes μ \mu μ to a product-field basis law of a direct sum of matroids of total rank s = ∑ C s C s=\sum_Cs_C s = ∑ C s C . If s = 0 s=0 s = 0 , the support is a singleton. Otherwise components of rank zero may be removed. For its deletion–replacement generator, a step chooses component C C C with probability s C / s s_C/s s C / s and exchanges only within that component. Thus Γ L ( A ) ⪯ ( 2 s ) − 1 ∑ C s C a C 2 I \Gamma_L(A)\preceq(2s)^{-1}\sum_Cs_Ca_C^2I Γ L ( A ) ⪯ ( 2 s ) − 1 ∑ C s C a C 2 I . The complete modified logarithmic Sobolev constant is at least 1 / s 1/s 1/ s . Apply (11.8) and pull the observable back through the complementation bijection. □
Matrix concentration for arbitrary-order subconfiguration averages
A subconfiguration average inherits the kernel’s matrix Laplace bound, with reciprocal deletion coefficient playing the role of sample size.
Let α \alpha α have a degree-r r r Lorentzian occupation law, and fix 1 ≤ ℓ ≤ r 1\leq\ell\leq r 1 ≤ ℓ ≤ r . For a Hermitian matrix kernel B β B_\beta B β on the retained ℓ \ell ℓ -occupations, define
U ( α ) = ∑ β D r → ℓ ( β ∣ α ) B β , B ‾ = E μ ℓ B β = E μ U . (97) U(\alpha)=\sum_{\beta}D_{r\to\ell}(\beta\mid\alpha)B_\beta,\qquad\overline{B}=\mathbb{E}_{\mu_\ell}B_\beta=\mathbb{E}_{\mu}U. \tag*{(97)} U ( α ) = β ∑ D r → ℓ ( β ∣ α ) B β , B = E μ ℓ B β = E μ U . (97)
For a matroid basis B B B this is
U ( B ) = ( r ℓ ) − 1 ∑ I ⊆ B ∣ I ∣ = ℓ B I . U(B)=\binom{r}{\ell}^{-1}\sum_{\substack{I\subseteq B\\ |I|=\ell}}B_I. U ( B ) = ( ℓ r ) − 1 I ⊆ B ∣ I ∣ = ℓ ∑ B I .
Thus the kernel may depend on an entire ℓ \ell ℓ -subset. Throughout this section one may take c = r / ℓ c=r/\ell c = r / ℓ , or the larger value c = a r / a ℓ c=a_r/a_\ell c = a r / a ℓ under the hypothesis of Theorem 3.4 .
Theorem 12.1 (Matrix Laplace comparison). For every Hermitian K K K and every real θ \theta θ ,
E Tr e K + θ U ≤ Tr exp ( K + c log E μ ℓ e θ B β / c ) . (98) \mathbb{E}\operatorname{Tr}e^{K+\theta U}\leq\operatorname{Tr}\exp\left(K+c\log\mathbb{E}_{\mu_\ell}e^{\theta B_\beta/c}\right). \tag*{(98)} E Tr e K + θ U ≤ Tr exp ( K + c log E μ ℓ e θ B β / c ) . (98)
In particular, with M θ = E μ ℓ e θ B β / c M_\theta= \mathbb{E}_{\mu_\ell}e^{\theta B_\beta/c} M θ = E μ ℓ e θ B β / c ,
E Tr e θ U ≤ Tr M θ c ≤ d ∥ M θ ∥ c . \mathbb{E}\operatorname{Tr}e^{\theta U} \le\operatorname{Tr}M_\theta^c \le d\lVert M_\theta\rVert^c. E Tr e θ U ≤ Tr M θ c ≤ d ∥ M θ ∥ c .
Proof. In Corollary 4.2, put A β = e θ B β / c A_\beta= e^{\theta B_\beta/c} A β = e θ B β / c . Then c D ∗ log A = θ U cD^* \log A = \theta U c D ∗ log A = θ U . The last comparisons follow from spectral calculus; the real exponent c c c need not be an integer. □
Theorem 12.2 (Chernoff and binomial bounds). Suppose 0 ⪯ B β ⪯ R I d 0 \preceq B_\beta\preceq R I_d 0 ⪯ B β ⪯ R I d , where R > 0 R > 0 R > 0 , and put m + = λ max ( B ‾ ) m_+ = \lambda_{\max}(\overline{B}) m + = λ m a x ( B ) and m − = λ min ( B ‾ ) m_- = \lambda_{\min}(\overline{B}) m − = λ m i n ( B ) . For δ ≥ 0 \delta\ge0 δ ≥ 0 and m + > 0 m_+ > 0 m + > 0 ,
Pr { λ max ( U ) ≥ ( 1 + δ ) m + } ≤ d ( e δ ( 1 + δ ) 1 + δ ) c m + / R . (99) \Pr\{\lambda_{\max}(U) \ge(1+\delta)m_+\} \le d\left(\frac{e^\delta}{(1+\delta)^{1+\delta}}\right)^{cm_+/R}. \tag*{(99)} Pr { λ m a x ( U ) ≥ ( 1 + δ ) m + } ≤ d ( ( 1 + δ ) 1 + δ e δ ) c m + / R . (99) For 0 < δ < 1 0 < \delta< 1 0 < δ < 1 and m − > 0 m_- > 0 m − > 0 ,
Pr { λ min ( U ) ≤ ( 1 − δ ) m − } ≤ d ( e − δ ( 1 − δ ) 1 − δ ) c m − / R . (100) \Pr\{\lambda_{\min}(U) \le(1-\delta)m_-\} \le d\left(\frac{e^{-\delta}}{(1-\delta)^{1-\delta}}\right)^{cm_-/R}. \tag*{(100)} Pr { λ m i n ( U ) ≤ ( 1 − δ ) m − } ≤ d ( ( 1 − δ ) 1 − δ e − δ ) c m − / R . (100) For s + = m + / R < u ≤ 1 s_+ = m_+/R < u \le1 s + = m + / R < u ≤ 1 and 0 ≤ u < s − = m − / R 0 \le u < s_- = m_-/R 0 ≤ u < s − = m − / R , the respective refinements are
Pr { λ max ( U ) ≥ R u } ≤ d e − c kl ( u ∥ s + ) , Pr { λ min ( U ) ≤ R u } ≤ d e − c kl ( u ∥ s − ) , \Pr\{\lambda_{\max}(U) \ge Ru\} \le de^{-c\operatorname{kl}(u\mathbin{\|}s_+)}, \qquad\Pr\{\lambda_{\min}(U) \le Ru\} \le de^{-c\operatorname{kl}(u\mathbin{\|}s_-)}, Pr { λ m a x ( U ) ≥ R u } ≤ d e − c kl ( u ∥ s + ) , Pr { λ m i n ( U ) ≤ R u } ≤ d e − c kl ( u ∥ s − ) , where kl ( u ∥ s ) = u log ( u / s ) + ( 1 − u ) log ( ( 1 − u ) / ( 1 − s ) ) \operatorname{kl}(u\mathbin{\|}s) = u\log(u/s) + (1-u)\log((1-u)/(1-s)) kl ( u ∥ s ) = u log ( u / s ) + ( 1 − u ) log (( 1 − u ) / ( 1 − s )) . Endpoint cases follow by limits, and all bounds may be capped at one.
Proof. For every real t t t , spectral calculus gives
e t B β / R ⪯ I + ( e t − 1 ) B β / R . e^{tB_\beta/R} \preceq I + (e^t - 1)B_\beta/R. e t B β / R ⪯ I + ( e t − 1 ) B β / R .
If t > 0 t > 0 t > 0 , the largest eigenvalue of the expectation on the right is at most 1 + s + ( e t − 1 ) 1 + s_+(e^t - 1) 1 + s + ( e t − 1 ) . Theorem 12.1 and the trace exponential Markov bound, with θ = c t / R \theta= ct/R θ = c t / R , give
Pr { λ max ( U ) ≥ R u } ≤ d inf t > 0 exp { c [ − t u + log ( 1 − s + + s + e t ) ] } . \Pr\{\lambda_{\max}(U) \ge Ru\} \le d\inf_{t>0}\exp\{c[-tu+\log(1-s_+ + s_+e^t)]\}. Pr { λ m a x ( U ) ≥ R u } ≤ d t > 0 inf exp { c [ − t u + log ( 1 − s + + s + e t )]} .
Optimizing gives the binary relative-entropy bound. For t < 0 t < 0 t < 0 , the largest eigenvalue is controlled by s − s_- s − , and the same argument with λ min ( U ) \lambda_{\min}(U) λ m i n ( U ) gives the lower bound. Finally use log ( 1 + s ( e t − 1 ) ) ≤ s ( e t − 1 ) \log(1+s(e^t-1)) \le s(e^t-1) log ( 1 + s ( e t − 1 )) ≤ s ( e t − 1 ) and take t = log ( 1 + δ ) t=\log(1+\delta) t = log ( 1 + δ ) or t = log ( 1 − δ ) t=\log(1-\delta) t = log ( 1 − δ ) to obtain the displayed Chernoff estimates. □
Theorem 12.3 (Centered Bennett and Bernstein bounds). Put C β = B β − B ‾ C_\beta= B_\beta- \overline{B} C β = B β − B and suppose ∥ C β ∥ ≤ R \lVert C_\beta\rVert\le R ∥ C β ∥ ≤ R on the support, where R > 0 R > 0 R > 0 . Define
V = E μ ℓ C β 2 , σ 2 = ∥ V ∥ , X = U − B ‾ . V = \mathbb{E}_{\mu_\ell}C_\beta^2,\qquad\sigma^2 = \lVert V\rVert,\qquad X = U-\overline{B}. V = E μ ℓ C β 2 , σ 2 = ∥ V ∥ , X = U − B . If σ 2 > 0 \sigma^2 > 0 σ 2 > 0 , then for every t > 0 t > 0 t > 0 ,
Pr { λ max ( X ) ≥ t } ≤ d exp [ − c σ 2 R 2 h ( R t σ 2 ) ] (101) \Pr\{\lambda_{\max}(X) \ge t\} \le d\exp\left[-\frac{c\sigma^2}{R^2}h\left(\frac{Rt}{\sigma^2}\right)\right] \tag*{(101)} Pr { λ m a x ( X ) ≥ t } ≤ d exp [ − R 2 c σ 2 h ( σ 2 R t ) ] (101) ≤ d exp [ − c t 2 2 σ 2 + 2 R t / 3 ] , (102) \le d\exp\left[-\frac{ct^2}{2\sigma^2+2Rt/3}\right], \tag*{(102)} ≤ d exp [ − 2 σ 2 + 2 R t /3 c t 2 ] , (102) where h ( v ) = ( 1 + v ) log ( 1 + v ) − v h(v)=(1+v)\log(1+v)-v h ( v ) = ( 1 + v ) log ( 1 + v ) − v . The same bounds hold for λ max ( − X ) \lambda_{\max}(-X) λ m a x ( − X ) , and a union bound gives a two-sided norm bound with prefactor 2 d 2d 2 d . If V = 0 V=0 V = 0 , then X = 0 X=0 X = 0 almost surely.
Proof. For s ≥ 0 s \ge0 s ≥ 0 and − R ≤ x ≤ R -R \le x \le R − R ≤ x ≤ R ,
e s x ≤ 1 + s x + e s R − 1 − s R R 2 x 2 . e^{sx} \le1+sx+\frac{e^{sR}-1-sR}{R^2}x^2. e s x ≤ 1 + s x + R 2 e s R − 1 − s R x 2 .
Apply this by spectral calculus and use E C β = 0 \mathbb{E}C_\beta=0 E C β = 0 . Since log \log log is operator monotone and log ( I + A ) ⪯ A \log(I+A)\preceq A log ( I + A ) ⪯ A for A ⪰ 0 A\succeq0 A ⪰ 0 ,
c log E e θ C β / c ⪯ c R 2 ( e θ R / c − 1 − θ R / c ) V . c\log\mathbb{E}e^{\theta C_\beta/c}\preceq\frac{c}{R^2}(e^{\theta R/c}-1-\theta R/c)V. c log E e θ C β / c ⪯ R 2 c ( e θ R / c − 1 − θ R / c ) V .
The Laplace comparison applies to the centered kernel itself, so
E Tr e θ X ≤ Tr exp [ c R 2 ( e θ R / c − 1 − θ R / c ) V ] . \mathbb{E}\operatorname{Tr}e^{\theta X}\leq\operatorname{Tr}\exp\left[\frac{c}{R^2}\left(e^{\theta R/c}-1-\theta R/c\right)V\right]. E Tr e θ X ≤ Tr exp [ R 2 c ( e θ R / c − 1 − θ R / c ) V ] .
The exponential trace is monotone on Hermitian matrices. Bounding the last expression by its largest eigenvalue and optimizing at θ = ( c / R ) log ( 1 + R t / σ 2 ) \theta=(c/R)\log(1+Rt/\sigma^2) θ = ( c / R ) log ( 1 + R t / σ 2 ) proves (101) . The scalar inequality h ( v ) ≥ v 2 / ( 2 + 2 v / 3 ) h(v)\geq v^2/(2+2v/3) h ( v ) ≥ v 2 / ( 2 + 2 v /3 ) gives (102) . If V = 0 V=0 V = 0 , positivity gives C β = 0 C_\beta=0 C β = 0 on the reference support.
Corollary 12.4 (Intrinsic dimension). Under Theorem 12.3 , suppose σ 2 > 0 \sigma^2>0 σ 2 > 0 and put d e f f = Tr V / ∥ V ∥ d_{\mathrm{eff}}=\operatorname{Tr}V/\lVert V\rVert d eff = Tr V / ∥ V ∥ . For t > 0 t>0 t > 0 ,
Pr { λ max ( X ) ≥ t } ≤ min { 1 , inf θ > 0 Tr ( e g c ( θ ) V − I ) e θ t − θ t − 1 } \Pr\{\lambda_{\max}(X)\geq t\}\leq\min\left\{1,\inf_{\theta>0}\frac{\operatorname{Tr}(e^{g_c(\theta)V}-I)}{e^{\theta t}-\theta t-1}\right\} Pr { λ m a x ( X ) ≥ t } ≤ min { 1 , θ > 0 inf e θ t − θ t − 1 Tr ( e g c ( θ ) V − I ) } ≤ e d e f f exp [ − c t 2 2 σ 2 + 2 R t / 3 ] , g c ( θ ) = c R 2 ( e θ R / c − 1 − θ R / c ) . \leq e\,d_{\mathrm{eff}}\exp\left[-\frac{ct^2}{2\sigma^2+2Rt/3}\right],\qquad g_c(\theta)=\frac{c}{R^2}\left(e^{\theta R/c}-1-\theta R/c\right). ≤ e d eff exp [ − 2 σ 2 + 2 R t /3 c t 2 ] , g c ( θ ) = R 2 c ( e θ R / c − 1 − θ R / c ) . The same upper bound holds for the opposite tail, and the norm bound has prefactor 2 e d e f f 2e\,d_{\mathrm{eff}} 2 e d eff .
Proof. Use the intrinsic-dimension argument of Corollary 11.2 , replacing θ 2 V / α \theta^2V/\alpha θ 2 V / α by g c ( θ ) V g_c(\theta)V g c ( θ ) V in its spectral test. The Laplace bound in Theorem 12.3 gives the first inequality and bounds its numerator by d e f f ( e g c ( θ ) σ 2 − 1 ) d_{\mathrm{eff}}(e^{g_c(\theta)\sigma^2}-1) d eff ( e g c ( θ ) σ 2 − 1 ) . Set
E = c t 2 2 σ 2 + 2 R t / 3 , θ = c t σ 2 + R t / 3 . E=\frac{ct^2}{2\sigma^2+2Rt/3},\qquad\theta=\frac{ct}{\sigma^2+Rt/3}. E = 2 σ 2 + 2 R t /3 c t 2 , θ = σ 2 + R t /3 c t .
Then θ t = 2 E \theta t=2E θ t = 2 E , θ R / c < 3 \theta R/c<3 θ R / c < 3 , and e u − 1 − u ≤ u 2 / [ 2 ( 1 − u / 3 ) ] e^u-1-u\leq u^2/[2(1-u/3)] e u − 1 − u ≤ u 2 / [ 2 ( 1 − u /3 )] gives g c ( θ ) σ 2 ≤ E g_c(\theta)\sigma^2\leq E g c ( θ ) σ 2 ≤ E . The same scalar ratio bound and small-exponent case from that corollary give e d e f f e − E e\,d_{\mathrm{eff}}e^{-E} e d eff e − E . Apply this to − X -X − X for the opposite tail.
Corollary 12.5 (Matrix sums on every matroid). Let B B B be a product-field random basis of any rank-r r r matroid, r ≥ 1 r\geq1 r ≥ 1 , and put Y = ∑ i ∈ B A i Y=\sum_{i\in B}A_i Y = ∑ i ∈ B A i , Q = E Y Q=\mathbb{E}Y Q = E Y , for d × d d\times d d × d matrices. If 0 ⪯ A i ⪯ R I 0\preceq A_i\preceq RI 0 ⪯ A i ⪯ R I with R > 0 R>0 R > 0 , then, for δ ≥ 0 \delta\geq0 δ ≥ 0 and q + = λ max ( Q ) > 0 q_+=\lambda_{\max}(Q)>0 q + = λ m a x ( Q ) > 0 ,
Pr { λ max ( Y ) ≥ ( 1 + δ ) q + } ≤ d ( e δ ( 1 + δ ) 1 + δ ) q + / R . \Pr\{\lambda_{\max}(Y)\geq(1+\delta)q_+\}\leq d\left(\frac{e^\delta}{(1+\delta)^{1+\delta}}\right)^{q_+/R}. Pr { λ m a x ( Y ) ≥ ( 1 + δ ) q + } ≤ d ( ( 1 + δ ) 1 + δ e δ ) q + / R . For 0 < δ < 1 0<\delta<1 0 < δ < 1 and q − = λ min ( Q ) > 0 q_-=\lambda_{\min}(Q)>0 q − = λ m i n ( Q ) > 0 ,
Pr { λ min ( Y ) ≤ ( 1 − δ ) q − } ≤ d ( e − δ ( 1 − δ ) 1 − δ ) q − / R . \Pr\{\lambda_{\min}(Y)\leq(1-\delta)q_-\}\leq d\left(\frac{e^{-\delta}}{(1-\delta)^{1-\delta}}\right)^{q_-/R}. Pr { λ m i n ( Y ) ≤ ( 1 − δ ) q − } ≤ d ( ( 1 − δ ) 1 − δ e − δ ) q − / R . For arbitrary Hermitian A i A_i A i , write p i = Pr ( i ∈ B ) p_i=\Pr(i\in B) p i = Pr ( i ∈ B ) and
V 0 = ∑ i p i A i 2 − Q 2 r , R 0 = max i : p i > 0 ∥ A i − Q / r ∥ . V_0=\sum_i p_iA_i^2-\frac{Q^2}{r},\qquad R_0=\max_{i:p_i>0}\lVert A_i-Q/r\rVert. V 0 = i ∑ p i A i 2 − r Q 2 , R 0 = i : p i > 0 max ∥ A i − Q / r ∥ . If ∥ V 0 ∥ > 0 \lVert V_0\rVert>0 ∥ V 0 ∥ > 0 , then for every t > 0 t>0 t > 0 ,
Pr { λ max ( Y − Q ) ≥ t } ≤ d exp ( − t 2 2 ∥ V 0 ∥ + 2 R 0 t / 3 ) . \Pr\{\lambda_{\max}(Y-Q)\geq t\}\leq d\exp\left(-\frac{t^2}{2\lVert V_0\rVert+2R_0t/3}\right). Pr { λ m a x ( Y − Q ) ≥ t } ≤ d exp ( − 2 ∥ V 0 ∥ + 2 R 0 t /3 t 2 ) . The same bound holds for the opposite tail.
Proof. Use the one-particle kernel B i = r A i B_i=rA_i B i = r A i in the preceding theorems. Then U = Y U=Y U = Y , c = r c=r c = r , its bound is r R rR r R , and its centered variance and radius are r V 0 rV_0 r V 0 and r R 0 rR_0 r R 0 . The rank cancels in each exponent. Positivity of V 0 V_0 V 0 follows from operator Jensen for the probabilities p i / r p_i/r p i / r .
Corollary 12.6 (Spectral approximation from arbitrary matroid bases). Let B B B be a product-field random basis of an arbitrary matroid, let P i ⪰ 0 P_i \succeq0 P i ⪰ 0 , and suppose Q = E ∑ i ∈ B P i Q = \mathbb{E}\sum_{i\in B} P_i Q = E ∑ i ∈ B P i has rank d 0 ≥ 1 d_0 \ge1 d 0 ≥ 1 . Interpret Q − 1 / 2 Q^{-1/2} Q − 1/2 on its range. If ∥ Q − 1 / 2 P i Q − 1 / 2 ∥ ≤ ρ \lVert Q^{-1/2}P_iQ^{-1/2}\rVert\le\rho ∥ Q − 1/2 P i Q − 1/2 ∥ ≤ ρ on supported elements, then for independent bases B 1 , … , B m B_1,\ldots,B_m B 1 , … , B m and 0 < ε < 1 0 < \varepsilon< 1 0 < ε < 1 ,
Pr { 1 m ∑ j = 1 m ∑ i ∈ B j P i ∉ [ ( 1 − ε ) Q , ( 1 + ε ) Q ] } ≤ 2 d 0 exp ( − m ε 2 3 ρ ) . \Pr\left\{\frac{1}{m}\sum_{j=1}^{m}\sum_{i\in B_j}P_i \notin[(1-\varepsilon)Q,(1+\varepsilon)Q]\right\}\le2d_0\exp\left(-\frac{m\varepsilon^2}{3\rho}\right). Pr ⎩ ⎨ ⎧ m 1 j = 1 ∑ m i ∈ B j ∑ P i ∈ / [( 1 − ε ) Q , ( 1 + ε ) Q ] ⎭ ⎬ ⎫ ≤ 2 d 0 exp ( − 3 ρ m ε 2 ) . Here interval membership is in the positive semidefinite order. If Q = 0 Q=0 Q = 0 , every supported summand vanishes and the approximation is exact.
Proof. Write r r r for the rank of the matroid. The joint basis law is the law of a direct sum of m m m matroids, of total degree m r mr m r . Give its one-particle label ( j , i ) (j,i) ( j , i ) the kernel B ( j , i ) = r Q − 1 / 2 P i Q − 1 / 2 B_{(j,i)}=rQ^{-1/2}P_iQ^{-1/2} B ( j , i ) = r Q − 1/2 P i Q − 1/2 . Its normalized average is the whitened matrix in the event, its mean is I I I , its kernel bound is R = r ρ R=r\rho R = r ρ , and c = m r c=mr c = m r . Thus c / R = m / ρ c/R=m/\rho c / R = m / ρ in Theorem 12.2. The standard scalar estimates for the two Chernoff functions give the claim. □
Theorems 12.1–12.3 apply to every matroid, without representability assumptions, and to all homogeneous Lorentzian occupation laws. The effective sample size for a kernel on ℓ \ell ℓ particles is r / ℓ r/\ell r / ℓ . The quadratic gap gives the larger value a r / a ℓ a_r/a_\ell a r / a ℓ for simple matroids and any other law satisfying (3.3).
Exact marginal and matrix Hölder regions
Selected-particle entropy controls the event that a coordinate is present. Applying it also to the complement controls the event that it is absent. Together they give the half-weight inequality for full Bernoulli coordinates. Component factorization then gives the whole admissible region, and a field concentrating on one basis exchange proves every necessary pair constraint. Section 14 then determines fixed-field sharpness and every equality density for the half vector.
Let M M M be a matroid on [ n ] [n] [ n ] , of rank r r r , and let
f ( z ) = ∑ B ∈ B ( M ) c B z B , c B > 0. f(z)=\sum_{B\in\mathcal{B}(M)}c_Bz^B,\qquad c_B>0. f ( z ) = B ∈ B ( M ) ∑ c B z B , c B > 0.
Assume that both f f f and its complement polynomial f ⋆ ( z ) = z [ n ] f ( z 1 − 1 , … , z n − 1 ) f^\star(z)=z^{[n]}f(z_1^{-1},\ldots,z_n^{-1}) f ⋆ ( z ) = z [ n ] f ( z 1 − 1 , … , z n − 1 ) are Lorentzian. For a positive field put
μ λ ( B ) = c B λ B f ( λ ) , λ i > 0. \mu_\lambda(B)=\frac{c_B\lambda^B}{f(\lambda)},\qquad\lambda_i>0. μ λ ( B ) = f ( λ ) c B λ B , λ i > 0.
This includes product-field basis measures of every matroid and all homogeneous strongly Rayleigh laws: a multiaffine stable polynomial and its complement are stable, hence Lorentzian. The coefficient weights are otherwise arbitrary subject to the two stated hypotheses. Write X i = 1 i ∈ B X_i=1_{i\in B} X i = 1 i ∈ B and let μ λ , i \mu_{\lambda,i} μ λ , i be its Bernoulli reference law.
Proposition 13.1 (Full Bernoulli marginals). For every positive semidefinite matrix function F F F on B ( M ) \mathcal{B}(M) B ( M ) ,
1 2 ∑ i = 1 n H μ λ , i ( E [ F ∣ X i ] ) ≤ H μ λ ( F ) . (103) \frac{1}{2}\sum_{i=1}^{n}\mathcal{H}_{\mu_{\lambda,i}}\bigl(\mathbb{E}[F\mid X_i]\bigr)\le\mathcal{H}_{\mu_\lambda}(F). \tag*{(103)} 2 1 i = 1 ∑ n H μ λ , i ( E [ F ∣ X i ] ) ≤ H μ λ ( F ) . (103) Proof. If r = 0 r=0 r = 0 or r = n r=n r = n , the law is a point mass and the assertion is immediate. Otherwise put p i = Pr ( X i = 1 ) p_i=\Pr(X_i=1) p i = Pr ( X i = 1 ) , G i = E [ F ∣ X i = 1 ] G_i=\mathbb{E}[F\mid X_i=1] G i = E [ F ∣ X i = 1 ] , K i = E [ F ∣ X i = 0 ] K_i=\mathbb{E}[F\mid X_i=0] K i = E [ F ∣ X i = 0 ] , and σ = E F \sigma=\mathbb{E}F σ = E F ; omit missing conditioning events. Applying Theorem 2.1 with ℓ = 1 \ell=1 ℓ = 1 and multiplying by r r r gives
∑ i p i Tr [ G i log G i ] − r Tr [ σ log σ ] ≤ H μ ( F ) . \sum_i p_i\operatorname{Tr}[G_i\log G_i]-r\operatorname{Tr}[\sigma\log\sigma]\le\mathcal{H}_\mu(F). i ∑ p i Tr [ G i log G i ] − r Tr [ σ log σ ] ≤ H μ ( F ) .
Complementary bases have the law of f ∗ f^* f ∗ at field λ − 1 \lambda^{-1} λ − 1 . The same argument therefore gives
∑ i ( 1 − p i ) Tr [ K i log K i ] − ( n − r ) Tr [ σ log σ ] ≤ H μ ( F ) . \sum_i (1-p_i)\operatorname{Tr}[K_i\log K_i]-(n-r)\operatorname{Tr}[\sigma\log\sigma]\leq\mathcal{H}_{\mu}(F). i ∑ ( 1 − p i ) Tr [ K i log K i ] − ( n − r ) Tr [ σ log σ ] ≤ H μ ( F ) .
Their sum is the left side of (103) before multiplication by 1 / 2 1/2 1/2 . Constant coordinates cause no problem in these identities. □
Lemma 13.2 (Factorization over matroid components). The coefficient law of a multiaffine Lorentzian polynomial factors over the connected components of its support matroid, at every positive field.
Proof. Choose adjacent bases in each of two distinct components and fix bases in the other components. Differentiate by the common elements and set all remaining coordinates except the four exchanged elements to zero. The resulting quadratic is Lorentzian and has the form
∑ i , j = 1 2 c i j x i y j , \sum_{i,j=1}^{2}c_{ij}x_i y_j, i , j = 1 ∑ 2 c ij x i y j ,
with c i j > 0 c_{ij}>0 c ij > 0 . Its Hessian has off-diagonal blocks C C C and C T C^{\mathsf T} C T ; its number of positive eigenvalues is rank C \operatorname{rank} C rank C . Hence c 11 c 22 = c 12 c 21 c_{11}c_{22}=c_{12}c_{21} c 11 c 22 = c 12 c 21 . These rectangle identities say that the coefficient ratio along a basis exchange in one component is unchanged by an exchange in another. Each component’s basis graph is connected. Fixing reference bases and following exchange paths therefore factors the coefficients, and positive fields preserve the factorization. Each factor inherits Lorentzianity by differentiating in fixed bases of the other components. When f ∗ f^* f ∗ is also Lorentzian, the same argument gives Lorentzianity of the complementary factors. □
Let C 1 , … , C m C_1,\ldots,C_m C 1 , … , C m be the nontrivial connected components of M M M ; loops and coloops are constant coordinates and may be omitted. Define
W ( M ) = { w ∈ R ≥ 0 n : w i + w j ≤ 1 whenever i ≠ j lie in the same C a } . \mathcal{W}(M)=\{w\in\mathbb{R}_{\geq0}^{n}:w_i+w_j\leq1\ \text{whenever }i\neq j\text{ lie in the same }C_a\}. W ( M ) = { w ∈ R ≥ 0 n : w i + w j ≤ 1 whenever i = j lie in the same C a } .
Theorem 13.3 (Sharp all-field quantum coefficient region). Assume f f f and f ∗ f^* f ∗ are Lorentzian as above. For w ≥ 0 w\geq0 w ≥ 0 , the following statements are equivalent.
(1) w ∈ W ( M ) w\in\mathcal{W}(M) w ∈ W ( M ) .
(2) For every positive external field, every finite quantum register Q Q Q , and every classical–quantum state supported on B ( M ) ⊗ Q \mathcal{B}(M)\otimes Q B ( M ) ⊗ Q ,
∑ i w i D ( ρ X i Q ∥ μ λ , i ⊗ ρ Q ) ≤ D ( ρ B Q ∥ μ λ ⊗ ρ Q ) . (104) \sum_i w_iD(\rho_{X_iQ}\Vert\mu_{\lambda,i}\otimes\rho_Q)\leq D(\rho_{BQ}\Vert\mu_\lambda\otimes\rho_Q). \tag*{(104)} i ∑ w i D ( ρ X i Q ∥ μ λ , i ⊗ ρ Q ) ≤ D ( ρ B Q ∥ μ λ ⊗ ρ Q ) . (104) (3) Statement (2) holds with a one-dimensional quantum register.
The same equivalence applies to every nonzero minor law, with the connected components of its support matroid.
Proof. First suppose M M M is connected with at least two elements. The half vector satisfies (104) by Proposition 13.1, and every coordinate unit vector does by data processing. If h = max i w i ≤ 1 / 2 h=\max_i w_i\leq1/2 h = max i w i ≤ 1/2 , the weight vector is dominated coordinatewise by the half vector. Otherwise choose j j j with w j = h > 1 / 2 w_j=h>1/2 w j = h > 1/2 . All other weights are at most 1 − h 1-h 1 − h , and h ≤ 1 h\leq1 h ≤ 1 . Therefore
w ≤ ( 2 h − 1 ) e j + ( 2 − 2 h ) 1 2 1. w\leq(2h-1)e_j+(2-2h)\frac{1}{2}\mathbf{1}. w ≤ ( 2 h − 1 ) e j + ( 2 − 2 h ) 2 1 1 .
This is a convex combination of two admissible coefficient vectors. Nonnegativity of relative entropy proves the claim by domination.
For several components, Lemma 13.2 makes μ λ \mu_\lambda μ λ a product of its component laws. If B a = B ∩ C a B_a=B\cap C_a B a = B ∩ C a , the complete product inequality
∑ a D ( ρ B a Q ∥ μ λ , a ⊗ ρ Q ) ≤ D ( ρ B Q ∥ μ λ ⊗ ρ Q ) (105) \sum_a D(\rho_{B_aQ}\Vert\mu_{\lambda,a}\otimes\rho_Q)\leq D(\rho_{BQ}\Vert\mu_\lambda\otimes\rho_Q) \tag*{(105)} a ∑ D ( ρ B a Q ∥ μ λ , a ⊗ ρ Q ) ≤ D ( ρ B Q ∥ μ λ ⊗ ρ Q ) (105)
follows from strong subadditivity [35 ] . Indeed, the difference of the right and left sides is ∑ a H ( B a ∣ Q ) − H ( B 1 ⋯ B m ∣ Q ) ≥ 0 \sum_a H(B_a\mid Q)-H(B_1\cdots B_m\mid Q)\geq0 ∑ a H ( B a ∣ Q ) − H ( B 1 ⋯ B m ∣ Q ) ≥ 0 . Apply the connected-component inequality to each B a B_a B a and use (105) . This proves (1)⇒ \Rightarrow ⇒ (2), and (2)⇒ \Rightarrow ⇒ (3) is immediate.
For necessity, two distinct elements i , j i,j i , j belong to the same connected component precisely when some circuit contains both. Extending that circuit minus j j j to a basis gives a basis $B such that B ′ = B ∖ { i } ∪ { j } B' = B \setminus\{i\} \cup\{j\} B ′ = B ∖ { i } ∪ { j } is also a basis. The segment joining their indicator vectors is an edge of the matroid base polytope. Expose this edge by a linear functional q q q and take λ k ( t ) = exp ( t q k ) \lambda_k(t) = \exp(tq_k) λ k ( t ) = exp ( t q k ) . The laws converge to a two-point law on { B , B ′ } \{B, B'\} { B , B ′ } assigning mass π = c B / ( c B + c B ′ ) ∈ ( 0 , 1 ) \pi= c_B/(c_B+c_{B'}) \in(0,1) π = c B / ( c B + c B ′ ) ∈ ( 0 , 1 ) to B B B . Test with ν = δ B \nu= \delta_B ν = δ B . The joint divergence and the two changing coordinate divergences tend to − log π > 0 -\log\pi> 0 − log π > 0 ; all other coordinate divergences tend to zero. Hence w i + w j ≤ 1 w_i+w_j \le1 w i + w j ≤ 1 . Differentiation and setting variables to zero preserve both hypotheses, since complementation interchanges these operations. The same proof therefore applies to every nonzero minor. □ \square □
Corollary 13.5 (The exact matrix Hölder region). Under the generating-and-complement hypotheses of Theorem 13.3 , fix w i ≥ 0 w_i \ge0 w i ≥ 0 . For positive definite A i , 0 , A i , 1 A_{i,0}, A_{i,1} A i , 0 , A i , 1 put M i = E μ i A i , X i M_i = \mathbb{E}_{\mu_i}A_{i,X_i} M i = E μ i A i , X i . The inequality
E μ Tr exp ( K + ∑ i w i log A i , X i ) ≤ Tr exp ( K + ∑ i w i log M i ) \mathbb{E}_\mu\operatorname{Tr}\exp\left(K+\sum_i w_i\log A_{i,X_i}\right) \le\operatorname{Tr}\exp\left(K+\sum_i w_i\log M_i\right) E μ Tr exp ( K + i ∑ w i log A i , X i ) ≤ Tr exp ( K + i ∑ w i log M i ) holds for all such matrices, in every matrix dimension, for every Hermitian K K K and every positive field, if and only if w ∈ W ( M ) w \in\mathcal{W}(M) w ∈ W ( M ) .
Proof. In the Gibbs variational formula, let F F F have normalized trace mean N N N and put J i = E [ F ∣ X i ] J_i = \mathbb{E}[F \mid X_i] J i = E [ F ∣ X i ] . The weighted marginal entropy bound gives H μ ( F ) ≥ ∑ i w i H μ i ( J i ) \mathcal{H}_\mu(F) \ge\sum_i w_i\mathcal{H}_{\mu_i}(J_i) H μ ( F ) ≥ ∑ i w i H μ i ( J i ) . Joint convexity gives
E μ i Tr [ J i ( log J i − log A i ) ] ≥ Tr [ N ( log N − log M i ) ] . \mathbb{E}_{\mu_i}\operatorname{Tr}[J_i(\log J_i-\log A_i)] \ge\operatorname{Tr}[N(\log N-\log M_i)]. E μ i Tr [ J i ( log J i − log A i )] ≥ Tr [ N ( log N − log M i )] . Substitution in the variational formula and optimization over N N N prove the trace comparison. For the converse take F > 0 F>0 F > 0 , normalized by Tr N = 1 \operatorname{Tr}N=1 Tr N = 1 , and set A i = J i A_i=J_i A i = J i and K = − ( ∑ i w i − 1 ) log N K=-(\sum_i w_i-1)\log N K = − ( ∑ i w i − 1 ) log N . The resulting normalized trace bound, followed by Klein’s inequality exactly as in Theorem 4.1, gives the weighted marginal entropy bound. Regularization includes semidefinite inputs. Its exact region is Theorem 13.3. □ \square □
Marginal equality, reciprocity, and sharp fields
The half vector of Section 13 is sharp at a fixed field precisely when the law and its complement have nonzero compatible translation directions. On connected support, exact integration gives two Bernoulli product marginals, the sharp-field locus, and every matrix extremizer.
A joint kernel and quantitative strictness
Absorb a fixed positive external field and write
f ( z ) = ∑ B ∈ B ( M ) μ ( B ) z B , f ( 1 ) = 1. f(z)=\sum_{B\in\mathcal{B}(M)}\mu(B)z^B,\qquad f(\mathbf{1})=1. f ( z ) = B ∈ B ( M ) ∑ μ ( B ) z B , f ( 1 ) = 1.
Assume that f f f and f ∗ ( z ) = z [ n ] f ( z − 1 ) f^*(z)=z^{[n]}f(z^{-1}) f ∗ ( z ) = z [ n ] f ( z − 1 ) are Lorentzian. Remove deterministic coordinates; the point-mass case is trivial. Thus r = rank M r=\operatorname{rank}M r = rank M , s = n − r s=n-r s = n − r are positive and p i = Pr ( i ∈ B ) p_i=\Pr(i\in B) p i = Pr ( i ∈ B ) , q i = 1 − p i q_i=1-p_i q i = 1 − p i belong to ( 0 , 1 ) (0,1) ( 0 , 1 ) . The selected-particle distribution assigns probability p i / r p_i/r p i / r to label i i i . Put
A = ∇ 2 f ( 1 ) , A ∗ = ∇ 2 f ∗ ( 1 ) , D = diag ( p i / q i ) , J = { h : D h f = 0 , D − D h f ∗ = 0 } . A=\nabla^2 f(\mathbf{1}),\qquad A^*=\nabla^2 f^*(\mathbf{1}),\qquad D=\operatorname{diag}(p_i/q_i),\qquad\mathcal{J}=\{h:D_hf=0,\ D_{-D_h}f^*=0\}. A = ∇ 2 f ( 1 ) , A ∗ = ∇ 2 f ∗ ( 1 ) , D = diag ( p i / q i ) , J = { h : D h f = 0 , D − D h f ∗ = 0 } .
Lemma 5.1 identifies the translation space with the Hessian kernel in degree at least two. In degree one it is the mean-zero hyperplane. Consequently, uniformly in r , s ≥ 1 r,s\ge1 r , s ≥ 1 ,
J = { h : p T h = 0 , A h = 0 , A ∗ D h = 0 } . \mathcal{J}=\{h:p^{\mathsf{T}}h=0,\ Ah=0,\ A^*Dh=0\}. J = { h : p T h = 0 , A h = 0 , A ∗ D h = 0 } .
For p T h = 0 p^{\mathsf T}h=0 p T h = 0 define
Q ( h ) = − h T A h r − ( D h ) T A ∗ ( D h ) s , V ( h ) = ∑ i p i q i h i 2 , γ = min h ≠ 0 , p T h = 0 Q ( h ) V ( h ) . Q(h)=-\frac{h^{\mathsf T}Ah}{r}-\frac{(Dh)^{\mathsf T}A^*(Dh)}{s},\qquad V(h)=\sum_i\frac{p_i}{q_i}h_i^2,\qquad\gamma=\min_{\substack{h\ne0,\ p^{\mathsf T}h=0}}\frac{Q(h)}{V(h)}. Q ( h ) = − r h T A h − s ( D h ) T A ∗ ( D h ) , V ( h ) = i ∑ q i p i h i 2 , γ = h = 0 , p T h = 0 min V ( h ) Q ( h ) .
Both summands of Q Q Q are nonnegative by the Lorentzian signature, since q T D h = p T h = 0 q^{\mathsf T}Dh=p^{\mathsf T}h=0 q T D h = p T h = 0 . Its kernel is J \mathcal{J} J ; hence γ > 0 \gamma>0 γ > 0 exactly when J = 0 \mathcal{J}=0 J = 0 .
For a positive semidefinite matrix function F F F on a register Q \mathcal{Q} Q , put σ = E μ F \sigma=\mathbb{E}_{\mu}F σ = E μ F and write
M μ ( F ) = ∑ i H Bern ( p i ) ( E [ F ∣ X i ] ) , X i = 1 i ∈ B . \mathcal{M}_{\mu}(F)=\sum_i\mathcal{H}_{\operatorname{Bern}(p_i)}\bigl(\mathbb{E}[F\mid X_i]\bigr),\qquad X_i=\mathbf{1}_{i\in B}. M μ ( F ) = i ∑ H Bern ( p i ) ( E [ F ∣ X i ] ) , X i = 1 i ∈ B .
Theorem 14.1 (Exact sharpness criterion and quantitative strictness). In every finite matrix dimension,
1 2 M μ ( F ) ≤ 1 1 + γ H μ ( F ) . \frac{1}{2}\mathcal{M}_{\mu}(F)\leq\frac{1}{1+\gamma}\mathcal{H}_{\mu}(F). 2 1 M μ ( F ) ≤ 1 + γ 1 H μ ( F ) . The optimal complete coefficient β μ \beta_{\mu} β μ in M μ ( F ) / 2 ≤ β μ H μ ( F ) \mathcal{M}_{\mu}(F)/2\leq\beta_{\mu}\mathcal{H}_{\mu}(F) M μ ( F ) /2 ≤ β μ H μ ( F ) , taken over all finite matrix dimensions, equals one if and only if J ≠ 0 \mathcal{J}\ne0 J = 0 . When J = 0 \mathcal{J}=0 J = 0 , this gives the dimension-independent estimate β μ ≤ ( 1 + γ ) − 1 < 1 \beta_{\mu}\leq(1+\gamma)^{-1}<1 β μ ≤ ( 1 + γ ) − 1 < 1 . The same bound and coefficient-one criterion hold when every entropy is replaced by H ν , a ( F ) = Tr E ν F a − Tr ( E ν F ) a \mathcal{H}_{\nu,a}(F)=\operatorname{Tr}\mathbb{E}_{\nu}F^a-\operatorname{Tr}(\mathbb{E}_{\nu}F)^a H ν , a ( F ) = Tr E ν F a − Tr ( E ν F ) a , 1 < a ≤ 2 1<a\leq2 1 < a ≤ 2 , also for pointwise trace-one inputs.
Proof. Compress to supp σ \operatorname{supp}\sigma supp σ . For σ > 0 \sigma>0 σ > 0 use the Bogoliubov–Kubo–Mori (BKM) inner product
⟨ U , V ⟩ σ = Tr [ U ∫ 0 ∞ ( σ + t I ) − 1 V ( σ + t I ) − 1 d t ] . \langle U,V\rangle_{\sigma}=\operatorname{Tr}\left[U\int_0^\infty(\sigma+tI)^{-1}V(\sigma+tI)^{-1}\,dt\right]. ⟨ U , V ⟩ σ = Tr [ U ∫ 0 ∞ ( σ + t I ) − 1 V ( σ + t I ) − 1 d t ] .
Put G i = E [ F ∣ X i = 1 ] G_i=\mathbb{E}[F\mid X_i=1] G i = E [ F ∣ X i = 1 ] , K i = E [ F ∣ X i = 0 ] K_i=\mathbb{E}[F\mid X_i=0] K i = E [ F ∣ X i = 0 ] , Z i = G i − σ Z_i=G_i-\sigma Z i = G i − σ , and Y i = K i − σ = − ( p i / q i ) Z i Y_i=K_i-\sigma=-(p_i/q_i)Z_i Y i = K i − σ = − ( p i / q i ) Z i . Write D g ( U ∥ V ) = Tr [ U ( log U − log V ) − U + V ] D_g(U\mathbin{\|}V)=\operatorname{Tr}[U(\log U-\log V)-U+V] D g ( U ∥ V ) = Tr [ U ( log U − log V ) − U + V ] . The profile identity of Theorem 3.1 and the quadratic remainder in Theorem 3.4 give
δ + : = H μ ( F ) − ∑ i p i D g ( G i ∥ σ ) ≥ − 1 r ∑ i j A i j ⟨ Z i , Z j ⟩ σ , \delta_+:=\mathcal{H}_{\mu}(F)-\sum_i p_iD_g(G_i\mathbin{\|}\sigma)\geq-\frac{1}{r}\sum_{ij}A_{ij}\langle Z_i,Z_j\rangle_{\sigma}, δ + := H μ ( F ) − i ∑ p i D g ( G i ∥ σ ) ≥ − r 1 ij ∑ A ij ⟨ Z i , Z j ⟩ σ ,
δ − : = H μ ( F ) − ∑ i q i D g ( K i ∥ σ ) ≥ − 1 s ∑ i j A i j ∗ ⟨ Y i , Y j ⟩ σ . \delta_-:=\mathcal{H}_{\mu}(F)-\sum_i q_iD_g(K_i\mathbin{\|}\sigma)\geq-\frac{1}{s}\sum_{ij}A^*_{ij}\langle Y_i,Y_j\rangle_{\sigma}. δ − := H μ ( F ) − i ∑ q i D g ( K i ∥ σ ) ≥ − s 1 ij ∑ A ij ∗ ⟨ Y i , Y j ⟩ σ .
Indeed H r − r H 1 ≥ ( r − 1 ) ( H 2 − 2 H 1 ) H_r-rH_1\geq(r-1)(H_2-2H_1) H r − r H 1 ≥ ( r − 1 ) ( H 2 − 2 H 1 ) , and the ordered pair matrix is A / [ r ( r − 1 ) ] A/[r(r-1)] A / [ r ( r − 1 )] . For r = 1 r=1 r = 1 , the first selected channel is the identity and its remainder is zero; likewise when s = 1 s=1 s = 1 . Since ∑ i p i Z i = 0 \sum_i p_iZ_i=0 ∑ i p i Z i = 0 , expansion in a BKM orthonormal basis yields
2 H μ ( F ) − M μ ( F ) = δ + + δ − ≥ γ ∑ i p i q i ∥ Z i ∥ σ 2 . 2\mathcal{H}_{\mu}(F)-\mathcal{M}_{\mu}(F)=\delta_++\delta_-\geq\gamma\sum_i\frac{p_i}{q_i}\lVert Z_i\rVert_{\sigma}^{2}. 2 H μ ( F ) − M μ ( F ) = δ + + δ − ≥ γ i ∑ q i p i ∥ Z i ∥ σ 2 .
The tangent inequality for log \log log , used in Theorem 3.4, gives D g ( U ∥ σ ) ≤ ∥ U − σ ∥ σ 2 D_g(U\mathbin{\|}\sigma)\leq\lVert U-\sigma\rVert_{\sigma}^{2} D g ( U ∥ σ ) ≤ ∥ U − σ ∥ σ 2 . Therefore
M μ ( F ) ≤ ∑ i ( p i ∥ Z i ∥ σ 2 + q i ∥ Y i ∥ σ 2 ) = ∑ i p i q i ∥ Z i ∥ σ 2 . \mathcal{M}_{\mu}(F)\leq\sum_i\left(p_i\lVert Z_i\rVert_{\sigma}^{2}+q_i\lVert Y_i\rVert_{\sigma}^{2}\right)=\sum_i\frac{p_i}{q_i}\lVert Z_i\rVert_{\sigma}^{2}. M μ ( F ) ≤ i ∑ ( p i ∥ Z i ∥ σ 2 + q i ∥ Y i ∥ σ 2 ) = i ∑ q i p i ∥ Z i ∥ σ 2 .
This proves the bound; regularization includes singular inputs. The scalar equality family constructed next proves logarithmic sharpness. For trace powers, apply this bound to F + t I F+tI F + t I and integrate the shifted-log formula from Theorem 8.1; continuity gives a = 2 a=2 a = 2 . If 0 ≠ h ∈ J 0\ne h\in\mathcal{J} 0 = h ∈ J , the tangent u ( B ) = ∑ i ∈ B h i u(B)=\sum_{i\in B}h_i u ( B ) = ∑ i ∈ B h i of that equality family has half-marginal variance ratio one. Thus the trace-one qubit perturbation F ε = 1 2 diag ( 1 + ε u , 1 − ε u ) F_{\varepsilon}=\frac{1}{2}\operatorname{diag}(1+\varepsilon u,1-\varepsilon u) F ε = 2 1 diag ( 1 + ε u , 1 − ε u ) has power-entropy ratio tending to one, with exact equality for a = 2 a=2 a = 2 . If J = 0 \mathcal{J}=0 J = 0 , the strict bound applies. □
Exact integration of the joint directions.
Theorem 14.2 (A pointwise primal–dual identity). The normalized scalar equality densities for the logarithmic half-weight inequality are exactly
F B = ∏ i ∈ B ( 1 + h i ) = ∏ i ∉ B ( 1 − p i q i h i ) , h ∈ J , − 1 ≤ h i ≤ q i / p i . F_B = \prod_{i\in B}(1+h_i)=\prod_{i\notin B}\left(1-\frac{p_i}{q_i}h_i\right),\qquad h\in\mathcal{J},\quad-1\le h_i\le q_i/p_i. F B = i ∈ B ∏ ( 1 + h i ) = i ∈ / B ∏ ( 1 − q i p i h i ) , h ∈ J , − 1 ≤ h i ≤ q i / p i . In particular the two products agree on every supported basis.
Proof. First let a i = 1 + h i > 0 a_i=1+h_i>0 a i = 1 + h i > 0 , b i = 1 − ( p i / q i ) h i > 0 b_i=1-(p_i/q_i)h_i>0 b i = 1 − ( p i / q i ) h i > 0 . Theorem 5.2 gives normalized densities F B + = a B F_B^+=a^B F B + = a B , F B − = b B c F_B^-=b^{B^c} F B − = b B c , with entropies
S + = ∑ i p i a i log a i , S − = ∑ i q i b i log b i . S_+=\sum_i p_i a_i\log a_i,\qquad S_-=\sum_i q_i b_i\log b_i. S + = i ∑ p i a i log a i , S − = i ∑ q i b i log b i . Their present and absent conditional means are a i , b i a_i,b_i a i , b i in both cases, because p i a i + q i b i = 1 p_i a_i+q_i b_i=1 p i a i + q i b i = 1 . Applying complementary deletion to F + F^+ F + and original deletion to F − F^- F − gives S − ≤ S + ≤ S − S_-\le S_+\le S_- S − ≤ S + ≤ S − . Moreover,
D ( μ F + ∥ μ F − ) = S + − E μ F + [ F + log F − ] = S + − S − = 0. D(\mu_{F^+}\Vert\mu_{F^-})=S_+-\mathbb{E}_{\mu_{F^+}}[F^+\log F^-]=S_+-S_-=0. D ( μ F + ∥ μ F − ) = S + − E μ F + [ F + log F − ] = S + − S − = 0. Thus the densities agree pointwise. Replace h h h by t h th t h and let t ↑ 1 t\uparrow1 t ↑ 1 for boundary parameters. Conversely, half equality forces both nonnegative selected deficits to vanish. Their scalar equality classifications and the common Bernoulli mean give precisely a , b , h a,b,h a , b , h above. In degree one, the same description is just the mean-zero condition.
For 0 ≠ h ∈ J 0\ne h\in\mathcal{J} 0 = h ∈ J , small nonzero multiples give nonconstant positive densities: the derivative u ( B ) = ∑ i ∈ B h i u(B)=\sum_{i\in B}h_i u ( B ) = ∑ i ∈ B h i has Var u = ∑ i p i h i 2 > 0 \operatorname{Var}u=\sum_i p_i h_i^2>0 Var u = ∑ i p i h i 2 > 0 , since A h = 0 Ah=0 A h = 0 . This completes Theorem 14.1. □
Automatic reciprocity and connected rigidity.
The next algebraic statement uses only nonnegative coefficients and the translation space N f = { h : D h f = 0 } \mathcal{N}_f=\{h:D_hf=0\} N f = { h : D h f = 0 } .
Theorem 14.3 (Automatic reciprocity and independent halves). Let f ≠ 0 f\ne0 f = 0 be homogeneous and multiaffine of degree m ≥ 1 m\ge1 m ≥ 1 on 2 m 2m 2 m variables, with nonnegative coefficients. Suppose h ∈ N f h\in\mathcal{N}_f h ∈ N f and h i ≠ 0 h_i\ne0 h i = 0 for every i i i . Then its two sign sets A , A ′ A,A' A , A ′ each have size m m m , and both are supported monomials. With λ i = ∣ h i ∣ \lambda_i=|h_i| λ i = ∣ h i ∣ , the normalized polynomial g ( z ) = f ( λ z ) / f ( λ ) g(z)=f(\lambda z)/f(\lambda) g ( z ) = f ( λ z ) / f ( λ ) satisfies
g ∗ = g , g ( z A , 1 A ′ ) = 2 − m ∏ i ∈ A ( 1 + z i ) , g ( 1 A , z A ′ ) = 2 − m ∏ j ∈ A ′ ( 1 + z j ) . g^*=g,\qquad g(z_A,1_{A'})=2^{-m}\prod_{i\in A}(1+z_i),\qquad g(1_A,z_{A'})=2^{-m}\prod_{j\in A'}(1+z_j). g ∗ = g , g ( z A , 1 A ′ ) = 2 − m i ∈ A ∏ ( 1 + z i ) , g ( 1 A , z A ′ ) = 2 − m j ∈ A ′ ∏ ( 1 + z j ) . Equivalently,
f ∗ ( z ) = ( ∏ i ∣ h i ∣ ) − 1 f ( h 1 2 z 1 , … , h 2 m 2 z 2 m ) . (106) f^*(z)=\left(\prod_i|h_i|\right)^{-1}f(h_1^2z_1,\ldots,h_{2m}^2z_{2m}). \tag*{(106)} f ∗ ( z ) = ( i ∏ ∣ h i ∣ ) − 1 f ( h 1 2 z 1 , … , h 2 m 2 z 2 m ) . (106) In particular, if f f f is Lorentzian, so is f ∗ f^* f ∗ . Each half of the balanced reference law consists of independent Bernoulli( 1 / 2 ) (1/2) ( 1/2 ) coordinates.
Proof. Put ϵ i = sign ( h i ) \epsilon_i=\operatorname{sign}(h_i) ϵ i = sign ( h i ) . Then D ϵ g = 0 D_\epsilon g=0 D ϵ g = 0 and g ( 1 + ϵ ) = g ( 1 − ϵ ) = g ( 1 ) = 1 g(1+\epsilon)=g(1-\epsilon)=g(1)=1 g ( 1 + ϵ ) = g ( 1 − ϵ ) = g ( 1 ) = 1 . Nonnegativity therefore makes each sign set contain a supported m m m -subset. There are only 2 m 2m 2 m labels, so each set has size m m m and its coefficient in g g g is 2 − m 2^{-m} 2 − m . Translation now gives
g ( z A , 1 A ′ ) = g ( z A + 1 A , 0 A ′ ) = 2 − m ∏ i ∈ A ( 1 + z i ) , g(z_A,1_{A'})=g(z_A+1_A,0_{A'})=2^{-m}\prod_{i\in A}(1+z_i), g ( z A , 1 A ′ ) = g ( z A + 1 A , 0 A ′ ) = 2 − m i ∈ A ∏ ( 1 + z i ) ,
and similarly for A ′ A' A ′ .
For reciprocity, the sign change places g g g at middle weight in the Boolean s l 2 \mathfrak{sl}_2 sl 2 representation on 2 m 2m 2 m variables. A middle-weight vector killed by lowering is invariant; the Weyl element then acts by signed complementation. Put g ^ ( w ) = g ( ( ϵ i w i ) i ) \widehat{g}(w)=g((\epsilon_iw_i)_i) g ( w ) = g (( ϵ i w i ) i ) . On the multiaffine vector space consider
F = ∑ i ∂ w i , E = ∑ i w i ( 1 − w i ∂ w i ) , [ E , F ] = 2 ∑ i w i ∂ w i − 2 m . F=\sum_i\partial_{w_i},\qquad E=\sum_iw_i(1-w_i\partial_{w_i}),\qquad[E,F]=2\sum_iw_i\partial_{w_i}-2m. F = i ∑ ∂ w i , E = i ∑ w i ( 1 − w i ∂ w i ) , [ E , F ] = 2 i ∑ w i ∂ w i − 2 m .
In the inner product making monomials orthonormal, E ∗ = F E^*=F E ∗ = F . Since g ^ \hat{g} g ^ has degree m m m and F g ^ = 0 F\hat{g}=0 F g ^ = 0 , the commutator gives ∥ E g ^ ∥ 2 = ⟨ F E g ^ , g ^ ⟩ = 0 \lVert E\hat{g}\rVert^2=\langle FE\hat{g},\hat{g}\rangle=0 ∥ E g ^ ∥ 2 = ⟨ F E g ^ , g ^ ⟩ = 0 . Thus both raising and lowering annihilate g ^ \hat{g} g ^ . On one variable, exp ( F ) exp ( − E ) exp ( F ) \exp(F)\exp(-E)\exp(F) exp ( F ) exp ( − E ) exp ( F ) sends 1 ↦ − w 1\mapsto-w 1 ↦ − w and w ↦ 1 w\mapsto1 w ↦ 1 . On degree m m m in 2 m 2m 2 m labels its tensor action is ( − 1 ) m (-1)^m ( − 1 ) m times coefficient complementation. Hence g ^ ∗ = ( − 1 ) m g ^ \hat{g}^*=(-1)^m\hat{g} g ^ ∗ = ( − 1 ) m g ^ . Exactly m m m signs ϵ i \epsilon_i ϵ i are negative, so ϵ S c = ( − 1 ) m ϵ S \epsilon^{S^c}=(-1)^m\epsilon^S ϵ S c = ( − 1 ) m ϵ S for every m m m -subset S S S . The signs cancel, proving g ^ ∗ = g ^ \hat{g}^*=\hat{g} g ^ ∗ = g ^ . Substituting in this identity gives (14.1); positive diagonal substitutions preserve Lorentzianity. □
The size 2 m 2m 2 m is minimal: the same sign argument shows that a translation with no zero coordinate in degree m m m requires at least m m m labels of each sign.
The coefficient law factors over matroid components by Lemma 13.2. Translation spaces, and hence J \mathcal{J} J , split over these components: each logarithmic directional derivative depends on its own variables and has homogeneity − 1 -1 − 1 .
Theorem 14.4 (Two bases and independent coordinate halves). For a nontrivial connected component C C C , of rank m m m , dim J C ≤ 1 \dim\mathcal{J}_C\leq1 dim J C ≤ 1 . If it is nonzero, there are two bases C = A ⊔ A ′ C=A\sqcup A' C = A ⊔ A ′ , each of size m m m , and π ∈ ( 0 , 1 ) \pi\in(0,1) π ∈ ( 0 , 1 ) , such that
p i = π ( i ∈ A ) , p i = 1 − π ( i ∈ A ′ ) , J C = R h , h i = { π − 1 , i ∈ A , − ( 1 − π ) − 1 , i ∈ A ′ . p_i=\pi\ (i\in A),\qquad p_i=1-\pi\ (i\in A'),\qquad\mathcal{J}_C=\mathbb{R}h,\qquad h_i=\begin{cases}\pi^{-1},&i\in A,\\-(1-\pi)^{-1},&i\in A'.\end{cases} p i = π ( i ∈ A ) , p i = 1 − π ( i ∈ A ′ ) , J C = R h , h i = { π − 1 , − ( 1 − π ) − 1 , i ∈ A , i ∈ A ′ . Within A A A the coordinates are independent Bernoulli( π ) (\pi) ( π ) ; within A ′ A' A ′ they are independent Bernoulli( 1 − π ) (1-\pi) ( 1 − π ) . In particular ∣ B ∩ A ∣ |B\cap A| ∣ B ∩ A ∣ is Binomial( m , π ) (m,\pi) ( m , π ) . For m ≥ 2 m\geq2 m ≥ 2 , the component is simple and cosimple.
Proof. Choose 0 ≠ h ∈ J C 0\ne h\in\mathcal{J}_C 0 = h ∈ J C and put k i = − p i h i / q i k_i=-p_i h_i/q_i k i = − p i h i / q i . Any two elements i , j i,j i , j of a connected matroid can be exchanged in some basis. Compare the pointwise identities of Theorem 14.2 on those bases, at parameter t h th t h :
( 1 + t h i ) ( 1 + t k i ) = ( 1 + t h j ) ( 1 + t k j ) . (1+th_i)(1+tk_i)=(1+th_j)(1+tk_j). ( 1 + t h i ) ( 1 + t k i ) = ( 1 + t h j ) ( 1 + t k j ) .
All pairs ( h i , k i ) (h_i,k_i) ( h i , k i ) have the same sum and product. A zero coordinate would force every coordinate to vanish. Otherwise their common unordered pair is { u , v } \{u,v\} { u , v } , with u > 0 > v u>0>v u > 0 > v , and the two sign classes have probabilities π = − v / ( u − v ) \pi=-v/(u-v) π = − v / ( u − v ) and 1 − π 1-\pi 1 − π . The identities ∑ i p i h i = 0 \sum_i p_i h_i=0 ∑ i p i h i = 0 and ∑ i p i = m \sum_i p_i=m ∑ i p i = m show that both classes have size m m m . Evaluation at any coordinate is injective on J C \mathcal{J}_C J C , proving its dimension bound.
After the positive diagonal rescaling to the translation ( 1 A , − 1 A ′ ) (1_A,-1_{A'}) ( 1 A , − 1 A ′ ) , Theorem 14.3 makes both halves bases and their coordinates independent. Rescaling back gives the stated probabilities and binomial count. A parallel pair must cross the halves. Its probabilities sum to one, so every basis contains exactly one member. That pair and its complement have ranks 1 1 1 and m − 1 m-1 m − 1 , making the pair a direct component. Connectedness excludes this for m ≥ 2 m\geq2 m ≥ 2 . The same argument in the dual proves cosimplicity. □
Sharp fields and product components
For a positive field λ \lambda λ , write β λ \beta_\lambda β λ for the optimal complete coefficient in M μ λ ( F ) / 2 ≤ β λ H μ λ ( F ) \mathcal{M}_{\mu_\lambda}(F)/2\leq\beta_\lambda\mathcal{H}_{\mu_\lambda}(F) M μ λ ( F ) /2 ≤ β λ H μ λ ( F ) .
Theorem 14.5 (The complete sharp-field locus). For a connected rank-m m m law with f , f ∗ f,f^* f , f ∗ Lorentzian, β λ = 1 \beta_\lambda=1 β λ = 1 at some positive field if and only if
n = 2 m and N f contains a vector h with every h i ≠ 0. n=2m\qquad\text{and}\qquad\mathcal{N}_f\text{ contains a vector }h\text{ with every }h_i\ne0. n = 2 m and N f contains a vector h with every h i = 0. When this holds, N f = R h \mathcal{N}_f=\mathbb{R}h N f = R h , and all sharp fields are exactly
λ i = a ∣ h i ∣ ( i ∈ A ) , λ i = b ∣ h i ∣ ( i ∈ A ′ ) , a , b > 0. \lambda_i=a|h_i|\quad(i\in A),\qquad\lambda_i=b|h_i|\quad(i\in A'),\qquad a,b>0. λ i = a ∣ h i ∣ ( i ∈ A ) , λ i = b ∣ h i ∣ ( i ∈ A ′ ) , a , b > 0. The partition is unique up to exchanging its halves. Modulo common scaling of the field, the sharp locus is one line in logarithmic field coordinates. The sufficient condition itself makes the hypothesis that f ∗ f^* f ∗ is Lorentzian automatic.
Proof. For necessity use the joint-kernel criterion and connected rigidity of Theorems 14.1 and 14.4 : a sharp field forces n = 2 m n=2m n = 2 m and a joint translation with no zero coordinate. Undoing the field gives the stated h h h . Conversely, Theorem 14.3 gives a self-complementary balanced law with all inclusion probabilities 1 / 2 1/2 1/2 . Its translations ϵ , − ϵ \epsilon,-\epsilon ϵ , − ϵ satisfy the joint criterion, so this field is sharp.
To prove uniqueness, let v ∈ N f v \in\mathcal{N}_f v ∈ N f . For small real t t t , h + t v h+tv h + t v has the same signs as h h h , and the fields ∣ h + t v ∣ |h+tv| ∣ h + t v ∣ all have inclusion vector ( 1 / 2 , … , 1 / 2 ) (1/2,\ldots,1/2) ( 1/2 , … , 1/2 ) . The exponential-family mean map is injective modulo common scaling on connected matroid support. Indeed, if two fields have the same inclusion vector, then
D ( μ λ ∥ μ ν ) + D ( μ ν ∥ μ λ ) = ∑ i ( p i ( λ ) − p i ( ν ) ) log ( λ i / ν i ) = 0. D(\mu_\lambda\Vert\mu_\nu)+D(\mu_\nu\Vert\mu_\lambda)=\sum_i(p_i(\lambda)-p_i(\nu))\log(\lambda_i/\nu_i)=0. D ( μ λ ∥ μ ν ) + D ( μ ν ∥ μ λ ) = i ∑ ( p i ( λ ) − p i ( ν )) log ( λ i / ν i ) = 0.
Their laws coincide, so ( λ / ν ) B (\lambda/\nu)^B ( λ / ν ) B is constant on all bases. Connected basis exchanges force all λ i / ν i \lambda_i/\nu_i λ i / ν i to agree. Consequently ∣ h + t v ∣ |h+tv| ∣ h + t v ∣ is proportional to ∣ h ∣ |h| ∣ h ∣ , and v ∈ R h v \in\mathbb{R}h v ∈ R h .
At an arbitrary sharp field, the primal translation is proportional to ( h i / λ i ) i (h_i/\lambda_i)_i ( h i / λ i ) i . Connected rigidity requires one common positive value and one common negative value, giving the displayed field locus. Conversely, at such a field the present probabilities are a / ( a + b ) a/(a+b) a / ( a + b ) on A A A and b / ( a + b ) b/(a+b) b / ( a + b ) on A ′ A' A ′ . The primal direction is ( 1 / a , − 1 / b ) (1/a,-1/b) ( 1/ a , − 1/ b ) ; the complementary direction required by the joint criterion is
( − 1 / b , 1 / a ) = − ( a b ) − 1 ( a , − b ) . (-1/b,1/a)=-(ab)^{-1}(a,-b). ( − 1/ b , 1/ a ) = − ( ab ) − 1 ( a , − b ) .
Reciprocity gives ( a , − b ) (a,-b) ( a , − b ) as a complementary translation. Thus every displayed field is sharp. □ \square □
Write G = 2 m g G=2^m g G = 2 m g , p = a / ( a + b ) p=a/(a+b) p = a / ( a + b ) , and q = 1 − p q=1-p q = 1 − p . We call G G G balanced: it is multiaffine and homogeneous of degree m m m on two m m m -element halves, has nonnegative coefficients, pure-half coefficients one, and D ( 1 , − 1 ) G = 0 D(1,-1)G=0 D ( 1 , − 1 ) G = 0 . The whole sharp family is f p ( z ) = G ( p z A , q z A ′ ) f_p(z)=G(pz_A,qz_{A'}) f p ( z ) = G ( p z A , q z A ′ ) , with
f p ( z A , 1 ) = ∏ i ∈ A ( p z i + q ) , f p ( 1 , z A ′ ) = ∏ j ∈ A ′ ( q z j + p ) . f_p(z_A,1)=\prod_{i\in A}(pz_i+q),\qquad f_p(1,z_{A'})=\prod_{j\in A'}(qz_j+p). f p ( z A , 1 ) = i ∈ A ∏ ( p z i + q ) , f p ( 1 , z A ′ ) = j ∈ A ′ ∏ ( q z j + p ) .
The count ∣ B ∩ A ∣ |B\cap A| ∣ B ∩ A ∣ is Binomial( m , p ) (m,p) ( m , p ) ; the conditional law given that count is independent of p p p . In particular
D ( μ p ∥ μ t ) = m D ( Bern ( p ) ∥ Bern ( t ) ) . D(\mu_p\Vert\mu_t)=mD(\operatorname{Bern}(p)\Vert\operatorname{Bern}(t)). D ( μ p ∥ μ t ) = m D ( Bern ( p ) ∥ Bern ( t )) .
These product marginals also explain the half bound directly. The complete product inequality and data processing give
∑ i ∈ A H μ i ( E [ F ∣ X i ] ) ≤ H μ A ( E [ F ∣ X A ] ) ≤ H μ ( F ) . \sum_{i\in A}\mathcal{H}_{\mu_i}(\mathbb{E}[F\mid X_i])\leq\mathcal{H}_{\mu_A}(\mathbb{E}[F\mid X_A])\leq\mathcal{H}_{\mu}(F). i ∈ A ∑ H μ i ( E [ F ∣ X i ]) ≤ H μ A ( E [ F ∣ X A ]) ≤ H μ ( F ) .
Apply the same argument to A ′ A' A ′ and average. This uses only the product strong subadditivity (13.3) and the two product marginals.
Theorem 14.6 (Exact component tensorization and generic strictness). Let μ = ⨂ C μ C \mu=\bigotimes_C\mu_C μ = ⨂ C μ C be the factorization over the nontrivial matroid components. Then
β μ = max C β μ C . \beta_\mu=\max_C\beta_{\mu_C}. β μ = C max β μ C . The same value is obtained when inputs are restricted to pointwise trace-one matrices.
Consequently the global set of sharp positive fields is the union of the component cylinders specified by Theorem 14.5, with the fields outside the selected component arbitrary. If a binary direct component exists, every positive field is sharp. Otherwise each nonempty cylinder has codimension ∣ C ∣ − 2 ≥ 2 |C|-2\geq2 ∣ C ∣ − 2 ≥ 2 in logarithmic field coordinates. The complete half bound is therefore strict on an open dense set of full measure, uniformly in matrix dimension; the improvement is uniform on compact subsets of that set.
Proof. For F C = E [ F ∣ B ∩ C ] F_C=\mathbb{E}[F\mid B\cap C] F C = E [ F ∣ B ∩ C ] , apply the component inequalities and (13.3):
1 2 M μ ( F ) ≤ ∑ C β μ C H μ C ( F C ) ≤ ( max C β μ C ) H μ ( F ) . \frac{1}{2}\mathcal{M}_{\mu}(F)\leq\sum_C\beta_{\mu_C}\mathcal{H}_{\mu_C}(F_C)\leq\left(\max_C\beta_{\mu_C}\right)\mathcal{H}_{\mu}(F). 2 1 M μ ( F ) ≤ C ∑ β μ C H μ C ( F C ) ≤ ( C max β μ C ) H μ ( F ) .
Inputs depending on one component prove the reverse inequality for the optimal coefficients. To compare unrestricted inputs with trace-one encodings, enlarge the register by one dimension and use the flag extension
F ^ ε ( B ) = diag ( ε F ( B ) , 1 − ε Tr F ( B ) ) . \widehat{F}_{\varepsilon}(B)=\operatorname{diag}\bigl(\varepsilon F(B),1-\varepsilon\operatorname{Tr}F(B)\bigr). F ε ( B ) = diag ( εF ( B ) , 1 − ε Tr F ( B ) ) .
For small positive ε \varepsilon ε this is a pointwise density matrix. The first block contributes exactly ε H μ ( F ) \varepsilon H_{\mu}(F) ε H μ ( F ) , and the scalar block contributes O ( ε 2 ) O(\varepsilon^{2}) O ( ε 2 ) ; the same expansion holds after every marginal channel. Thus every nonzero entropy ratio is approached by trace-one ratios. The reverse comparison is immediate by restriction. The suprema agree; exact attainment is treated below. Lemma 13.2 and Theorem 14.5 give the stated cylinders. Binary cylinders fill their field space; other eligible components have at least four labels. A finite union of proper affine subspaces in logarithmic coordinates is closed and null. Continuity of γ > 0 \gamma>0 γ > 0 off this union gives compact uniformity.
Matrix equality and exact attainment.
Theorem 14.7 (Only binary components carry noncommuting factors). Let S S S be the nonbinary connected components with J C ≠ 0 J_{C}\ne0 J C = 0 , and let U U U be the binary U 1 , 2 U_{1,2} U 1 , 2 direct components. All logarithmic half-weight equality densities with Tr E μ F = 1 \operatorname{Tr}E_{\mu}F=1 Tr E μ F = 1 have exactly the form
F B = ⨁ z w z ( ∏ C ∈ S ∏ i ∈ B ∩ C ( 1 + t C , z h i C ) τ 0 z ⊗ ⨂ D ∈ U A Y D ( B ) D , z ) . (107) F_{B}=\bigoplus_{z}w_{z}\left(\prod_{C\in S}\prod_{i\in B\cap C}(1+t_{C,z}h_{i}^{C})\tau_{0}^{z}\otimes\bigotimes_{D\in U}A_{Y_{D}(B)}^{D,z}\right). \tag*{(107)} F B = z ⨁ w z ( C ∈ S ∏ i ∈ B ∩ C ∏ ( 1 + t C , z h i C ) τ 0 z ⊗ D ∈ U ⨂ A Y D ( B ) D , z ) . (107) Here w z > 0 w_{z}>0 w z > 0 , ∑ z w z = 1 \sum_{z}w_{z}=1 ∑ z w z = 1 ; Y D Y_{D} Y D is the binary choice with reference law π D \pi_{D} π D ; ∑ y π D ( y ) A y D , z = τ D z \sum_{y}\pi_{D}(y)A_{y}^{D,z}=\tau_{D}^{z} ∑ y π D ( y ) A y D , z = τ D z ; the A y D , z A_{y}^{D,z} A y D , z are positive semidefinite and the τ 0 z , τ D z \tau_{0}^{z},\tau_{D}^{z} τ 0 z , τ D z are faithful density matrices on their respective tensor factors. For each C ∈ S C\in S C ∈ S , take h C h^{C} h C and π C \pi_{C} π C from Theorem 14.4 ; then − π C ≤ t C , z ≤ 1 − π C -\pi_{C}\le t_{C,z}\le1-\pi_{C} − π C ≤ t C , z ≤ 1 − π C . Empty products are permitted, with 0 0 = 1 0^{0}=1 0 0 = 1 .
Proof. Suppose first that the global rank is at least two. Primal and dual equality give ( G i − σ ) i ∈ J ⊗ Herm ( Q ) (G_{i}-\sigma)_{i}\in\mathcal{J}\otimes\operatorname{Herm}(Q) ( G i − σ ) i ∈ J ⊗ Herm ( Q ) , by (33) ; for a degree-one complement, its null identity is the centered mean relation. Apply Theorem 7.6 . Outside surviving components G i = σ G_{i}=\sigma G i = σ . On each surviving nonbinary component, Theorem 14.4 makes every parallel class a singleton. Its matrices are therefore scalar multiples of σ \sigma σ in each central block, with parameters on the one-dimensional joint line. On a binary component the class mean is σ \sigma σ , leaving exactly its two matrix factors. Every remaining class factor is constant. This is (14.2). Conversely, entropy is additive over normalized independent tensor factors and affine over these orthogonal blocks. Each scalar factor has half equality, and the two coordinates of a binary component both reveal its entire choice. Hence (14.2) has equality. For global rank one, G i = F { i } G_{i}=F_{\{i\}} G i = F { i } . Joint nullity therefore makes F F F constant unless the component is a binary pair, where the assertion allows an arbitrary two-state ensemble.
Corollary 14.8 (Binary rigidity and unattained Holevo sharpness). For positive product-field matroid laws, β μ = 1 \beta_{\mu}=1 β μ = 1 exactly when a binary direct component exists. For arbitrary coefficients under the two Lorentzian hypotheses, an informative equality encoding satisfying Tr F B = 1 \operatorname{Tr}F_{B}=1 Tr F B = 1 exists exactly when a binary component exists. Nevertheless the optimal coefficient restricted to such encodings equals one exactly when J ≠ 0 \mathcal{J}\ne0 J = 0 . Thus, if J ≠ 0 \mathcal{J}\ne0 J = 0 and there is no binary component, scalar densities attain the bound while informative trace-one encodings approach it without attaining it.
Proof. For product-field laws, the translation classification in Theorem 5.6, using the Hessian strictness of [38 ] , gives only weighted parallel-class contrasts. Its dual gives series-class contrasts. A circuit { i , j } \{i,j\} { i , j } and cocircuit { i , k } \{i,k\} { i , k } cannot meet in exactly one element, so a common nonzero coordinate must belong to a binary direct component. Conversely such a component has a joint contrast.
Under pointwise trace one, Corollary 7.9 removes every quotient scalar tilt in (14.2); only binary factors remain. Any nonconstant trace-one ensemble on a binary component suffices. For 0 ≠ h ∈ J 0 \ne h \in\mathcal{J} 0 = h ∈ J , put u ( B ) = ∑ i ∈ B h i u(B)=\sum_{i\in B}h_i u ( B ) = ∑ i ∈ B h i and
F ε ( B ) = 1 2 diag ( 1 + ε u ( B ) , 1 − ε u ( B ) ) . F_{\varepsilon}(B)=\frac{1}{2}\operatorname{diag}(1+\varepsilon u(B),1-\varepsilon u(B)). F ε ( B ) = 2 1 diag ( 1 + ε u ( B ) , 1 − ε u ( B )) .
Differentiating the pointwise identity gives u ( B ) = ∑ i ∉ B ( − D h ) i u(B)=\sum_{i\notin B}(-Dh)_i u ( B ) = ∑ i ∈ / B ( − D h ) i . Thus u u u is simultaneously a primal and dual variance equality mode. The second-order entropy expansion makes the half-marginal ratio tend to one. When J = 0 \mathcal{J}=0 J = 0 , the strict complete bound already applies to these encodings.
For product-field matroid laws, marginal entropy equality and the unit-edge eigenvalue-two space thus single out the same binary direct components (Theorem 10.7 ).
Determinantal models and the cubic classification
For any m × m m \times m m × m unitary U U U , put
G U ( x , y ) = det ( diag ( x ) + U diag ( y ) U ∗ ) , f p ( z ) = G U ( p z A , ( 1 − p ) z A ′ ) , 0 < p < 1. G_U(x,y)=\det(\operatorname{diag}(x)+U\operatorname{diag}(y)U^*), \qquad f_p(z)=G_U(pz_A,(1-p)z_{A'}), \qquad0<p<1. G U ( x , y ) = det ( diag ( x ) + U diag ( y ) U ∗ ) , f p ( z ) = G U ( p z A , ( 1 − p ) z A ′ ) , 0 < p < 1.
Cauchy–Binet gives nonnegative coefficients. Upper-half-plane variables make the matrix’s imaginary part positive definite, so G U G_U G U is stable and hence Lorentzian [13 ] . Its full-support translation is ( 1 A , − 1 A ′ ) (1_A,-1_{A'}) ( 1 A , − 1 A ′ ) , and Theorem 14.3 gives complement reciprocity. The law of f p f_p f p is the projection determinantal law associated with the row space of [ p I , 1 − p U ] [\sqrt{p}I,\sqrt{1-p}U] [ p I , 1 − p U ] . For generic U U U , every square minor is nonzero, giving support U m , 2 m U_{m,2m} U m , 2 m , simple and cosimple for m ≥ 2 m \ge2 m ≥ 2 . Corollary 14.8 gives scalar half equality and sharp but unattained informative trace-one ratios.
For m = 3 m=3 m = 3 , self-complementarity and translation force
G M ( x , y ) = x 1 x 2 x 3 + y 1 y 2 y 3 + ∑ i , j = 1 3 M i j ( x [ 3 ] ∖ { i } y j + x i y [ 3 ] ∖ { j } ) (108) G_M(x,y)=x_1x_2x_3+y_1y_2y_3+\sum_{i,j=1}^{3}M_{ij}\left(x_{[3]\setminus\{i\}}y_j+x_i y_{[3]\setminus\{j\}}\right) \tag*{(108)} G M ( x , y ) = x 1 x 2 x 3 + y 1 y 2 y 3 + i , j = 1 ∑ 3 M ij ( x [ 3 ] ∖ { i } y j + x i y [ 3 ] ∖ { j } ) (108)
where M M M is doubly stochastic. Indeed complementation pairs the mixed coefficients; the coefficients of x [ 3 ] ∖ { i } x_{[3]\setminus\{i\}} x [ 3 ] ∖ { i } and y [ 3 ] ∖ { j } y_{[3]\setminus\{j\}} y [ 3 ] ∖ { j } in ( ∑ ∂ x − ∑ ∂ y ) G M = 0 (\sum\partial_x-\sum\partial_y)G_M=0 ( ∑ ∂ x − ∑ ∂ y ) G M = 0 give its row and column sums. Conversely those sums give the translation identity.
Theorem 14.9 (Complete cubic classification). For a 3 × 3 3 \times3 3 × 3 doubly stochastic matrix M M M , the following conditions are equivalent:
G M G_M G M is Lorentzian.
The Hessian of the single quadratic ∂ x 1 G M \partial_{x_1}G_M ∂ x 1 G M has at most one positive eigenvalue.
The three numbers M 2 j M 3 j \sqrt{M_{2j}M_{3j}} M 2 j M 3 j , 1 ≤ j ≤ 3 1 \le j \le3 1 ≤ j ≤ 3 , are the side lengths of a possibly degenerate triangle.
M i j = ∣ U i j ∣ 2 M_{ij}=|U_{ij}|^2 M ij = ∣ U ij ∣ 2 for a unitary matrix U U U .
G M = det ( diag ( x ) + U diag ( y ) U ∗ ) G_M=\det(\operatorname{diag}(x)+U\operatorname{diag}(y)U^*) G M = det ( diag ( x ) + U diag ( y ) U ∗ ) for a unitary U U U .
G M G_M G M is real stable.
Thus every connected rank-three sharp law is a two-orthonormal-bases projection determinantal law after positive rescaling.
Proof. Lorentzianity implies (2), and (5) implies stability by positivity of the imaginary part of the determinant matrix. Stability implies Lorentzianity by [13 ] . We prove the remaining implications explicitly.
Set a j = M 2 j a_j=M_{2j} a j = M 2 j , b j = M 3 j b_j=M_{3j} b j = M 3 j , and ℓ j = a j b j \ell_j=\sqrt{a_jb_j} ℓ j = a j b j . In the Hessian of ∂ x 1 G M \partial_{x_1}G_M ∂ x 1 G M , omitting its zero x 1 x_1 x 1 row and ordered by x 2 , x 3 , y 1 , y 2 , y 3 x_2,x_3,y_1,y_2,y_3 x 2 , x 3 , y 1 , y 2 , y 3 , the first block is ( 0 1 1 0 ) \begin{pmatrix}0&1\\1&0\end{pmatrix} ( 0 1 1 0 ) . Its Schur complement S S S satisfies
S 1 = 0 , S j j = − 2 ℓ j 2 , S i j = ℓ i 2 + ℓ j 2 − ℓ k 2 ( { i , j , k } = { 1 , 2 , 3 } ) . S_{1}=0,\qquad S_{jj}=-2\ell_j^2,\qquad S_{ij}=\ell_i^2+\ell_j^2-\ell_k^2\quad(\{i,j,k\}=\{1,2,3\}). S 1 = 0 , S j j = − 2 ℓ j 2 , S ij = ℓ i 2 + ℓ j 2 − ℓ k 2 ({ i , j , k } = { 1 , 2 , 3 }) .
Every principal 2 × 2 2 \times2 2 × 2 minor of S S S equals
Δ = 2 ∑ i < j ℓ i 2 ℓ j 2 − ∑ i ℓ i 4 = ( ℓ 1 + ℓ 2 + ℓ 3 ) ( − ℓ 1 + ℓ 2 + ℓ 3 ) ( ℓ 1 − ℓ 2 + ℓ 3 ) ( ℓ 1 + ℓ 2 − ℓ 3 ) . \begin{aligned}
\Delta=2\sum_{i<j}\ell_i^2\ell_j^2-\sum_i\ell_i^4 \\
&=(\ell_1+\ell_2+\ell_3)(-\ell_1+\ell_2+\ell_3)(\ell_1-\ell_2+\ell_3)(\ell_1+\ell_2-\ell_3).
\end{aligned} Δ = 2 i < j ∑ ℓ i 2 ℓ j 2 − i ∑ ℓ i 4 = ( ℓ 1 + ℓ 2 + ℓ 3 ) ( − ℓ 1 + ℓ 2 + ℓ 3 ) ( ℓ 1 − ℓ 2 + ℓ 3 ) ( ℓ 1 + ℓ 2 − ℓ 3 ) .
Since det S = 0 \det S=0 det S = 0 , the principal-minor criterion gives S ⪯ 0 S\preceq0 S ⪯ 0 exactly when Δ ≥ 0 \Delta\ge0 Δ ≥ 0 . The saddle block has inertia ( + , − ) (+,-) ( + , − ) , so Schur-complement inertia proves (2)⇔ \Leftrightarrow ⇔ (3). When the triangle exists, Δ \Delta Δ is sixteen times its squared area. Choose phases satisfying ∑ j a j b j e i θ j = 0 \sum_j \sqrt{a_jb_j}e^{i\theta_j}=0 ∑ j a j b j e i θ j = 0 . The unit rows ( a j ) j (\sqrt{a_j})_j ( a j ) j and ( b j e i θ j ) j (\sqrt{b_j}e^{i\theta_j})_j ( b j e i θ j ) j are orthogonal. Complete them to a unitary U U U , assigning them rows two and three. Column norms force ∣ U 1 j ∣ 2 = 1 − a j − b j = M 1 j |U_{1j}|^2=1-a_j-b_j=M_{1j} ∣ U 1 j ∣ 2 = 1 − a j − b j = M 1 j . This is the classical unistochastic triangle construction [8 Section 3, equations (21)–(25)] ; it also covers degenerate triangles and zero entries.
Finally Cauchy–Binet gives coefficient ∣ U i j ∣ 2 |U_{ij}|^2 ∣ U ij ∣ 2 for x [ 3 ] ∖ { i } y j x^{[3]\setminus\{i\}}y_j x [ 3 ] ∖ { i } y j . Complementary minors of a unitary give the same coefficient for x i y [ 3 ] ∖ { j } x_i y^{[3]\setminus\{j\}} x i y [ 3 ] ∖ { j } . The pure coefficients are one, proving (4)⇒(5). The converse follows from the first coefficients. □
Real orthogonal representations correspond exactly to degenerate triangles. In rank two, the matrix
( t 1 − t 1 − t t ) \begin{pmatrix} t & 1-t \\ 1-t & t \end{pmatrix} ( t 1 − t 1 − t t )
is realized by a rotation; rank one is immediate. The determinantal classification therefore holds through rank three.
For σ ∈ S m \sigma\in S_m σ ∈ S m , put P σ ( x , y ) = ∏ i ( x i + y σ ( i ) ) P_\sigma(x,y)=\prod_i(x_i+y_{\sigma(i)}) P σ ( x , y ) = ∏ i ( x i + y σ ( i ) ) . Through rank three, every nonnegative balanced polynomial is a convex combination of these matching products: its coefficient matrix is doubly stochastic, Birkhoff’s theorem expresses that matrix as a convex combination of permutation matrices, and the polynomial depends affinely on it [8 Section 2] . Brändén obtained obstructions to determinantal representability from matroid representability [12 ] . For the symmetric family below, two-by-two minors force an impossible Gram matrix.
Proposition 14.10 (Two incomparable models from rank four). For every m ≥ 4 m \ge4 m ≥ 4 , connected real stable sharp laws on U m , 2 m U_{m,2m} U m , 2 m that admit positive matching decompositions and those that admit homogeneous positive-semidefinite determinantal representations form incomparable classes. In particular, the determinantal classification ends exactly at rank three.
Proof. One direction has the explicit witness
G s y m ( x , y ) = 1 m ! per ( [ x i + y j ] i , j = 1 m ) = ∑ k = 0 m e m − k ( x ) e k ( y ) ( m k ) . G_{\mathrm{sym}}(x,y)=\frac{1}{m!}\operatorname{per}\left([x_i+y_j]_{i,j=1}^{m}\right)=\sum_{k=0}^{m}\frac{e_{m-k}(x)e_k(y)}{\binom{m}{k}}. G sym ( x , y ) = m ! 1 per ( [ x i + y j ] i , j = 1 m ) = k = 0 ∑ m ( k m ) e m − k ( x ) e k ( y ) .
Here per \operatorname{per} per denotes the permanent, equal to ∑ σ ∈ S m P σ ( x , y ) \sum_{\sigma\in S_m}P_\sigma(x,y) ∑ σ ∈ S m P σ ( x , y ) . Counting permutations sending a fixed k k k -subset to another gives the second identity. This polynomial is the separate polarization of ( x + y ) m (x+y)^m ( x + y ) m , so it is real stable [11 Section 2.2, Proposition 2.4] . It has every degree-m m m squarefree monomial with positive coefficient and translation ( 1 , − 1 ) (1,-1) ( 1 , − 1 ) , hence gives connected sharp laws. To sample G s y m ( p x , ( 1 − p ) y ) G_{\mathrm{sym}}(px,(1-p)y) G sym ( p x , ( 1 − p ) y ) , draw K ∼ Binomial ( m , p ) K\sim\operatorname{Binomial}(m,p) K ∼ Binomial ( m , p ) , then independently choose uniform subsets of sizes K K K and m − K m-K m − K from the two halves, conditional on K K K .
A two-basis representation G s y m = det ( diag ( x ) + U diag ( y ) U ∗ ) G_{\mathrm{sym}}=\det(\operatorname{diag}(x)+U\operatorname{diag}(y)U^*) G sym = det ( diag ( x ) + U diag ( y ) U ∗ ) forces ∣ U i j ∣ 2 = 1 / m |U_{ij}|^2=1/m ∣ U ij ∣ 2 = 1/ m . For the unit phase ratios z 1 , … , z m z_1,\ldots,z_m z 1 , … , z m of two rows, the two-by-two minor coefficients give
2 − 2 Re ( z j z k ‾ ) m 2 = ( m 2 ) − 1 , Re ( z j z k ‾ ) = − 1 m − 1 ( j ≠ k ) . \frac{2-2\operatorname{Re}(z_j\overline{z_k})}{m^2}=\binom{m}{2}^{-1},\qquad\operatorname{Re}(z_j\overline{z_k})=-\frac{1}{m-1}\quad(j\ne k). m 2 2 − 2 Re ( z j z k ) = ( 2 m ) − 1 , Re ( z j z k ) = − m − 1 1 ( j = k ) .
Thus the z j z_j z j would form a regular ( m − 1 ) (m-1) ( m − 1 ) -simplex in R 2 \mathbb{R}^2 R 2 , with Gram matrix of rank m − 1 m-1 m − 1 . This forces m ≤ 3 m\le3 m ≤ 3 .
This also rules out every homogeneous PSD determinant representation G = det ( ∑ i x i A i + ∑ j y j B j ) G=\det(\sum_i x_iA_i+\sum_j y_jB_j) G = det ( ∑ i x i A i + ∑ j y j B j ) , including a positive scalar prefactor. Indeed, a nonzero such determinant has size m m m . At the positive matrix P = ∑ i A i + ∑ j B j P=\sum_i A_i+\sum_j B_j P = ∑ i A i + ∑ j B j , multiaffinity forces each A i , B j A_i,B_j A i , B j to have rank one, since det ( P + ( t − 1 ) A i ) \det(P+(t-1)A_i) det ( P + ( t − 1 ) A i ) has degree rank A i \operatorname{rank}A_i rank A i . Translation makes det ( P + t ( ∑ i A i − ∑ j B j ) ) \det(P+t(\sum_i A_i-\sum_j B_j)) det ( P + t ( ∑ i A i − ∑ j B j )) constant. The normalized difference is Hermitian and all its eigenvalues vanish, so ∑ i A i − ∑ j B j = : R > 0 \sum_i A_i-\sum_j B_j=:R>0 ∑ i A i − ∑ j B j =: R > 0 . After conjugation by R − 1 / 2 R^{-1/2} R − 1/2 , each group consists of m m m rank-one matrices summing to I m I_m I m , hence is an orthonormal basis. The pure-half coefficient one fixes the scalar factor and reduces the representation to the two-basis form already excluded.
Appendix B gives the reverse obstruction in every rank m ≥ 4 m\ge4 m ≥ 4 , persisting under perturbation to uniform support. □
The exact transverse infinitesimal gap
Let f p = G ( p z A , q z A ′ ) f_p=G(pz_A,qz_{A'}) f p = G ( p z A , q z A ′ ) be a connected sharp law of rank m ≥ 2 m\ge2 m ≥ 2 , and let M i j M_{ij} M ij be the coefficient of x A ∖ { i } y j x_{A\setminus\{i\}}y_j x A ∖ { i } y j in G G G . Translation and complementation make M M M doubly stochastic. On centered L 2 ( μ p ) L^2(\mu_p) L 2 ( μ p ) define
L u = 1 2 ∑ i ∈ A ∪ A ′ E [ u ∣ X i ] , K = ∣ B ∩ A ∣ . \mathcal{L}u=\frac{1}{2}\sum_{i\in A\cup A'}\mathbb{E}[u\mid X_i],\qquad K=|B\cap A|. L u = 2 1 i ∈ A ∪ A ′ ∑ E [ u ∣ X i ] , K = ∣ B ∩ A ∣.
Proposition 14.11 (A field-independent local spectrum). Write the singular values of M M M as 1 = s 1 > s 2 ≥ ⋯ ≥ s m ≥ 0 1=s_1>s_2\ge\cdots\ge s_m\ge0 1 = s 1 > s 2 ≥ ⋯ ≥ s m ≥ 0 . The nonzero spectrum of L \mathcal{L} L , with multiplicities, is
1 , 1 + s j 2 , 1 − s j 2 ( 2 ≤ j ≤ m ) . 1,\qquad\frac{1+s_j}{2},\qquad\frac{1-s_j}{2}\qquad(2\le j\le m). 1 , 2 1 + s j , 2 1 − s j ( 2 ≤ j ≤ m ) . Its eigenvalue-one space is R ( K − m p ) \mathbb{R}(K-mp) R ( K − m p ) , and it vanishes on the orthogonal complement of the coordinate span. Thus the optimal infinitesimal half-marginal entropy coefficient on ( K − m p ) ⊥ (K-mp)^\perp ( K − m p ) ⊥ is ( 1 + s 2 ) / 2 (1+s_2)/2 ( 1 + s 2 ) /2 , with gap ( 1 − s 2 ) / 2 (1-s_2)/2 ( 1 − s 2 ) /2 , independently of p p p . For matrix perturbations F ε = σ + ε H F_\varepsilon=\sigma+\varepsilon H F ε = σ + ε H about a faithful constant state, the same coefficient holds on the transverse space
E H = 0 , E [ ( K − m p ) H ] = 0. \mathbb{E}H=0,\qquad\mathbb{E}[(K-mp)H]=0. E H = 0 , E [( K − m p ) H ] = 0. It is attained even with pointwise trace-one perturbations.
Proof. The independent halves give orthonormal families
Z i = X i − p p q ( i ∈ A ) , W j = 1 − X j − p p q ( j ∈ A ′ ) . Z_i=\frac{X_i-p}{\sqrt{pq}}\quad(i\in A),\qquad W_j=\frac{1-X_j-p}{\sqrt{pq}}\quad(j\in A'). Z i = pq X i − p ( i ∈ A ) , W j = pq 1 − X j − p ( j ∈ A ′ ) .
Translate ∂ x i ∂ y j G \partial_{x_i}\partial_{y_j}G ∂ x i ∂ y j G from ( p 1 , q 1 ) (p\mathbf{1},q\mathbf{1}) ( p 1 , q 1 ) to ( 1 , 0 ) (\mathbf{1},\mathbf{0}) ( 1 , 0 ) . There it equals 1 − M i j 1-M_{ij} 1 − M ij , so Cov ( X i , 1 − X j ) = p q M i j \operatorname{Cov}(X_i,1-X_j)=pqM_{ij} Cov ( X i , 1 − X j ) = pq M ij . Consequently the Gram matrix of ( Z , W ) (Z,W) ( Z , W ) is
( I M M T I ) . \begin{pmatrix}
I & M\\
M^{\mathsf T} & I
\end{pmatrix}. ( I M T M I ) .
The operator L \mathcal{L} L is half its frame operator; its nonzero eigenvalues are therefore half the nonzero Gram eigenvalues.
For simplicity of s 1 s_1 s 1 , put R = diag ( p I , q I ) R=\operatorname{diag}(pI,qI) R = diag ( p I , q I ) . The same covariance calculation gives
R − 1 ∇ 2 f p ( 1 ) R − 1 = 11 T − ( I M M T I ) . R^{-1}\nabla^2 f_p(\mathbf{1})R^{-1}
=
\mathbf{1}\mathbf{1}^{\mathsf T}
-
\begin{pmatrix}
I & M\\
M^{\mathsf T} & I
\end{pmatrix}. R − 1 ∇ 2 f p ( 1 ) R − 1 = 1 1 T − ( I M T M I ) .
The Hessian kernel equals the translation space by Lemma 5.1 , which is one-dimensional by Theorem 14.5 . Since m ≥ 2 m\ge2 m ≥ 2 , the constant vector has nonzero Hessian eigenvalue in this congruence; the singular-vector decomposition then makes s 1 = 1 s_1=1 s 1 = 1 simple and gives the stated spectrum. The common coordinate sum is ∑ i Z i = ∑ j W j = ( K − m p ) / p q \sum_i Z_i=\sum_j W_j=(K-mp)/\sqrt{pq} ∑ i Z i = ∑ j W j = ( K − m p ) / pq .
Finally the second entropy variation at a faithful constant σ \sigma σ is the BKM quadratic form 1 2 E Tr [ H D log ( σ ) [ H ] ] \frac{1}{2}\mathbb{E}\operatorname{Tr}[H D\log(\sigma)[H]] 2 1 E Tr [ H D log ( σ ) [ H ]] . Conditional expectations act only on the classical variable, so the same frame calculation is tensored with the matrix tangent space. Scalar perturbations, or H ( B ) = u ( B ) T H(B)=u(B)T H ( B ) = u ( B ) T with a traceless qubit matrix T T T , attain the corresponding Rayleigh quotients. □
The value ( 1 + s 2 ) / 2 (1+s_2)/2 ( 1 + s 2 ) /2 is the optimal coefficient for the second variation on the stated transverse space. On the full tangent space, the directions ( K − m p ) T (K-mp)T ( K − m p ) T give coefficient one, including under pointwise trace one when T T T is traceless.
For G s y m G_{\mathrm{sym}} G sym , M = m − 1 11 T M=m^{-1}\mathbf{1}\mathbf{1}^{\mathsf T} M = m − 1 1 1 T , so for m ≥ 2 m\ge2 m ≥ 2 the coordinate spectrum is 1 1 1 once and 1 / 2 1/2 1/2 with multiplicity 2 m − 2 2m-2 2 m − 2 . Its maximal transverse gap 1 / 2 1/2 1/2 coexists, for m ≥ 4 m\ge4 m ≥ 4 , with the absence of a homogeneous PSD determinant representation.
Exact realization of entropy profiles
The profile inequalities in § 3 are complete when the reference law, finite ground set, and encoding vary. The following construction realizes every allowed profile using only a classical register and a partition matroid.
Theorem A.1 (Exact realization of every entropy profile). Fix r ≥ 1 r\ge1 r ≥ 1 . As the finite ground set, Lorentzian reference law, and finite-dimensional encoding vary, the profiles
( I ( A 0 : Q ) , … , I ( A r : Q ) ) \bigl(I(A_0:Q),\ldots,I(A_r:Q)\bigr) ( I ( A 0 : Q ) , … , I ( A r : Q ) ) ends_mid=1
with classical marginal equal to the reference law are exactly the sequences h = ( h 0 , … , h r ) h=(h_0,\ldots,h_r) h = ( h 0 , … , h r ) satisfying
h 0 = 0 , h 1 ≥ 0 , h j + 1 − 2 h j + h j − 1 ≥ 0 ( 1 ≤ j < r ) . (109) h_0=0,\qquad h_1\ge0,\qquad h_{j+1}-2h_j+h_{j-1}\ge0\quad(1\le j<r). \tag*{(109)} h 0 = 0 , h 1 ≥ 0 , h j + 1 − 2 h j + h j − 1 ≥ 0 ( 1 ≤ j < r ) . (109)
Every such profile has an exact realization with a classical register and the uniform basis law of a partition matroid.
Proof. Necessity is Theorem 3.1 . For sufficiency, decompose h h h into ramps:
h k = ∑ j = 0 r − 1 c j ( k − j ) + , c 0 = h 1 , c j = h j + 1 − 2 h j + h j − 1 ( j ≥ 1 ) . h_k=\sum_{j=0}^{r-1}c_j(k-j)_+,\qquad c_0=h_1,\qquad c_j=h_{j+1}-2h_j+h_{j-1}\quad(j\ge1). h k = j = 0 ∑ r − 1 c j ( k − j ) + , c 0 = h 1 , c j = h j + 1 − 2 h j + h j − 1 ( j ≥ 1 ) .
Choose a prime power q ≥ r q\ge r q ≥ r with log q ≥ ∑ j c j \log q\ge\sum_j c_j log q ≥ ∑ j c j , and let X 1 , … , X r X_1,\ldots,X_r X 1 , … , X r be independent uniform elements of F q \mathbb{F}_q F q . The associated partition matroid chooses one label ( a , X a ) (a,X_a) ( a , X a ) from each of r r r classes; its normalized basis polynomial is
f ( z ) = q − r ∏ a = 1 r ( ∑ v ∈ F q z a , v ) . f(z)=q^{-r}\prod_{a=1}^{r}\left(\sum_{v\in\mathbb{F}_q}z_{a,v}\right). f ( z ) = q − r a = 1 ∏ r v ∈ F q ∑ z a , v .
For 0 ≤ j < r 0\le j<r 0 ≤ j < r , choose a linear map A j : F q r → F q r − j A_j:\mathbb{F}_q^r\to\mathbb{F}_q^{r-j} A j : F q r → F q r − j whose kernel is an [ r , j ] [r,j] [ r , j ] MDS code, and put Q j = A j X Q_j=A_jX Q j = A j X . Such codes are obtained by evaluating polynomials of degree less than j j j at r r r distinct field points; for j = 0 j=0 j = 0 use the zero code. Conditional on Q j Q_j Q j , the vector X X X is uniform on a coset of this code. Its projection onto any k k k coordinates has dimension min ( k , j ) \min(k,j) min ( k , j ) . Hence, for every S ⊆ [ r ] S\subseteq[r] S ⊆ [ r ] with ∣ S ∣ = k |S|=k ∣ S ∣ = k ,
I ( X S : Q j ) = ( k − min ( k , j ) ) log q = ( k − j ) + log q . I(X_S:Q_j)=(k-\min(k,j))\log q=(k-j)_+\log q. I ( X S : Q j ) = ( k − min ( k , j )) log q = ( k − j ) + log q .
Uniform retention reveals S S S and X S X_S X S , and S S S is independent of ( X , Q j ) (X,Q_j) ( X , Q j ) . Thus the same formula holds for I ( A k : Q j ) I(A_k:Q_j) I ( A k : Q j ) . This is the MDS ramp-secret-sharing construction [31 ] in the language of retained bases.
Choose an independent visible flag J J J with Pr ( J = j ) = c j / log q \Pr(J=j)=c_j/\log q Pr ( J = j ) = c j / log q , and assign its remaining probability to an uninformative encoding. More explicitly, on this extra flag value take Q J Q_J Q J to be constant. Since J J J is independent of the input, the chain rule gives
I ( A k : J , Q J ) = ∑ j = 0 r − 1 c j log q I ( A k : Q j ) = h k . I(A_k:J,Q_J)=\sum_{j=0}^{r-1}\frac{c_j}{\log q}I(A_k:Q_j)=h_k. I ( A k : J , Q J ) = j = 0 ∑ r − 1 log q c j I ( A k : Q j ) = h k .
All alphabets are finite; arbitrary real values of the c j c_j c j are carried by the flag probabilities. This proves exact realization. □
The theorem also identifies the cone of all matrix-entropy profiles when the classical marginal may vary: every such profile satisfies (A.1), and every point of that cone already has the fixed-marginal classical realization above. The cone is simplicial, with extreme rays ( ( k − j ) + ) k = 0 r ((k-j)_+)_{k=0}^{r} (( k − j ) + ) k = 0 r . Quantum side information realizes no additional profiles. In particular, a linear inequality
∑ k = 1 r a k H k ≥ 0 \sum_{k=1}^{r}a_kH_k\ge0 k = 1 ∑ r a k H k ≥ 0
holds for all these profiles if and only if
∑ k = j + 1 r a k ( k − j ) ≥ 0 ( 0 ≤ j < r ) . \sum_{k=j+1}^{r}a_k(k-j)\ge0\qquad(0\le j<r). k = j + 1 ∑ r a k ( k − j ) ≥ 0 ( 0 ≤ j < r ) .
This exact finite realization complements the Shannon descriptions of partition-symmetric entropy regions in [15 ] .
Corollary A.2 (Sharpness of every chord). For 0 < j < ℓ < r 0<j<\ell<r 0 < j < ℓ < r , the coefficient ( ℓ − j ) / ( r − j ) (\ell-j)/(r-j) ( ℓ − j ) / ( r − j ) in Corollary 3.2 is attained. More generally, for j > 0 j>0 j > 0 and every feasible pair H j = a ≥ 0 H_j=a\ge0 H j = a ≥ 0 , H r = b ≥ r a / j H_r=b\ge ra/j H r = b ≥ r a / j , equality in (13) is simultaneously attained at all j < ℓ < r j<\ell<r j < ℓ < r . Every coefficient
of the Green-kernel deficit formula (3.2) is separately attained by a profile with just that second difference nonzero.
Proof. Use the j j j th ramp for the first assertion. For the second, realize
h k = { a k / j , k ≤ j , a + ( b − a ) ( k − j ) / ( r − j ) , k ≥ j . h_k =
\begin{cases}
ak/j, & k \le j, \\
a + (b-a)(k-j)/(r-j), & k \ge j.
\end{cases} h k = { ak / j , a + ( b − a ) ( k − j ) / ( r − j ) , k ≤ j , k ≥ j .
This sequence is convex precisely under the stated endpoint condition. A single ramp isolates one second difference and proves the last claim.
Determinantal laws without positive matching decompositions
Here is the determinantal obstruction used in Proposition 14.10 , with the matching products P σ P_\sigma P σ defined there.
Proposition B.1 (An exact obstruction to matching mixtures). For every m ≥ 4 m \ge4 m ≥ 4 , there are connected sharp laws with uniform support U m , 2 m U_{m,2m} U m , 2 m whose balanced polynomials are not nonnegative sums of the P σ P_\sigma P σ .
Proof. Start with the rational orthogonal matrix
U = 1 45 ( 32 − 6 − 2 31 − 30 0 − 15 30 − 10 − 15 40 10 − 1 − 42 − 14 − 8 ) , U U T = I , U = \frac{1}{45}
\begin{pmatrix}
32 & -6 & -2 & 31 \\
-30 & 0 & -15 & 30 \\
-10 & -15 & 40 & 10 \\
-1 & -42 & -14 & -8
\end{pmatrix},
\qquad UU^{\mathsf{T}} = I, U = 45 1 32 − 30 − 10 − 1 − 6 0 − 15 − 42 − 2 − 15 40 − 14 31 30 10 − 8 , U U T = I ,
and put G U = det ( diag ( x ) + U diag ( y ) U T ) G_U = \det(\operatorname{diag}(x) + U\operatorname{diag}(y)U^{\mathsf{T}}) G U = det ( diag ( x ) + U diag ( y ) U T ) . If G U = ∑ σ w σ P σ G_U = \sum_\sigma w_\sigma P_\sigma G U = ∑ σ w σ P σ , w σ ≥ 0 w_\sigma\ge0 w σ ≥ 0 , the pure x x x coefficient gives ∑ σ w σ = 1 \sum_\sigma w_\sigma= 1 ∑ σ w σ = 1 . Comparing coefficients of x I c y J x_Ic y_J x I c y J gives
∣ det U I , J ∣ 2 = Pr ( σ ( I ) = J ) , ∣ I ∣ = ∣ J ∣ . \left|\det U_{I,J}\right|^2 = \Pr(\sigma(I) = J), \qquad|I| = |J|. ∣ det U I , J ∣ 2 = Pr ( σ ( I ) = J ) , ∣ I ∣ = ∣ J ∣.
Every permutation of four labels satisfies
{ σ { 1 , 3 } = { 1 , 4 } } ⊆ { σ ( 2 ) = 2 } ∪ { σ ( 3 ) = 4 } ∪ { σ { 1 , 2 } = { 3 , 4 } } . \{\sigma\{1,3\} = \{1,4\}\} \subseteq\{\sigma(2) = 2\} \cup\{\sigma(3) = 4\} \cup\{\sigma\{1,2\} = \{3,4\}\}. { σ { 1 , 3 } = { 1 , 4 }} ⊆ { σ ( 2 ) = 2 } ∪ { σ ( 3 ) = 4 } ∪ { σ { 1 , 2 } = { 3 , 4 }} .
Indeed, outside the middle event the left event forces σ ( 3 ) = 1 , σ ( 1 ) = 4 \sigma(3)=1,\sigma(1)=4 σ ( 3 ) = 1 , σ ( 1 ) = 4 ; then σ ( 2 ) \sigma(2) σ ( 2 ) is either 2 2 2 or 3 3 3 . Its required probability inequality is violated by the exact values
∣ det U 13 , 14 ∣ 2 = 196 2025 > ∣ U 22 ∣ 2 + ∣ U 34 ∣ 2 + ∣ det U 12 , 34 ∣ 2 = 0 + 100 2025 + 81 2025 . \left|\det U_{13,14}\right|^2 = \frac{196}{2025} > |U_{22}|^2 + |U_{34}|^2 + \left|\det U_{12,34}\right|^2 = 0 + \frac{100}{2025} + \frac{81}{2025}. ∣ det U 13 , 14 ∣ 2 = 2025 196 > ∣ U 22 ∣ 2 + ∣ U 34 ∣ 2 + ∣ det U 12 , 34 ∣ 2 = 0 + 2025 100 + 2025 81 .
The gap is 1 / 135 1/135 1/135 . It persists under sufficiently small orthogonal perturbations. A generic such perturbation has all square minors nonzero, so its two-basis law has uniform support U 4 , 8 U_{4,8} U 4 , 8 , and is connected. It remains a sharp determinantal law by the preceding field construction. For m > 4 m > 4 m > 4 , embed the displayed matrix as U ⊕ I m − 4 U \oplus I_{m-4} U ⊕ I m − 4 . The same permutation inclusion holds with the additional event { σ ( 2 ) > 4 } \{\sigma(2) > 4\} { σ ( 2 ) > 4 } on the right. Its determinantal probability is ∑ j = 5 m ∣ U 2 j ∣ 2 = 0 \sum_{j=5}^{m}|U_{2j}|^2 = 0 ∑ j = 5 m ∣ U 2 j ∣ 2 = 0 at this embedding, so the strict violation again persists under generic orthogonal perturbation to uniform support. □
This excludes positive matching mixtures with the prescribed common Bernoulli weights.
Acknowledgment
The author used ChatGPT for mathematical exploration and proof development, literature searches, computational experiments, and code for numerical certificates. Proofs developed with this assistance were carefully checked, adapted, and rewritten by the author, who also rewrote the exposition. The author takes full responsibility for the mathematical content, computations, and any errors.
References
Email address: eren321@gmail.com