betaFeedback

WorksPhysics and Astronomy

Positivity Improving Quantum Markov Operators

Abstract

A Markov operator on a matrix algebra, that is, a unital completely positive map, is called positivity improving if it maps every nonzero positive element to an invertible one. This is the noncommutative counterpart of a stochastic matrix with strictly positive entries. We show that positivity improvement is self-dual and equivalent to a faithful quantum Markov–Dobrushin–Doeblin minorization. We also compare the scalar trace minorization with the optimal common-lower-bound construction of the quantum Markov–Dobrushin coefficient. We then characterize positivity improvement through the Kraus space of the map: a Markov operator is positivity improving if and only if the Hilbert–Schmidt orthogonal complement of its Kraus space contains no operator of rank one. As a consequence the Kraus rank of a positivity improving Markov operator on the algebra of d×dd\times d matrices is at least 2d−12d-1, and an explicit Hankel-type channel shows that this bound is attained in every dimension. Although deciding positivity improvement is NP-hard in general, we obtain complete and effectively checkable classifications for qubits and qutrits, together with families of examples (Pauli, Weyl and Werner–Holevo channels) showing that all our estimates are optimal. Finally we show that, in infinite dimensions, positivity improvement no longer implies any contraction of the Markov operator.

Introduction

Let p=(p(x,y))x,yp = (p(x,y))_{x,y} be a stochastic matrix on a finite set SS. If every entry of pp is strictly positive, then pp sends every nonzero nonnegative vector to a strictly positive one, and the Perron–Frobenius theorem yields a unique strictly positive invariant distribution together with geometric convergence towards it. The rate of convergence is governed by the Dobrushin ergodicity coefficient

δ(p)=12max⁡x,x′∑y∣p(x,y)−p(x′,y)∣,\delta(p) = \frac{1}{2} \max_{x,x'} \sum_y \lvert p(x,y) - p(x',y) \rvert,

introduced by Dobrushin [7] in his study of Markov chains. The basic estimate behind this theory is the Markov–Dobrushin–Doeblin minorization: if p(x,⋅)≥κν(⋅)p(x,\cdot) \ge\kappa\nu(\cdot) for all xx and some probability measure ν\nu, then δ(p)≤1−κ\delta(p) \le1-\kappa; see Seneta [24]. Operators that send nonzero positive elements to strictly positive ones are called positivity improving. In infinite dimensions this notion is at the heart of the uniqueness theory of ground states in quantum field theory and statistical mechanics, through the work of Gross [14], Faris [11], and Reed–Simon [20].

In quantum probability the role of stochastic matrices is played by quantum Markov operators, i.e. unital completely positive maps on a C∗C^{*}-algebra, which are the transition operators of quantum Markov chains in the sense of Accardi [1, 2], and the transfer operators of finitely correlated (matrix product) states [9]. The noncommutative Perron–Frobenius theory of such maps was initiated by Evans and Høegh-Krohn [8], who introduced irreducibility for positive maps on C∗C^{*}-algebras, and developed further by Farenick [10]. Sanz, Pérez-García, Wolf and Cirac [23] studied primitive maps, i.e. maps some power of which is strictly positive (in our terminology, positivity improving), and proved a quantum version of Wielandt’s inequality; see also Wolf’s lecture notes [32]. On the ergodic side, Dobrushin’s coefficient was extended to positive contractions on preduals of von Neumann algebras by Mukhamedov [16]. A quantum Markov–Dobrushin inequality was established by Accardi, Lu and Souissi [26] and subsequently refined in the study of inhomogeneous quantum processes [25]. Related contraction techniques are relevant to exponential mixing and matrix product states [27]; see also the ergodic theory of quantum processes [28]. Dobrushin coefficients for Markov operators on general cones were investigated by Gaubert and Qu [12]. Contraction properties of quantum channels were also studied through the trace norm by Ruskai [22] and Raginsky [19], and through Hilbert’s projective metric by Reeb, Kastoryano and Wolf [21]. Quantum Markov chains and their ergodic properties on lattices and trees have been studied from the operator-algebraic viewpoint in [3, 17].

Two features make positivity improvement in the quantum setting much more subtle than in the classical one. First, positivity of a matrix in Md(C)M_d(\mathbb{C}) is not an entrywise property, so there is no “positive entries” criterion. Second, Gaubert and Qu [13] proved that deciding whether a Kraus map is strictly positive is NP-hard. Hence one cannot expect simple criteria in arbitrary dimension, and it becomes natural to ask for (a) structural characterizations that explain where the difficulty lies, (b) sharp necessary conditions, and (c) complete and effectively checkable classifications in the physically most relevant low dimensions.

The purpose of this paper is to answer these questions. Our main results are as follows.

  • Self-duality and the Markov–Dobrushin condition (Proposition 2.3 and Theorem 3.9). A Markov operator PP on B(H)\mathcal{B}(\mathcal{H}), dim⁡H=d<∞\dim\mathcal{H}=d<\infty, is positivity improving if and only if its predual (the corresponding quantum channel) is, and if and only if it satisfies the quantum Markov–Dobrushin condition

P(a)≥κTr⁡(σa) 1,a≥0,P(a) \ge\kappa\operatorname{Tr}(\sigma a)\,1,\qquad a \ge0,

for some κ>0\kappa>0 and some faithful state σ\sigma. The optimal constant for σ=1/d\sigma=1/d is κ=dλ(P)\kappa=d\lambda(P), where λ(P)\lambda(P) is the smallest eigenvalue of an output of a pure state.

  • Ergodic consequences (Theorem 3.12). If PP is positivity improving, then the quantum Dobrushin coefficient satisfies δ(P)≤1−dλ(P)<1\delta(P)\le1-d\lambda(P)<1, PP has a unique invariant state, this state is faithful, 11 is an algebraically simple eigenvalue, and every other eigenvalue μ\mu satisfies ∣μ∣≤1−dλ(P)|\mu|\le1-d\lambda(P). All these bounds are attained (Proposition 5.2 and Example 6.3).

  • Kraus-space characterization (Theorem 4.2). PP is positivity improving if and only if the Hilbert–Schmidt orthogonal complement K(P)⊥\mathcal{K}(P)^\perp of its Kraus space contains no rank-one operator. Equivalently, the kernel of the Choi matrix of PP is a completely entangled subspace in the sense of Parthasarathy [18]. This links the problem to the geometry of the Segre variety and explains its computational hardness.

  • Sharp Kraus-rank bound (Corollary 4.3 and Proposition 4.4). The Kraus rank of a positivity improving Markov operator on Md(C)M_d(\mathbb{C}) is at least 2d−12d-1. This bound is attained in every dimension by an explicit bistochastic “Hankel channel”.

  • Classification for d=2d = 2 and d=3d = 3 (Theorems 5.1 and 6.1). For qubits, PP is positivity improving if and only if either its Choi matrix is invertible, or its Kraus rank equals 3 and the generator of K(P)⊥\mathcal{K}(P)^\perp is invertible. For qutrits we obtain an analogous classification in terms of the adjugates of the elements of K(P)⊥\mathcal{K}(P)^\perp, which reduces the problem to the absence of common zeros of at most nine quadratic forms in at most four variables.

  • Infinite dimensions (§ 8). The Markov–Dobrushin estimate survives, but positivity improvement no longer implies any contraction: we construct a positivity improving normal channel with δ=1\delta= 1.

We emphasize that several natural guesses are false. Invertibility of all Kraus operators is neither necessary nor sufficient for positivity improvement in any dimension (Remark 2.5). Moreover, strict positivity of the images of the rank-one projections of a single orthonormal basis does not suffice (Example 3.11).

The paper is organized as follows. § 2 fixes notation and proves the basic structural properties of positivity improving maps. § 3 contains the Markov–Dobrushin theory and its ergodic consequences. § 4 is devoted to the Kraus-space and Choi-matrix characterization and to the sharp rank bound. § 6 and § 7 give the classifications in dimensions two and three, and § 8 discusses the infinite-dimensional situation.

Preliminaries and Basic Properties

Throughout, except in § 8, H\mathcal{H} is a complex Hilbert space of finite dimension d≥2d \ge2, with inner product ⟨⋅,⋅⟩\langle\cdot,\cdot\rangle linear in the second variable, and B(H)≅Md(C)\mathcal{B}(\mathcal{H}) \cong M_d(\mathbb{C}) is the algebra of all operators on H\mathcal{H}. We write a≥0a \ge0 for positive semidefinite operators and a>0a > 0 for positive invertible ones. The set of density matrices (states) is S(H)={ρ≥0:Tr⁡ρ=1}\mathcal{S}(\mathcal{H}) = \{\rho\ge0 : \operatorname{Tr}\rho= 1\}, and for a unit vector ψ\psi we put eψ:=∣ψ⟩⟨ψ∣e_\psi:= |\psi\rangle\langle\psi|. We use the trace norm ∥x∥1=Tr⁡∣x∣\|x\|_1 = \operatorname{Tr}|x|, the operator norm ∥a∥\|a\|, and the Hilbert–Schmidt inner product ⟨x,y⟩2=Tr⁡(x∗y)\langle x,y\rangle_2 = \operatorname{Tr}(x^*y). For a self-adjoint aa we denote by λmin⁡(a)\lambda_{\min}(a) and λmax⁡(a)\lambda_{\max}(a) its extreme eigenvalues and put osc⁡(a):=λmax⁡(a)−λmin⁡(a)\operatorname{osc}(a) := \lambda_{\max}(a) - \lambda_{\min}(a). We fix an orthonormal basis (fj)j=1d(f_j)_{j=1}^d of H\mathcal{H}, write Ejk=∣fj⟩⟨fk∣E_{jk} = |f_j\rangle\langle f_k|, and for ψ=∑jψjfj\psi= \sum_j \psi_j f_j we let ψ‾:=∑jψj‾fj\overline{\psi} := \sum_j \overline{\psi_j}f_j. The transpose xTx^{\mathsf{T}} is taken with respect to (fj)(f_j).

Definition 2.1. A quantum Markov operator on B(H)\mathcal{B}(\mathcal{H}) is a linear map P:B(H)→B(H)P : \mathcal{B}(\mathcal{H}) \to\mathcal{B}(\mathcal{H}) that is completely positive and unital, P(1)=1P(1) = 1. Its predual P∗:B(H)→B(H)P_* : \mathcal{B}(\mathcal{H}) \to\mathcal{B}(\mathcal{H}) is defined by

Tr⁡(P∗(x)a)=Tr⁡(xP(a)),x,a∈B(H).\operatorname{Tr}(P_*(x)a) = \operatorname{Tr}(xP(a)), \qquad x,a \in\mathcal{B}(\mathcal{H}).

Thus PP describes the evolution of observables (Heisenberg picture), and P∗P_* is a completely positive trace-preserving map, i.e. a quantum channel (Schrödinger picture). By the theorems of Stinespring [29], Kraus [15] and Choi [5], there are K1,…,Kr∈B(H)K_1,\ldots,K_r \in\mathcal{B}(\mathcal{H}) such that

P∗(x)=∑i=1rKixKi∗,P(a)=∑i=1rKi∗aKi,∑i=1rKi∗Ki=1.(1)P_*(x) = \sum_{i=1}^{r} K_i x K_i^*, \qquad P(a) = \sum_{i=1}^{r} K_i^*aK_i, \qquad\sum_{i=1}^{r} K_i^*K_i = 1. \tag*{(1)}

The Kraus space K(P):=span⁡{K1,…,Kr}⊂B(H)\mathcal{K}(P) := \operatorname{span}\{K_1,\ldots,K_r\} \subset\mathcal{B}(\mathcal{H}) does not depend on the chosen Kraus decomposition, since two decompositions are related by an isometric matrix [5, 32]. Its dimension is the Kraus rank r(P)r(P), which is the minimal number of Kraus operators. The Choi matrix of PP is

CP:=∑j,k=1dEjk⊗P∗(Ejk)∈B(H⊗H).(2)C_P := \sum_{j,k=1}^{d} E_{jk} \otimes P_*(E_{jk}) \in\mathcal{B}(\mathcal{H} \otimes\mathcal{H}). \tag*{(2)}

It is positive, and rank⁡CP=r(P)\operatorname{rank} C_P = r(P) (Lemma 4.1 below).

Definition 2.2. A positive linear map T:B(H)→B(H)T:\mathcal{B}(\mathcal{H}) \to\mathcal{B}(\mathcal{H}) is positivity improving if

a≥0,a≠0⟹T(a)>0,a \ge0,\quad a \ne0 \quad\Longrightarrow\quad T(a)>0,

that is, T(a)T(a) is invertible, or equivalently its support projection is 11. A Markov operator PP is positivity improving if PP is. (By Proposition 2.3 this holds if and only if P∗P_\ast is positivity improving.)

In the literature on Perron–Frobenius theory for cones, such maps are also called strictly positive [13, 23]. For a positive map TT we write T♯T^\sharp for its adjoint with respect to the trace pairing, Tr⁡(T♯(x)a)=Tr⁡(xT(a))\operatorname{Tr}(T^\sharp(x)a)=\operatorname{Tr}(xT(a)). In particular P♯=P∗P^\sharp=P_\ast and (P∗)♯=P(P_\ast)^\sharp=P.

Proposition 2.3 (Pure-state criterion and self-duality). For a positive linear map TT on B(H)\mathcal{B}(\mathcal{H}) the following are equivalent:

  • (i) TT is positivity improving;

  • (ii) T(eψ)>0T(e_\psi)>0 for every unit vector ψ∈H\psi\in\mathcal{H};

  • (iii) ⟨φ,T(eψ)φ⟩>0\langle\varphi,T(e_\psi)\varphi\rangle>0 for all unit vectors φ,ψ∈H\varphi,\psi\in\mathcal{H};

  • (iv) T♯T^\sharp is positivity improving.

In particular, a Markov operator PP is positivity improving if and only if the quantum channel P∗P_\ast is positivity improving.

Proof. (i)⇒\Rightarrow(ii) is obvious. (ii)⇔\Leftrightarrow(iii) holds because a positive semidefinite operator bb is invertible if and only if ⟨φ,bφ⟩>0\langle\varphi,b\varphi\rangle>0 for all φ≠0\varphi\ne0. (iii)⇒\Rightarrow(i): let a≥0a\ge0, a≠0a\ne0, and let ψ\psi be a unit eigenvector of aa for the eigenvalue ∥a∥>0\lVert a\rVert>0. Then a≥∥a∥eψa\ge\lVert a\rVert e_\psi, so positivity of TT gives T(a)≥∥a∥T(eψ)T(a)\ge\lVert a\rVert T(e_\psi). Since (iii) implies (ii), the right-hand side is invertible, hence so is T(a)T(a).

(iii)⇔\Leftrightarrow(iv): for unit vectors φ,ψ\varphi,\psi,

⟨φ,T(eψ)φ⟩=Tr⁡(eφT(eψ))=Tr⁡(T♯(eφ)eψ)=⟨ψ,T♯(eφ)ψ⟩.\langle\varphi,T(e_\psi)\varphi\rangle =\operatorname{Tr}(e_\varphi T(e_\psi)) =\operatorname{Tr}(T^\sharp(e_\varphi)e_\psi) =\langle\psi,T^\sharp(e_\varphi)\psi\rangle.

Hence condition (iii) for TT is the same as condition (iii) for T♯T^\sharp, with the roles of φ\varphi and ψ\psi exchanged. Applying (i)⇔\Leftrightarrow(iii) to T♯T^\sharp concludes the proof. □\square

Proposition 2.4 (Structural properties). Let PP be a positivity improving Markov operator on B(H)\mathcal{B}(\mathcal{H}).

  • (i) For every Markov operator QQ on B(H)\mathcal{B}(\mathcal{H}), the composition Q∘PQ\circ P is positivity improving. In particular, PnP^n is positivity improving for all n≥1n\ge1.

  • (ii) For every Markov operator QQ and every t∈[0,1)t\in[0,1), the operator (1−t)P+tQ(1-t)P+tQ is positivity improving. In particular, the positivity improving Markov operators form a convex set.

  • (iii) P∗P_\ast is irreducible in the sense of Evans and Høegh-Krohn [8]: there is no projection p∉{0,1}p\notin\{0,1\} with P∗(pB(H)p)⊆pB(H)pP_\ast(p\mathcal{B}(\mathcal{H})p)\subseteq p\mathcal{B}(\mathcal{H})p. Moreover P∗P_\ast is primitive in the sense of [23].

  • (iv) The set of positivity improving Markov operators is relatively open in the set of all Markov operators on B(H)\mathcal{B}(\mathcal{H}).

Proof. (i) If a≥0a\ge0, a≠0a\ne0, then P(a)≥c1P(a)\ge c1 for some c>0c>0. Since QQ is positive and unital, Q(P(a))≥cQ(1)=c1Q(P(a))\ge cQ(1)=c1. (ii) (1−t)P(a)+tQ(a)≥(1−t)P(a)>0(1-t)P(a)+tQ(a)\ge(1-t)P(a)>0. (iii) If p≠0p\ne0 is a projection with P∗(pB(H)p)⊆pB(H)pP_\ast(p\mathcal{B}(\mathcal{H})p)\subseteq p\mathcal{B}(\mathcal{H})p, then P∗(p)≤∥P∗(p)∥pP_\ast(p)\le\lVert P_\ast(p)\rVert p. Since P∗(p)P_\ast(p) is invertible by Proposition 2.3, this forces p=1p=1. Primitivity, i.e. positivity improvement of some power of P∗P_\ast, holds trivially with the first power. (iv) The map ψ↦λmin⁡(P(eψ))\psi\mapsto\lambda_{\min}(P(e_\psi)) depends continuously on the pair (P,ψ)(P,\psi) by Weyl’s inequality ∣λmin⁡(b)−λmin⁡(c)∣≤∥b−c∥|\lambda_{\min}(b)-\lambda_{\min}(c)|\le\lVert b-c\rVert. By Proposition 2.3 and compactness of the unit sphere, PP is positivity improving if and only if

λ(P):=min⁡∥ψ∥=1λmin⁡(P∗(eψ))>0.\lambda(P):=\min_{\lVert\psi\rVert=1}\lambda_{\min}(P_\ast(e_\psi))>0.

This quantity depends continuously on PP. □\square

Remark 2.5. Invertibility of all Kraus operators is neither sufficient nor necessary for positivity improvement, in every dimension d≥2d \ge2. If UU is unitary, then P(a)=U∗aUP(a) = U^{*}aU has the single invertible Kraus operator UU, yet P∗(eψ)=eUψP_{*}(e_{\psi}) = e_{U\psi} has rank one. Conversely, the completely depolarizing channel P∗(x)=Tr⁡(x)1/dP_{*}(x) = \operatorname{Tr}(x)1/d admits the Kraus decomposition Kjk=d−1/2EjkK_{jk} = d^{-1/2}E_{jk}, 1≤j,k≤d1 \le j,k \le d, into rank-one operators, and it is clearly positivity improving.

Example 2.6 (Classical Markov operators). Let pp be a d×dd \times d stochastic matrix and let P∗(x)=∑x′,yp(x′,y)⟨fx′,xfx′⟩EyyP_{*}(x) = \sum_{x',y}p(x',y)\langle f_{x'},xf_{x'}\rangle E_{yy} be the associated measure-and-prepare channel, with Kraus operators p(x′,y)∣fy⟩⟨fx′∣\sqrt{p(x',y)}\lvert f_y\rangle\langle f_{x'}\rvert. Then P∗(eψ)=∑y(∑x′∣ψx′∣2p(x′,y))EyyP_{*}(e_{\psi}) = \sum_y\left(\sum_{x'}\lvert\psi_{x'}\rvert^2p(x',y)\right)E_{yy}. Hence P∗P_{*} is positivity improving if and only if p(x′,y)>0p(x',y)>0 for all x′,yx',y, that is, if and only if pp is positivity improving in the classical sense.

The Quantum Markov–Dobrushin Condition

The quantum Dobrushin coefficient

Following [12, 16, 19, 22], the Dobrushin ergodicity coefficient of a Markov operator PP on B(H)\mathcal{B}(\mathcal{H}) is

δ(P):=sup⁡ρ,σ∈S(H)12∥P∗(ρ)−P∗(σ)∥1∈[0,1].(3)\delta(P) := \sup_{\rho,\sigma\in\mathcal{S}(\mathcal{H})} \frac{1}{2}\lVert P_{*}(\rho)-P_{*}(\sigma)\rVert_{1} \in[0,1]. \tag*{(3)}

For the channel of Example 2.6 restricted to diagonal states, (3) is exactly Dobrushin’s classical coefficient of pp.

Lemma 3.1. Let P,QP,Q be Markov operators on B(H)\mathcal{B}(\mathcal{H}) and let T0:={X=X∗∈B(H):Tr⁡X=0}\mathcal{T}_0 := \{X=X^{*}\in\mathcal{B}(\mathcal{H}):\operatorname{Tr}X=0\}.

(i) δ(P)=sup⁡{∥P∗(X)∥1/∥X∥1:X∈T0, X≠0}\delta(P)=\sup\{\lVert P_{*}(X)\rVert_{1}/\lVert X\rVert_{1}:X\in\mathcal{T}_0,\ X\ne0\}.

(ii) δ(P∘Q)≤δ(P)δ(Q)\delta(P\circ Q)\le\delta(P)\delta(Q).

(iii) (Heisenberg form) δ(P)=sup⁡{osc⁡(P(a)):a=a∗, osc⁡(a)≤1}\delta(P)=\sup\{\operatorname{osc}(P(a)):a=a^{*},\ \operatorname{osc}(a)\le1\}.

Proof. (i) Let X∈T0∖{0}X\in\mathcal{T}_0\setminus\{0\} have Jordan decomposition X=X+−X−X=X_{+}-X_{-}. Then Tr⁡X+=Tr⁡X−=∥X∥1/2=:t>0\operatorname{Tr}X_{+}=\operatorname{Tr}X_{-}=\lVert X\rVert_{1}/2=:t>0, and X=t(ρ−σ)X=t(\rho-\sigma) with ρ=X+/t\rho=X_{+}/t, σ=X−/t∈S(H)\sigma=X_{-}/t\in\mathcal{S}(\mathcal{H}). Hence ∥P∗(X)∥1/∥X∥1=12∥P∗(ρ)−P∗(σ)∥1\lVert P_{*}(X)\rVert_{1}/\lVert X\rVert_{1}=\frac{1}{2}\lVert P_{*}(\rho)-P_{*}(\sigma)\rVert_{1}. Conversely, ρ−σ∈T0\rho-\sigma\in\mathcal{T}_0 and ∥ρ−σ∥1≤2\lVert\rho-\sigma\rVert_{1}\le2. Together these give (i). (ii) Since (P∘Q)∗=Q∗∘P∗(P\circ Q)_{*}=Q_{*}\circ P_{*} and P∗P_{*} leaves T0\mathcal{T}_0 invariant (being Hermiticity- and trace-preserving), (ii) follows from (i). (iii) For Y∈T0Y\in\mathcal{T}_0 and a=a∗a=a^{*}, writing Y=t(ρ−σ)Y=t(\rho-\sigma) as above,

∣Tr⁡(Ya)∣=t∣Tr⁡(ρa)−Tr⁡(σa)∣≤12∥Y∥1osc⁡(a),(4)\lvert\operatorname{Tr}(Ya)\rvert=t\lvert\operatorname{Tr}(\rho a)-\operatorname{Tr}(\sigma a)\rvert\le\frac{1}{2}\lVert Y\rVert_{1}\operatorname{osc}(a), \tag*{(4)}

with equality when aa is the spectral projection of YY onto (0,∞)(0,\infty), for then osc⁡(a)≤1\operatorname{osc}(a)\le1 and Tr⁡(Ya)=Tr⁡Y+=12∥Y∥1\operatorname{Tr}(Ya)=\operatorname{Tr}Y_{+}=\frac{1}{2}\lVert Y\rVert_{1}. Since osc⁡(b)=sup⁡ρ,σ∈S(H)Tr⁡((ρ−σ)b)\operatorname{osc}(b)=\sup_{\rho,\sigma\in\mathcal{S}(\mathcal{H})}\operatorname{Tr}((\rho-\sigma)b) for b=b∗b=b^{*}, we get

sup⁡osc⁡(a)≤1osc⁡(P(a))=sup⁡osc⁡(a)≤1sup⁡ρ,σTr⁡((P∗(ρ)−P∗(σ))a)=sup⁡ρ,σ12∥P∗(ρ)−P∗(σ)∥1.\sup_{\operatorname{osc}(a)\le1}\operatorname{osc}(P(a))=\sup_{\operatorname{osc}(a)\le1}\sup_{\rho,\sigma}\operatorname{Tr}((P_{*}(\rho)-P_{*}(\sigma))a)=\sup_{\rho,\sigma}\frac{1}{2}\lVert P_{*}(\rho)-P_{*}(\sigma)\rVert_{1}.

Here Q∗Q_{*} denotes the predual of QQ. Item (iii) is the exact noncommutative analogue of the classical formula δ(p)=sup⁡{osc⁡(pf):osc⁡(f)≤1}\delta(p)=\sup\{\operatorname{osc}(pf):\operatorname{osc}(f)\le1\}; compare [12].

Optimal common lower bounds and the quantum Markov–Dobrushin inequality

The quantum Markov–Dobrushin inequality was introduced by Accardi, Lu and Souissi [26]; a maximal-trace common-lower-bound formulation was subsequently used in the study of inhomogeneous quantum processes [25]. Related mixing conditions and applications to matrix product states appear in [27]. We formulate the variational coefficient carefully, distinguishing an optimal minorization bound from the actual trace-distance contraction coefficient.

Let

P1(H):={∣ξ⟩⟨ξ∣:ξ∈H, ∥ξ∥=1}.\mathcal{P}_1(\mathcal{H}) := \{|\xi\rangle\langle\xi| : \xi\in\mathcal{H},\ \lVert\xi\rVert= 1\}.

For a quantum channel Φ:B(H)→B(H)\Phi: \mathcal{B}(\mathcal{H}) \to\mathcal{B}(\mathcal{H}), define the compact convex set

LMD(Φ):={B∈B(H):B≥0, B≤Φ(E) for every E∈P1(H)}.(5)\mathcal{L}_{\mathrm{MD}}(\Phi) := \{B \in\mathcal{B}(\mathcal{H}) : B \ge0,\ B \le\Phi(E)\ \text{for every } E \in\mathcal{P}_1(\mathcal{H})\}. \tag*{(5)}

Lemma 3.2 (Existence of maximal-trace common lower bounds). The set LMD(Φ)\mathcal{L}_{\mathrm{MD}}(\Phi) is nonempty and compact. Consequently max⁡{Tr⁡B:B∈LMD(Φ)}\max\{\operatorname{Tr}B : B \in\mathcal{L}_{\mathrm{MD}}(\Phi)\} exists and lies in [0,1][0,1].

Proof. The zero operator is admissible. The set is closed, and every admissible BB satisfies 0≤B≤Φ(E0)0 \le B \le\Phi(E_0) for any fixed rank-one projection E0E_0. It is therefore bounded and, in finite dimension, compact. Taking traces shows 0≤Tr⁡B≤Tr⁡Φ(E0)=10 \le\operatorname{Tr}B \le\operatorname{Tr}\Phi(E_0) = 1.

Definition 3.3 (Markov–Dobrushin infimum set). The Markov–Dobrushin infimum set is the set of maximal-trace common lower bounds

IMDTr(Φ):={κΦ∈B(H):0≤κΦ≤Φ(E)  ∀E∈P1(H), Tr⁡κΦ≥Tr⁡B  ∀B∈LMD(Φ)}.(6)\mathcal{I}_{\mathrm{MD}}^{\mathrm{Tr}}(\Phi) := \{\kappa_\Phi\in\mathcal{B}(\mathcal{H}) : 0 \le\kappa_\Phi\le\Phi(E)\ \ \forall E \in\mathcal{P}_1(\mathcal{H}),\ \operatorname{Tr}\kappa_\Phi\ge\operatorname{Tr}B\ \ \forall B \in\mathcal{L}_{\mathrm{MD}}(\Phi)\}. \tag*{(6)}

For any κΦ∈IMDTr(Φ)\kappa_\Phi\in\mathcal{I}_{\mathrm{MD}}^{\mathrm{Tr}}(\Phi), put

ηMD(Φ):=1−Tr⁡κΦ.(7)\eta_{\mathrm{MD}}(\Phi) := 1 - \operatorname{Tr}\kappa_\Phi. \tag*{(7)}

This number is independent of the choice of the maximizing operator.

The construction avoids assuming that the family {Φ(E):E∈P1(H)}\{\Phi(E) : E \in\mathcal{P}_1(\mathcal{H})\} has a greatest lower bound in the Löwner order. It optimizes the trace over all common positive lower bounds, not the trace-distance contraction coefficient itself.

Theorem 3.4 (Quantum Markov–Dobrushin inequality). *For all states ρ,σ∈S(H)\rho, \sigma\in\mathcal{S}(\mathcal{H}),

∥Φ(ρ)−Φ(σ)∥TV≤ηMD(Φ)∥ρ−σ∥TV,∥X∥TV:=12∥X∥1(X=X∗, Tr⁡X=0).(8)\lVert\Phi(\rho) - \Phi(\sigma)\rVert_{\mathrm{TV}} \le\eta_{\mathrm{MD}}(\Phi)\lVert\rho- \sigma\rVert_{\mathrm{TV}},\qquad\lVert X\rVert_{\mathrm{TV}} := \tfrac{1}{2}\lVert X\rVert_1\quad(X = X^*,\ \operatorname{Tr}X = 0). \tag*{(8)}

More precisely, the optimal trace-norm contraction coefficient δ(Φ):=sup⁡ρ≠σ∥Φ(ρ)−Φ(σ)∥1/∥ρ−σ∥1\delta(\Phi) := \sup_{\rho\ne\sigma}\lVert\Phi(\rho)-\Phi(\sigma)\rVert_1/\lVert\rho-\sigma\rVert_1 satisfies

δ(Φ)≤ηMD(Φ)≤1.\delta(\Phi) \le\eta_{\mathrm{MD}}(\Phi) \le1.
{#eq:9}*

Proof. Choose κΦ∈IMDTr(Φ)\kappa_\Phi\in\mathcal{I}_{\mathrm{MD}}^{\mathrm{Tr}}(\Phi) and put c=Tr⁡κΦc = \operatorname{Tr}\kappa_\Phi. Since every state is a convex combination of rank-one projections, Φ(ρ)≥κΦ\Phi(\rho) \ge\kappa_\Phi for every state ρ\rho. If c=1c = 1, positivity and equality of traces force Φ(ρ)=κΦ\Phi(\rho) = \kappa_\Phi for every state, so δ(Φ)=0\delta(\Phi) = 0. If c<1c < 1, define

R(X)=Φ(X)−Tr⁡(X)κΦ1−c.R(X) = \frac{\Phi(X) - \operatorname{Tr}(X)\kappa_\Phi}{1-c}.

For positive XX, the spectral decomposition of XX shows R(X)≥0R(X) \ge0; moreover RR preserves the trace. For traceless self-adjoint XX, Φ(X)=(1−c)R(X)\Phi(X) = (1-c)R(X), and positivity and trace preservation give ∥R(X)∥1≤∥X∥1\lVert R(X)\rVert_1 \le\lVert X\rVert_1. Thus δ(Φ)≤1−c\delta(\Phi) \le1-c.

Remark 3.5 (Optimality and its limitation). The value ηMD(Φ)\eta_{\mathrm{MD}}(\Phi) is the smallest contraction bound obtainable by subtracting a single common positive lower bound from all outputs. It need not equal δ(Φ)\delta(\Phi): for a classical measure-and-prepare channel with output distributions p1=(1/2,1/2,0)p_1=(1/2,1/2,0), p2=(1/2,0,1/2)p_2=(1/2,0,1/2), and p3=(0,1/2,1/2)p_3=(0,1/2,1/2), the only common nonnegative lower bound is zero, so ηMD=1\eta_{\mathrm{MD}}=1, whereas the classical Dobrushin coefficient is 1/21/2. In particular, calling ηMD\eta_{\mathrm{MD}} the minimal contraction factor without restricting to common-lower-bound estimates would be incorrect.

Proposition 3.6 (Comparison with scalar minorization). For Φ=P∗\Phi=P_* and λ(P)\lambda(P) as in (11),

δ(P)≤ηMD(P∗)≤1−dλ(P).\delta(P)\leq\eta_{\mathrm{MD}}(P_*)\leq1-d\lambda(P).

If PP is positivity improving, then ηMD(P∗)<1\eta_{\mathrm{MD}}(P_*)<1.

Proof. Theorem 3.4 gives the first inequality. By definition P∗(eψ)≥λ(P)1P_*(e_\psi)\geq\lambda(P)1 for every unit ψ\psi, so λ(P)1∈LMD(P∗)\lambda(P)1\in\mathcal{L}_{\mathrm{MD}}(P_*); maximality of the trace gives the second inequality. If PP is positivity improving, then λ(P)>0\lambda(P)>0, so ηMD(P∗)<1\eta_{\mathrm{MD}}(P_*)<1.

Markov–Dobrushin condition

Definition 3.7. Let κ∈(0,1]\kappa\in(0,1] and σ∈S(H)\sigma\in\mathcal{S}(\mathcal{H}). A Markov operator PP satisfies the Markov–Dobrushin condition MD(κ,σ)\mathrm{MD}(\kappa,\sigma) if

P(a)≥κTr⁡(σa)1for all a≥0,(9)P(a)\geq\kappa\operatorname{Tr}(\sigma a)1\qquad\text{for all }a\geq0, \tag*{(9)}

or, equivalently (by duality, testing (10) against states), if P∗(ρ)≥κσP_*(\rho)\geq\kappa\sigma for all ρ∈S(H)\rho\in\mathcal{S}(\mathcal{H}).

For the classical channel of Example 2.6, MD(κ,σ)\mathrm{MD}(\kappa,\sigma) with diagonal σ=∑yν(y)Eyy\sigma=\sum_y\nu(y)E_{yy} is exactly the Doeblin minorization p(x,⋅)≥κν(⋅)p(x,\cdot)\geq\kappa\nu(\cdot).

Theorem 3.8 (Markov–Dobrushin inequality). If PP satisfies MD(κ,σ)\mathrm{MD}(\kappa,\sigma), then δ(P)≤1−κ\delta(P)\leq1-\kappa.

Proof. If κ=1\kappa=1, then P∗(ρ)−σP_*(\rho)-\sigma is positive with trace zero, hence P∗(ρ)=σP_*(\rho)=\sigma for all ρ\rho, and δ(P)=0\delta(P)=0. Let κ<1\kappa<1 and define

R(x):=11−κ(P∗(x)−κTr⁡(x)σ),x∈B(H).R(x):=\frac{1}{1-\kappa}\left(P_*(x)-\kappa\operatorname{Tr}(x)\sigma\right),\qquad x\in\mathcal{B}(\mathcal{H}).

By MD(κ,σ)\mathrm{MD}(\kappa,\sigma) the map RR is positive, and it is trace preserving. A positive trace-preserving map contracts the trace norm on self-adjoint elements: ∥R(X)∥1≤Tr⁡R(X+)+Tr⁡R(X−)=∥X∥1\lVert R(X)\rVert_1\leq\operatorname{Tr}R(X_+)+\operatorname{Tr}R(X_-)=\lVert X\rVert_1. For X∈T0X\in\mathcal{T}_0 we have P∗(X)=(1−κ)R(X)P_*(X)=(1-\kappa)R(X), so ∥P∗(X)∥1≤(1−κ)∥X∥1\lVert P_*(X)\rVert_1\leq(1-\kappa)\lVert X\rVert_1, and Lemma 3.1(i) concludes the proof.

The next result, a corrected and sharpened form of the “strict positivity on rank-one projections” principle, identifies positivity improvement with the Markov–Dobrushin condition. Define

λ(P):=min⁡ψ∈H∥ψ∥=1λmin⁡(P∗(eψ)).(10)\lambda(P):=\min_{\substack{\psi\in\mathcal{H}\\\lVert\psi\rVert=1}}\lambda_{\min}\left(P_*(e_\psi)\right). \tag*{(10)}

The minimum exists by continuity and compactness (see the proof of Proposition 2.4(iv)), and 0≤dλ(P)≤Tr⁡P∗(eψ)=10\leq d\lambda(P)\leq\operatorname{Tr}P_*(e_\psi)=1.

Theorem 3.9 (Positivity improvement = Markov–Dobrushin). For a Markov operator PP on B(H)\mathcal{B}(\mathcal{H}) the following are equivalent:

(i) PP is positivity improving;

(ii) PP satisfies MD⁡(κ,σ)\operatorname{MD}(\kappa,\sigma) for some κ>0\kappa> 0 and some faithful state σ\sigma;

(iii) λ(P)>0\lambda(P) > 0.

In this case PP satisfies MD⁡(dλ(P),1/d)\operatorname{MD}(d\lambda(P),1/d), and dλ(P)d\lambda(P) is the largest constant κ\kappa for which MD⁡(κ,1/d)\operatorname{MD}(\kappa,1/d) holds.

Proof. (i)⇒\Rightarrow(iii) follows from Proposition 2.3 and the attainment of the minimum in (11). (iii)⇒\Rightarrow(ii): every ρ∈S(H)\rho\in\mathcal{S}(\mathcal{H}) has a spectral decomposition ρ=∑kpkeψk\rho= \sum_k p_k e_{\psi_k} with pk≥0p_k \ge0, ∑kpk=1\sum_k p_k = 1. Hence P∗(ρ)≥∑kpkλ(P)1=dλ(P)⋅1/dP_*(\rho) \ge\sum_k p_k\lambda(P)1 = d\lambda(P)\cdot1/d, which is MD⁡(dλ(P),1/d)\operatorname{MD}(d\lambda(P),1/d). (ii)⇒\Rightarrow(i): P∗(ρ)≥κσ>0P_*(\rho) \ge\kappa\sigma> 0 for every state ρ\rho, so P∗P_* is positivity improving, and so is PP by Proposition 2.3. Finally, if MD⁡(κ,1/d)\operatorname{MD}(\kappa,1/d) holds, then taking ρ=eψ\rho= e_\psi gives λmin⁡(P∗(eψ))≥κ/d\lambda_{\min}(P_*(e_\psi)) \ge\kappa/d for all ψ\psi, i.e. κ≤dλ(P)\kappa\le d\lambda(P).

The following lemma states precisely what a positivity condition on a finite family of rank-one projections yields.

Lemma 3.10. Let FF be a finite family of unit vectors such that P∗(eψ)>0P_*(e_\psi) > 0 for all ψ∈F\psi\in F. Then λF:=min⁡ψ∈Fλmin⁡(P∗(eψ))>0\lambda_F := \min_{\psi\in F}\lambda_{\min}(P_*(e_\psi)) > 0, and P∗(ρ)≥λF1P_*(\rho) \ge\lambda_F1 for every ρ\rho in the convex hull of {eψ:ψ∈F}\{e_\psi: \psi\in F\}.

Proof. Each λmin⁡(P∗(eψ))\lambda_{\min}(P_*(e_\psi)) is strictly positive and FF is finite. The second assertion follows by convexity.

The convex hull in Lemma 3.10 is a proper subset of S(H)\mathcal{S}(\mathcal{H}). The next example shows that the hypothesis of Lemma 3.10, even for an orthonormal basis, does not imply positivity improvement. Hence it does not imply δ(P)<1\delta(P) < 1 via Theorem 3.8.

Example 3.11. On C2\mathbb{C}^2 let ∣±⟩=(f1±f2)/2|\pm\rangle= (f_1 \pm f_2)/\sqrt{2} and let P∗(x)=⟨+∣x∣+⟩121+⟨−∣x∣−⟩ef1P_*(x) = \langle+|x|+\rangle\frac{1}{2}1 + \langle-|x|-\rangle e_{f_1}, an entanglement-breaking channel. Then P∗(ef1)=P∗(ef2)=141+12ef1>0P_*(e_{f_1}) = P_*(e_{f_2}) = \frac{1}{4}1 + \frac{1}{2}e_{f_1} > 0, but P∗(e−)=ef1P_*(e_-) = e_{f_1} is not invertible.

Ergodic consequences

Theorem 3.12. Let PP be a positivity improving Markov operator on B(H)\mathcal{B}(\mathcal{H}) and put λ=λ(P)\lambda= \lambda(P). Then:

(i) δ(P)≤1−dλ<1\delta(P) \le1 - d\lambda< 1;

(ii) there is a unique ρ∗∈S(H)\rho_* \in\mathcal{S}(\mathcal{H}) with P∗(ρ∗)=ρ∗P_*(\rho_*) = \rho_*; it is faithful, and in fact ρ∗≥λ1\rho_* \ge\lambda1;

(iii) for all n≥1n \ge1, all ρ∈S(H)\rho\in\mathcal{S}(\mathcal{H}) and all a=a∗∈B(H)a = a^* \in\mathcal{B}(\mathcal{H}),

∥P∗n(ρ)−ρ∗∥1≤2(1−dλ)n,∥Pn(a)−Tr⁡(ρ∗a)1∥≤(1−dλ)nosc⁡(a);\lVert P_*^n(\rho) - \rho_*\rVert_1 \le2(1-d\lambda)^n,\qquad\lVert P^n(a) - \operatorname{Tr}(\rho_*a)1\rVert\le(1-d\lambda)^n\operatorname{osc}(a);

(iv) 11 is an algebraically simple eigenvalue of PP (equivalently of P∗P_*), and every eigenvalue μ≠1\mu\ne1 satisfies ∣μ∣≤δ(P)≤1−dλ|\mu| \le\delta(P) \le1-d\lambda. In particular, the spectral gap

γ(P):=1−max⁡{∣μ∣:μ∈spec⁡(P)∖{1}}\gamma(P) := 1-\max\{|\mu| : \mu\in\operatorname{spec}(P) \setminus\{1\}\}

satisfies γ(P)≥1−δ(P)≥dλ(P)>0\gamma(P) \ge1-\delta(P) \ge d\lambda(P) > 0.

Proof. (i) follows from Theorems 3.9 and 3.8. (ii) By (i), P∗P_* is a strict contraction of the complete metric space (S(H),∥⋅∥1)(\mathcal{S}(\mathcal{H}),\lVert\cdot\rVert_1), so it has a unique fixed point by Banach’s theorem. Moreover ρ∗=P∗(ρ∗)≥λ1\rho_* = P_*(\rho_*) \ge\lambda1 by (11) and the proof of Theorem 3.9. (iii) By Lemma 3.1,

∥P∗n(ρ)−ρ∗∥1=∥P∗n(ρ−ρ∗)∥1≤δ(P)n∥ρ−ρ∗∥1≤2δ(P)n\lVert P_*^n(\rho)-\rho_*\rVert_1=\lVert P_*^n(\rho-\rho_*)\rVert_1\le\delta(P)^n\lVert\rho-\rho_*\rVert_1\le2\delta(P)^n. For a=a∗a=a^* and ρ∈S(H)\rho\in\mathcal{S}(\mathcal{H}), by (4),

∣Tr⁡(ρ(Pn(a)−Tr⁡(ρ∗a)1))∣=∣Tr⁡((P∗n(ρ)−ρ∗)a)∣≤12∥P∗n(ρ)−ρ∗∥1osc⁡(a)≤δ(P)nosc⁡(a).\left|\operatorname{Tr}\left(\rho\left(P^n(a)-\operatorname{Tr}(\rho_*a)1\right)\right)\right|=\left|\operatorname{Tr}\left((P_*^n(\rho)-\rho_*)a\right)\right|\le\frac{1}{2}\lVert P_*^n(\rho)-\rho_*\rVert_1\operatorname{osc}(a)\le\delta(P)^n\operatorname{osc}(a).

Taking the supremum over ρ\rho gives the operator-norm estimate. (iv) PP and P∗P_* are transposes of each other with respect to the nondegenerate bilinear form (x,a)↦Tr⁡(xa)(x,a)\mapsto\operatorname{Tr}(xa), so they have the same eigenvalues with the same algebraic multiplicities. Let P∗(X)=μXP_*(X)=\mu X with X≠0X\ne0 and μ≠1\mu\ne1. Since P∗P_* preserves the trace, μTr⁡X=Tr⁡X\mu\operatorname{Tr}X=\operatorname{Tr}X, so Tr⁡X=0\operatorname{Tr}X=0. Write X=H+iKX=H+iK with H=(X+X∗)/2H=(X+X^*)/2 and K=(X−X∗)/(2i)K=(X-X^*)/(2i) in T0\mathcal{T}_0. As P∗P_* preserves Hermiticity, Lemma 3.1 gives

∣μ∣n∥X∥1=∥P∗n(X)∥1≤δ(P)n(∥H∥1+∥K∥1)≤2δ(P)n∥X∥1,|\mu|^n\lVert X\rVert_1=\lVert P_*^n(X)\rVert_1\le\delta(P)^n(\lVert H\rVert_1+\lVert K\rVert_1)\le2\delta(P)^n\lVert X\rVert_1,

whence ∣μ∣≤21/nδ(P)→δ(P)|\mu|\le2^{1/n}\delta(P)\to\delta(P). For simplicity, note that B(H)=Cρ∗⊕{X:Tr⁡X=0}\mathcal{B}(\mathcal{H})=\mathbb{C}\rho_*\oplus\{X:\operatorname{Tr}X=0\} is a decomposition into P∗P_*-invariant subspaces. By the same estimate, the spectral radius of P∗P_* restricted to the second summand is at most δ(P)<1\delta(P)<1. Hence the eigenvalue 11 has algebraic multiplicity one.

Remark 3.13. (a) Theorem 3.12 provides the precise form of the relation between positivity improvement and spectral gaps: the gap is bounded below by 1−δ(P)1-\delta(P), not by δ(P)\delta(P). All inequalities in Theorem 3.12(i) and (iv) are equalities for the Pauli channels of Proposition 5.2 and the Werner–Holevo channels of Example 6.3. (b) By the Birkhoff–Hopf theorem, positivity improving maps are also strict contractions for Hilbert’s projective metric on the cone of positive definite matrices. This yields contraction rates for a wide family of distinguishability measures; see Reeb, Kastoryano and Wolf [21]. (c) The quantity λ(P)\lambda(P) is the optimal value of a minimization of a biquadratic form over the product of two unit spheres, since λ(P)=min⁡φ,ψ⟨φ,P∗(eψ)φ⟩\lambda(P)=\min_{\varphi,\psi}\langle\varphi,P_*(e_\psi)\varphi\rangle. This is the variational reason for the computational hardness established in [13].

Kraus-Space and Choi-Matrix Characterization

Recall that a linear subspace S⊆H⊗H\mathcal{S}\subseteq\mathcal{H}\otimes\mathcal{H} is completely entangled if it contains no nonzero product vector ξ⊗η\xi\otimes\eta [4, 18].

Lemma 4.1. Let V:B(H)→H⊗HV:\mathcal{B}(\mathcal{H})\to\mathcal{H}\otimes\mathcal{H}, V(K):=∑j=1dfj⊗KfjV(K):=\sum_{j=1}^{d}f_j\otimes Kf_j. Then:

(i) VV is a unitary from (B(H),⟨⋅,⋅⟩2)(\mathcal{B}(\mathcal{H}),\langle\cdot,\cdot\rangle_2) onto H⊗H\mathcal{H}\otimes\mathcal{H}, and V(∣φ⟩⟨ψ∣)=ψ‾⊗φV(|\varphi\rangle\langle\psi|)=\overline{\psi}\otimes\varphi. In particular, VV maps the nonzero rank-one operators onto the nonzero product vectors.

(ii) CP=∑i∣V(Ki)⟩⟨V(Ki)∣C_P=\sum_i|V(K_i)\rangle\langle V(K_i)| for any Kraus family (Ki)(K_i) of PP. Hence ran⁡CP=V(K(P))\operatorname{ran}C_P=V(\mathcal{K}(P)), ker⁡CP=V(K(P)⊥)\ker C_P=V(\mathcal{K}(P)^\perp), and rank⁡CP=r(P)\operatorname{rank}C_P=r(P).

(iii) ⟨φ,P∗(eψ)φ⟩=∑i∣⟨φ,Kiψ⟩∣2=⟨ψ‾⊗φ,CP(ψ‾⊗φ)⟩\langle\varphi,P_*(e_\psi)\varphi\rangle=\sum_i|\langle\varphi,K_i\psi\rangle|^2=\langle\overline{\psi}\otimes\varphi,C_P(\overline{\psi}\otimes\varphi)\rangle for all φ,ψ∈H\varphi,\psi\in\mathcal{H}.

Proof. (i) V(Eab)=fb⊗faV(E_{ab})=f_b\otimes f_a, so VV maps an orthonormal basis onto an orthonormal basis. Moreover V(∣φ⟩⟨ψ∣)=∑j⟨ψ,fj⟩fj⊗φ=ψ‾⊗φV(|\varphi\rangle\langle\psi|)=\sum_j\langle\psi,f_j\rangle f_j\otimes\varphi=\overline{\psi}\otimes\varphi. (ii) ∑i∣V(Ki)⟩⟨V(Ki)∣=∑j,kEjk⊗∑iKiEjkKi∗=CP\sum_i|V(K_i)\rangle\langle V(K_i)|=\sum_{j,k}E_{jk}\otimes\sum_iK_iE_{jk}K_i^*=C_P. The range of ∑i∣vi⟩⟨vi∣\sum_i|v_i\rangle\langle v_i| is span⁡{vi}\operatorname{span}\{v_i\}, and VV is unitary. (iii) The first equality follows from (1). For the second, ⟨ψ‾⊗φ,V(Ki)⟩=∑j⟨ψ‾,fj⟩⟨φ,Kifj⟩=⟨φ,Kiψ⟩\langle\overline{\psi}\otimes\varphi,V(K_i)\rangle=\sum_j\langle\overline{\psi},f_j\rangle\langle\varphi,K_if_j\rangle=\langle\varphi,K_i\psi\rangle, and we sum over ii using (ii).

Theorem 4.2 (Kraus-space characterization). Let PP be a Markov operator on B(H)\mathcal{B}(\mathcal{H}) with Kraus family (Ki)i=1r(K_i)_{i=1}^r. The following are equivalent:

(i) PP is positivity improving;

(ii) for every ψ≠0\psi\ne0 the vectors K1ψ,…,KrψK_{1}\psi,\ldots,K_{r}\psi span H\mathcal{H}; equivalently, the bilinear system ⟨φ,Kiψ⟩=0\langle\varphi,K_{i}\psi\rangle=0, i=1,…,ri=1,\ldots,r, has no solution with φ≠0≠ψ\varphi\ne0\ne\psi;

(iii) the Hilbert–Schmidt orthogonal complement K(P)⊥\mathcal{K}(P)^{\perp} contains no operator of rank one;

(iv) ker⁡CP\ker C_{P} is a completely entangled subspace of H⊗H\mathcal{H}\otimes\mathcal{H}.

In particular, positivity improvement depends only on the Kraus space K(P)\mathcal{K}(P). If PP and QQ are Markov operators with K(P)⊆K(Q)\mathcal{K}(P)\subseteq\mathcal{K}(Q) and PP is positivity improving, then so is QQ.

Proof. By Proposition 2.3 and Lemma 4.1(iii), PP is positivity improving if and only if ∑i∣⟨φ,Kiψ⟩∣2>0\sum_i|\langle\varphi,K_i\psi\rangle|^2>0 for all φ,ψ≠0\varphi,\psi\ne0, which is (ii). The first form of (ii) is equivalent to the second because φ⊥Kiψ\varphi\perp K_i\psi for all ii exactly when φ⊥span⁡{Kiψ}\varphi\perp\operatorname{span}\{K_i\psi\}. (ii)⇔\Leftrightarrow(iii): ⟨φ,Kiψ⟩=Tr⁡(∣ψ⟩⟨φ∣Ki)=⟨∣φ⟩⟨ψ∣,Ki⟩\langle\varphi,K_i\psi\rangle=\operatorname{Tr}(|\psi\rangle\langle\varphi|K_i)=\langle|\varphi\rangle\langle\psi|,K_i\rangle. Hence ∣φ⟩⟨ψ∣∈K(P)⊥|\varphi\rangle\langle\psi|\in\mathcal{K}(P)^{\perp} if and only if ⟨φ,Kiψ⟩=0\langle\varphi,K_i\psi\rangle=0 for all ii. (iii)⇔\Leftrightarrow(iv) follows from Lemma 4.1(i),(ii). The last assertion follows from (iii), since K(Q)⊥⊆K(P)⊥\mathcal{K}(Q)^{\perp}\subseteq\mathcal{K}(P)^{\perp}.

Condition (ii) is the formulation in terms of bilinear systems used by Gaubert and Qu [13]. Condition (iv) connects positivity improvement with the theory of completely entangled subspaces initiated by Parthasarathy [18]; see also [4, 6, 30]. Geometrically, (iii) says that the projectivization of K(P)⊥\mathcal{K}(P)^{\perp} does not meet the Segre variety Pd−1×Pd−1⊂P(B(H))\mathbb{P}^{d-1}\times\mathbb{P}^{d-1}\subset\mathbb{P}(\mathcal{B}(\mathcal{H})) of rank-one operators.

Corollary 4.3 (Sharp Kraus-rank bound). Let PP be a Markov operator on Md(C)M_d(\mathbb{C}).

(i) If PP is positivity improving, then r(P)≥2d−1r(P)\ge2d-1.

(ii) If r(P)=d2r(P)=d^2, i.e. CPC_P is invertible, then PP is positivity improving.

(iii) For each r∈{2d−1,…,d2}r\in\{2d-1,\ldots,d^2\} there exist positivity improving Markov operators of Kraus rank rr.

Proof. (i) By Theorem 4.2, ker⁡CP\ker C_P is completely entangled. By Parthasarathy’s theorem [18] (see also [4, 30]), a completely entangled subspace of Cd⊗Cd\mathbb{C}^{d}\otimes\mathbb{C}^{d} has dimension at most (d−1)2(d-1)^2. Hence d2−r(P)≤(d−1)2d^2-r(P)\le(d-1)^2. (ii) Here K(P)⊥={0}\mathcal{K}(P)^{\perp}=\{0\}. (iii) By Proposition 4.4 below there is a positivity improving HdH_d with r(Hd)=2d−1r(H_d)=2d-1. Let Wab=XaZbW_{ab}=X^aZ^b (a,b∈Zda,b\in\mathbb{Z}_d) be the Weyl unitaries, which form an orthogonal basis of Md(C)M_d(\mathbb{C}). Given rr with 2d−1≤r≤d22d-1\le r\le d^2, choose Wa1b1,…,WambmW_{a_1b_1},\ldots,W_{a_mb_m}, m=r−(2d−1)m=r-(2d-1), such that K(Hd)+span⁡{Waℓbℓ}\mathcal{K}(H_d)+\operatorname{span}\{W_{a_\ell b_\ell}\} has dimension rr. This is possible because the WabW_{ab} span Md(C)M_d(\mathbb{C}). Then 12Hd+12m∑ℓWaℓbℓ∗(⋅)Waℓbℓ\frac{1}{2}H_d+\frac{1}{2m}\sum_\ell W_{a_\ell b_\ell}^{*}(\mathord{\cdot})W_{a_\ell b_\ell} is a Markov operator with Kraus space of dimension rr containing K(Hd)\mathcal{K}(H_d), hence positivity improving by Theorem 4.2.

Proposition 4.4 (Hankel channel)

For k=2,…,2dk=2,\ldots,2d let Jk:=∑i+j=kEijJ_k:=\sum_{i+j=k}E_{ij} (anti-diagonal partial permutations, indices 1≤i,j≤d1\le i,j\le d), and define

Hd(a):=1d∑k=22dJkaJk,a∈Md(C).H_d(a):=\frac{1}{d}\sum_{k=2}^{2d}J_k a J_k,\qquad a\in M_d(\mathbb{C}).

Then HdH_d is a bistochastic Markov operator (Hd(H_d and (Hd)∗=Hd(H_d)_*=H_d are both unital). Its Kraus space is the space of Hankel matrices, its Kraus rank is 2d−12d-1, and it is positivity improving. Hence the bound of Corollary 4.3(i) is attained in every dimension.

Proof. Each JkJ_k is real symmetric, so Jk∗=JkJ_k^*=J_k. Moreover Jk∗Jk=∑j:1≤k−j≤dEjjJ_k^*J_k=\sum_{j:1\le k-j\le d}E_{jj}. For fixed jj, the indices kk with 1≤k−j≤d1\le k-j\le d are exactly k=j+1,…,j+dk=j+1,\ldots,j+d, so 1d∑kJk∗Jk=1\frac{1}{d}\sum_kJ_k^*J_k=1. Since JkJk∗=Jk∗JkJ_kJ_k^*=J_k^*J_k, also (Hd)∗(1)=1(H_d)_*(1)=1. The JkJ_k have disjoint supports, so they are linearly independent and span the Hankel matrices; thus r(Hd)=2d−1r(H_d)=2d-1. For positivity improvement we use Theorem 4.2(ii). Identify φ=∑iφifi\varphi=\sum_i\varphi_i f_i and ψ=∑jψjfj\psi=\sum_j\psi_j f_j with the polynomials φ~(t)=∑iφi‾ti\widetilde{\varphi}(t)=\sum_i\overline{\varphi_i}t^i and ψ~(t)=∑jψjtj\widetilde{\psi}(t)=\sum_j\psi_jt^j. Then

⟨φ,Jkψ⟩=∑i+j=kφi‾ψj\langle\varphi,J_k\psi\rangle=\sum_{i+j=k}\overline{\varphi_i}\psi_j

is the coefficient of tkt^k in φ~ψ~\widetilde{\varphi}\widetilde{\psi}. If all these coefficients vanished, then φ~ψ~=0\widetilde{\varphi}\widetilde{\psi}=0 in C[t]\mathbb{C}[t], and since C[t]\mathbb{C}[t] is an integral domain, φ=0\varphi=0 or ψ=0\psi=0.

In the language of [4, 18], V(K(Hd)⊥)V(\mathcal{K}(H_d)^\perp) is a completely entangled subspace of the maximal dimension (d−1)2(d-1)^2. It consists of the matrices all of whose anti-diagonal sums vanish.

Classification in Dimension Two

Theorem 5.1 (Qubits). Let PP be a Markov operator on M2(C)M_2(\mathbb{C}) with Kraus rank r=r(P)r=r(P).

(i) If r≤2r\leq2, then PP is not positivity improving.

(ii) If r=4r=4, then PP is positivity improving.

(iii) If r=3r=3, let K(P)⊥=CZ\mathcal{K}(P)^\perp=\mathbb{C}Z with Z≠0Z\neq0. Then PP is positivity improving if and only if det⁡Z≠0\det Z\neq0, equivalently, if and only if ker⁡CP\ker C_P is spanned by an entangled vector (of Schmidt rank two).

Proof. A nonzero 2×22\times2 matrix has rank one if and only if it is singular. Hence, by Theorem 4.2, PP is positivity improving if and only if every nonzero Z∈K(P)⊥Z\in\mathcal{K}(P)^\perp is invertible. (i) If r≤2r\leq2, then dim⁡K(P)⊥≥2\dim\mathcal{K}(P)^\perp\geq2. Choose linearly independent Z1,Z2∈K(P)⊥Z_1,Z_2\in\mathcal{K}(P)^\perp. The function q(s,t)=det⁡(sZ1+tZ2)q(s,t)=\det(sZ_1+tZ_2) is a binary quadratic form (possibly degenerate or identically zero), so it has a zero (s,t)≠(0,0)(s,t)\neq(0,0) over C\mathbb{C}. Then sZ1+tZ2sZ_1+tZ_2 is a nonzero singular element of K(P)⊥\mathcal{K}(P)^\perp. (ii) Here K(P)⊥={0}\mathcal{K}(P)^\perp=\{0\}. (iii) The nonzero elements of K(P)⊥\mathcal{K}(P)^\perp are the multiples cZcZ, c≠0c\neq0, which are invertible if and only if ZZ is. The Schmidt rank of V(Z)V(Z) equals rank⁡Z\operatorname{rank}Z.

Thus, for qubits, the minimal Kraus rank 3=2d−13=2d-1 is necessary, but it is sufficient only under the nondegeneracy condition det⁡Z≠0\det Z\neq0. The following family shows that all cases occur and that the estimates of Theorem 3.12 are optimal.

Proposition 5.2 (Pauli channels). Let σ0=1\sigma_0=1, σ1,σ2,σ3\sigma_1,\sigma_2,\sigma_3 be the Pauli matrices and let P∗(x)=∑j=03pjσjxσjP_*(x)=\sum_{j=0}^{3}p_j\sigma_jx\sigma_j with pj≥0p_j\geq0, ∑jpj=1\sum_jp_j=1. For {j,k,l}={1,2,3}\{j,k,l\}=\{1,2,3\} put sj:=p0+pj−pk−pls_j:=p_0+p_j-p_k-p_l and s:=max⁡j∣sj∣s:=\max_j|s_j|. Then:

(i) PP is positivity improving if and only if at least three of the pjp_j are strictly positive, if and only if s<1s<1;

(ii) λ(P)=(1−s)/2\lambda(P)=(1-s)/2, δ(P)=s\delta(P)=s, and spec⁡(P∗)={1,s1,s2,s3}\operatorname{spec}(P_*)=\{1,s_1,s_2,s_3\};

(iii) consequently, if PP is positivity improving, then δ(P)=1−dλ(P)\delta(P)=1-d\lambda(P) and γ(P)=dλ(P)\gamma(P)=d\lambda(P), i.e. equality holds in Theorem 3.12(i) and (iv).

Proof. The Pauli matrices are Hilbert–Schmidt orthogonal, so r(P)=#{j:pj>0}r(P)=\#\{j:p_j>0\}, and K(P)\mathcal{K}(P) is spanned by the σj\sigma_j with pj>0p_j>0. If exactly three pjp_j are positive, then K(P)⊥\mathcal{K}(P)^\perp is spanned by the remaining Pauli matrix, which is invertible. Theorem 5.1 gives the first equivalence in (i). Since σkσjσk=σj\sigma_k\sigma_j\sigma_k=\sigma_j if k∈{0,j}k\in\{0,j\} and σkσjσk=−σj\sigma_k\sigma_j\sigma_k=-\sigma_j otherwise, P∗(σj)=sjσjP_*(\sigma_j)=s_j\sigma_j for j=1,2,3j=1,2,3, and P∗(1)=1P_*(1)=1. This gives the spectrum. Writing a state as ρ=12(1+n⋅σ)\rho=\frac{1}{2}(1+n\cdot\sigma) with ∣n∣≤1|n|\leq1, we get P∗(ρ)=12(1+(Λn)⋅σ)P_*(\rho)=\frac{1}{2}(1+(\Lambda n)\cdot\sigma) with Λ=diag⁡(s1,s2,s3)\Lambda=\operatorname{diag}(s_1,s_2,s_3). Pure states correspond to ∣n∣=1|n|=1, and λmin⁡(P∗(eψ))=12(1−∣Λn∣)\lambda_{\min}(P_*(e_\psi))=\frac{1}{2}(1-|\Lambda n|), whose minimum over the unit sphere is 12(1−s)\frac{1}{2}(1-s). Next, 12∥P∗(ρ)−P∗(ρ′)∥1=12∣Λ(n−n′)∣≤s\frac{1}{2}\lVert P_*(\rho)-P_*(\rho')\rVert_1=\frac{1}{2}\lvert\Lambda(n-n')\rvert\le s, with equality for n=−n′n=-n' a unit vector along a direction where ∣sj∣=s\lvert s_j\rvert=s. Hence δ(P)=s\delta(P)=s, and γ(P)=1−s\gamma(P)=1-s when s<1s<1. Finally, ∣sj∣=1\lvert s_j\rvert=1 means pk=pl=0p_k=p_l=0 (if sj=1s_j=1) or p0=pj=0p_0=p_j=0 (if sj=−1s_j=-1). So s<1s<1 if and only if no two of the pjp_j vanish, which completes (i). Part (iii) follows from (ii) with d=2d=2.

Example 5.3 (Depolarizing channel). P∗(x)=(1−p)x+pTr⁡(x)12P_*(x)=(1-p)x+p\operatorname{Tr}(x)\frac{1}{2} is completely positive if and only if 0≤p≤430\le p\le\frac{4}{3}. It is the Pauli channel with p0=1−3p4p_0=1-\frac{3p}{4}, p1=p2=p3=p4p_1=p_2=p_3=\frac{p}{4}. By Proposition 5.2 it is positivity improving if and only if 0<p≤430<p\le\frac{4}{3}; then λ(P)=min⁡{p/2,1−p/2}\lambda(P)=\min\{p/2,1-p/2\} and δ(P)=∣1−p∣\delta(P)=\lvert1-p\rvert. For p=43p=\frac{4}{3} the Kraus rank is 33 and K(P)⊥=C1\mathcal{K}(P)^\perp=\mathbb{C}1.

Example 5.4 (Rank three without positivity improvement). Let b,c∈C∖{0}b,c\in\mathbb{C}\setminus\{0\} with ∣b∣2+∣c∣2=1\lvert b\rvert^2+\lvert c\rvert^2=1, and let

K1=∣f2⟩⟨f1∣,K2=b∣f1⟩⟨f2∣,K3=c∣f2⟩⟨f2∣.K_1=\lvert f_2\rangle\langle f_1\rvert,\qquad K_2=b\lvert f_1\rangle\langle f_2\rvert,\qquad K_3=c\lvert f_2\rangle\langle f_2\rvert.

Then ∑iKi∗Ki=1\sum_i K_i^*K_i=1, r(P)=3r(P)=3, and K(P)⊥=CE11\mathcal{K}(P)^\perp=\mathbb{C}E_{11} with det⁡E11=0\det E_{11}=0. Hence PP is not positivity improving; indeed P∗(ef1)=ef2P_*(e_{f_1})=e_{f_2}.

For d=2d=2, the Hankel channel of Proposition 4.4 is the Pauli channel with (p0,p1,p2,p3)=(14,12,0,14)(p_0,p_1,p_2,p_3)=(\frac{1}{4},\frac{1}{2},0,\frac{1}{4}). Thus λ(H2)=14\lambda(H_2)=\frac{1}{4} and δ(H2)=γ(H2)=12\delta(H_2)=\gamma(H_2)=\frac{1}{2}.

Classification in Dimension Three

For Z∈M3(C)Z\in M_3(\mathbb{C}) we use the classical fact that rank⁡Z≤1\operatorname{rank}Z\le1 if and only if all 2×22\times2 minors of ZZ vanish, i.e. if and only if adj⁡(Z)=0\operatorname{adj}(Z)=0, where adj⁡\operatorname{adj} denotes the adjugate matrix.

Theorem 6.1 (Qutrits). Let PP be a Markov operator on M3(C)M_3(\mathbb{C}) with Kraus rank r=r(P)r=r(P), and let Z1,…,ZmZ_1,\ldots,Z_m be a basis of K(P)⊥\mathcal{K}(P)^\perp, m=9−rm=9-r.

(i) If r≤4r\le4, then PP is not positivity improving.

(ii) If r=9r=9, then PP is positivity improving.

(iii) If r=8r=8, then PP is positivity improving if and only if rank⁡Z1≥2\operatorname{rank}Z_1\ge2, i.e. adj⁡(Z1)≠0\operatorname{adj}(Z_1)\ne0.

(iv) If 5≤r≤75\le r\le7, then PP is positivity improving if and only if the nine quadratic forms

qab(t1,…,tm):=[adj⁡(∑ℓ=1mtℓZℓ)]ab,1≤a,b≤3,q_{ab}(t_1,\ldots,t_m):=\left[\operatorname{adj}\left(\sum_{\ell=1}^{m}t_\ell Z_\ell\right)\right]_{ab},\qquad1\le a,b\le3,

have no common zero in Cm∖{0}\mathbb{C}^m\setminus\{0\}. For r=7r=7 (m=2m=2) this holds if and only if the binary forms qabq_{ab} are not all identically zero and have no common linear factor, i.e. their greatest common divisor is constant.

Proof. By Theorem 4.2, PP is positivity improving if and only if every nonzero Z∈K(P)⊥Z\in\mathcal{K}(P)^\perp satisfies rank⁡Z≥2\operatorname{rank}Z\ge2, i.e. adj⁡(Z)≠0\operatorname{adj}(Z)\ne0. Writing Z=∑ℓtℓZℓZ=\sum_\ell t_\ell Z_\ell with t≠0t\ne0 gives (iii) and the first part of (iv). (i) is Corollary 4.3(i) with d=3d=3, since 2d−1=52d-1=5. (ii) is Corollary 4.3(ii). For m=2m=2, each qabq_{ab} is a binary quadratic form or zero. Over C\mathbb{C}, a family of binary forms has a common zero in C2∖{0}\mathbb{C}^2\setminus\{0\} if and only if all of them vanish there, which happens if and only if they share a linear factor or all vanish identically.

Remark 6.2. (a) Theorem 4.2 shows that, in every dimension, the spanning condition “{Kiψ}\{K_i\psi\} spans H\mathcal{H} for all ψ≠0\psi\ne0” is by itself equivalent to positivity improvement. Irreducibility is then automatic by Proposition 2.4(iii), so it is not an independent condition. (b) For d=3d=3 there is no positivity improving Markov operator with three Kraus operators. Indeed, ψ↦det⁡[K1ψ K2ψ K3ψ]\psi\mapsto\det[K_1\psi\ K_2\psi\ K_3\psi] is a homogeneous cubic form on C3\mathbb{C}^3, and every such form has a nontrivial zero, at which K1ψK_1\psi, K2ψK_2\psi, K3ψK_3\psi fail to span C3\mathbb{C}^3. This is a special case of Theorem 6.1(i).

(c) The criterion of Theorem 6.1(iv) is decidable by elimination theory (resultants or Gröbner bases) in a fixed number of variables, in contrast with the NP-hardness for general dd [13].

Example 6.3 (Werner–Holevo channels). Let d≥2d \ge2 and let

W∗±(x)=Tr⁡(x)1±xTd±1W_*^{\pm}(x)=\frac{\operatorname{Tr}(x)\mathbf{1}\pm x^{T}}{d\pm1}

be the Werner–Holevo channels [31]. From ∑j,kEjk⊗Ekj=F\sum_{j,k} E_{jk}\otimes E_{kj}=F (the flip) one gets CW±=2d±1Π±CW_{\pm}=\frac{2}{d\pm1}\Pi_{\pm}, where Π+\Pi_+ and Π−\Pi_- are the projections onto the symmetric and antisymmetric subspaces. Hence W±W^{\pm} are Markov operators, and V−1(ran⁡Π±)V^{-1}(\operatorname{ran}\Pi_{\pm}) identifies K(W+)\mathcal{K}(W^+) with the symmetric matrices and K(W−)\mathcal{K}(W^-) with the antisymmetric matrices.

(a) K(W+)⊥\mathcal{K}(W^+)^\perp consists of antisymmetric matrices, which have even rank, so W+W^+ is positivity improving, with r(W+)=d(d+1)/2r(W^+)=d(d+1)/2 (=6=6 for d=3d=3). Directly, W∗+(eψ)=(1+eψˉ)/(d+1)W_*^+(e_\psi)=(1+e_{\bar{\psi}})/(d+1), so λ(W+)=1/(d+1)\lambda(W^+)=1/(d+1). Since W∗+(ρ−σ)=(ρ−σ)T/(d+1)W_*^+(\rho-\sigma)=(\rho-\sigma)^T/(d+1) and the transpose is a trace-norm isometry, δ(W+)=1/(d+1)=1−dλ(W+)\delta(W^+)=1/(d+1)=1-d\lambda(W^+). The restriction of W∗+W_*^+ to traceless matrices is x↦xT/(d+1)x\mapsto x^T/(d+1), with eigenvalues ±1/(d+1)\pm1/(d+1), so γ(W+)=d/(d+1)=dλ(W+)\gamma(W^+)=d/(d+1)=d\lambda(W^+). All bounds of Theorem 3.12 are therefore attained in every dimension.

(b) W−W^- is never positivity improving, since W∗−(eψ)ψˉ=0W_*^-(e_\psi)\bar{\psi}=0. For d=3d=3, r(W−)=3<5r(W^-)=3<5.

Example 6.4 (Weyl channels on C3\mathbb{C}^3). Let ω=e2πi/3\omega=e^{2\pi i/3}, let Xfj=fj+1Xf_j=f_{j+1} and Zfj=ωjfjZf_j=\omega^j f_j (indices mod 3), and let Wab=XaZbW_{ab}=X^aZ^b. For a probability vector (pab)(p_{ab}) put P∗(x)=∑a,bpabWabxWab∗P_*(x)=\sum_{a,b}p_{ab}W_{ab}xW_{ab}^*. Since the WabW_{ab} are orthogonal, K(P)=span⁡{Wab:pab>0}\mathcal{K}(P)=\operatorname{span}\{W_{ab}:p_{ab}>0\} and r(P)=#{(a,b):pab>0}r(P)=\#\{(a,b):p_{ab}>0\}.

  • r=9r=9: positivity improving (completely depolarizing channel for pab≡19p_{ab}\equiv\frac{1}{9}).

  • r=8r=8, pab=0p_{ab}=0 for exactly one (a,b)(a,b): K(P)⊥=CWab\mathcal{K}(P)^\perp=\mathbb{C}W_{ab} with WabW_{ab} unitary, so PP is positivity improving. For uniform weights and (a,b)=(0,0)(a,b)=(0,0), P∗(x)=18(3Tr⁡(x)1−x)P_*(x)=\frac{1}{8}(3\operatorname{Tr}(x)\mathbf{1}-x) and λ(P)=14\lambda(P)=\frac{1}{4}.

  • r=7r=7, p01=p02=0p_{01}=p_{02}=0: K(P)⊥=span⁡{Z,Z2}\mathcal{K}(P)^\perp=\operatorname{span}\{Z,Z^2\}, and sZ+tZ2=diag⁡(s+t,sω+tω2,sω2+tω)sZ+tZ^2=\operatorname{diag}(s+t,s\omega+t\omega^2,s\omega^2+t\omega). Any two of the three diagonal entries vanish simultaneously only for s=t=0s=t=0, since the corresponding 2×22\times2 coefficient matrices have determinants ω2−ω\omega^2-\omega, ω−ω2\omega-\omega^2 and ω2−ω4≠0\omega^2-\omega^4\ne0. Hence every nonzero element has rank ≥2\ge2, and PP is positivity improving.

  • r=6r=6, pab>0p_{ab}>0 if and only if a≠0a\ne0: K(P)⊥=span⁡{1,Z,Z2}\mathcal{K}(P)^\perp=\operatorname{span}\{\mathbf{1},Z,Z^2\} is the diagonal algebra, which contains E11E_{11}. So PP is not positivity improving; indeed P∗(ef1)P_*(e_{f_1}) is supported on span⁡{f2,f3}\operatorname{span}\{f_2,f_3\}. Together with Example 6.3(a), this shows that for d=3d=3 and r=6r=6 both alternatives occur.

  • r=3r=3: the dephasing channel (p0b=13)(p_{0b}=\frac{1}{3}) and the cyclic twirl (pa0=13)(p_{a0}=\frac{1}{3}) are not positivity improving. For the cyclic twirl P∗(x)=13∑aXaxX−aP_*(x)=\frac{1}{3}\sum_a X^a xX^{-a}, the Fourier vectors ψ\psi (eigenvectors of XX) satisfy P∗(eψ)=eψP_*(e_\psi)=e_\psi.

Example 6.5 (Minimal rank for qutrits). The Hankel channel H3H_3 of Proposition 4.4 has Kraus rank 5 and is positivity improving. Its Kraus operators are 13Jk\frac{1}{\sqrt{3}}J_k, k=2,…,6k=2,\ldots,6, where J2=E11J_2=E_{11}, J3=E12+E21J_3=E_{12}+E_{21}, J4=E13+E22+E31J_4=E_{13}+E_{22}+E_{31}, J5=E23+E32J_5=E_{23}+E_{32} and J6=E33J_6=E_{33}. The complement K(H3)⊥\mathcal{K}(H_3)^\perp is the four-dimensional space of matrices with vanishing anti-diagonal sums. For instance, Z=E12−E21Z=E_{12}-E_{21} and Z=E13−E22Z=E_{13}-E_{22} have rank 2, in agreement with Theorem 6.1(iv). Numerically one finds λ(H3)≈0.0370\lambda(H_3)\approx0.0370, so Theorem 3.12 yields δ(H3)≤1−3λ(H3)≈0.889\delta(H_3)\le1-3\lambda(H_3)\approx0.889.

Remarks on the Infinite-Dimensional Case

Let now H\mathcal{H} be an infinite-dimensional separable Hilbert space, let PP be a normal unital completely positive map on B(H)\mathcal{B}(\mathcal{H}), and let P∗P_{*} be its predual on the trace class T(H)\mathcal{T}(\mathcal{H}). Since positive trace-class operators cannot be boundedly invertible, the appropriate replacement for strict positivity is faithfulness (trivial kernel). We say that PP is positivity improving if P(a)P(a) is faithful for every nonzero a≥0a \ge0 in B(H)\mathcal{B}(\mathcal{H}). By trace duality and testing rank-one projections, this is equivalent to faithfulness of P∗(x)P_{*}(x) for every nonzero positive trace-class xx.

Proposition 7.1. (i) If P∗(x)=∑iKixKi∗P_{*}(x) = \sum_i K_i x K_i^{*} with a trace-norm convergent Kraus series for trace-class xx, then PP is positivity improving if and only if for all nonzero φ,ψ∈H\varphi,\psi\in\mathcal{H} there is ii with ⟨φ,Kiψ⟩≠0\langle\varphi,K_i\psi\rangle\ne0.

(ii) Theorem 3.8 holds verbatim: if P∗(ρ)≥κσP_{*}(\rho) \ge\kappa\sigma for all normal states ρ\rho, then δ(P)≤1−κ\delta(P) \le1-\kappa.

(iii) The infimum of the spectrum of every positive trace-class output in infinite dimension is zero; hence the direct spectral analogue of λ(P)\lambda(P) vanishes. Moreover, positivity improvement does not imply δ(P)<1\delta(P) < 1.

Proof. (i) For x≥0x \ge0 nonzero in T(H)\mathcal{T}(\mathcal{H}), the operator P∗(x)P_{*}(x) dominates a positive multiple of P∗(eψ)P_{*}(e_{\psi}) for an eigenvector ψ\psi of xx with eigenvalue ∥x∥\lVert x\rVert. As in Proposition 2.3, faithfulness of all P∗(x)P_{*}(x) is equivalent to ⟨φ,P∗(eψ)φ⟩=∑i∣⟨φ,Kiψ⟩∣2>0\langle\varphi,P_{*}(e_{\psi})\varphi\rangle= \sum_i \lvert\langle\varphi,K_i\psi\rangle\rvert^2 > 0 for all nonzero φ,ψ\varphi,\psi. (ii) The proof of Theorem 3.8 uses only positivity and trace preservation. (iii) P∗(eψ)P_{*}(e_{\psi}) is a compact positive operator, so λmin⁡\lambda_{\min} must be replaced by inf⁡spec⁡\inf\operatorname{spec}, which is 0. For the second assertion, let (fn)n≥1(f_n)_{n\ge1} be an orthonormal basis, let τ\tau be a faithful normal state, let 0<εn→00 < \varepsilon_n \to0, and put σn=(1−εn)efn+εnτ\sigma_n = (1-\varepsilon_n)e_{f_n} + \varepsilon_n\tau. The measure-and-prepare map

P∗(x)=∑n≥1⟨fn,xfn⟩ σnP_{*}(x) = \sum_{n\ge1} \langle f_n,x f_n\rangle\,\sigma_n

is a normal quantum channel. For ψ≠0\psi\ne0 we have P∗(eψ)≥(∑n∣⟨fn,ψ⟩∣2εn)τP_{*}(e_{\psi}) \ge\left(\sum_n \lvert\langle f_n,\psi\rangle\rvert^2\varepsilon_n\right)\tau, which is faithful because the coefficient is strictly positive. Hence PP is positivity improving. On the other hand,

12∥P∗(efn)−P∗(efm)∥1=12∥σn−σm∥1≥1−εn−εm⟶1(n≠m→∞),\frac{1}{2}\lVert P_{*}(e_{f_n})-P_{*}(e_{f_m})\rVert_1 = \frac{1}{2}\lVert\sigma_n-\sigma_m\rVert_1 \ge1-\varepsilon_n-\varepsilon_m \longrightarrow1 \qquad(n \ne m \to\infty),

so δ(P)=1\delta(P)=1.

Thus, in infinite dimensions, the equivalence of Theorem 3.9 breaks down, and the Markov–Dobrushin condition with a faithful reference state must be imposed as a genuinely stronger hypothesis. This is in line with the classical theory, where positivity improvement must be supplemented by compactness or Doeblin-type conditions to obtain uniform ergodicity [16, 24]. Infinite-dimensional transition operators of quantum Markov chains on trees and lattices [3, 17], and transfer operators of finitely correlated states [9], are natural settings in which these refined conditions can be checked. This will be pursued elsewhere.

Conclusion

We have shown that positivity improvement of a quantum Markov operator on a matrix algebra is a self-dual property that coincides with the quantum Markov–Dobrushin minimization with respect to the trace. It implies an explicit and optimal contraction of the Dobrushin ergodicity coefficient, a spectral gap γ(P)≥dλ(P)\gamma(P) \ge d\lambda(P), and uniform exponential convergence to a unique faithful invariant state. Structurally, positivity improvement is exactly the absence of rank-one operators in the orthogonal complement of the Kraus space, i.e. complete entanglement of the kernel of the Choi matrix. This yields the sharp bound r(P)≥2d−1r(P) \ge2d - 1, attained by the Hankel channel. For d=2d = 2 and d=3d = 3 we obtained complete and effectively checkable classifications, and the Pauli, Weyl and Werner–Holevo families show that all cases occur and that our estimates cannot be improved.

References

References

  1. [1]L. Accardi, The noncommutative Markovian property, Funct. Anal. Appl. 9 (1975) 1–8.DOI
  2. [2]L. Accardi and A. Frigerio, Markovian cocycles, Proc. R. Ir. Acad. 83A(2) (1983) 251–263.
  3. [3]L. Accardi, A. Souissi and E. G. Soueidy, Quantum Markov chains: A unification approach, Infin. Dimens. Anal. Quantum Probab. Relat. Top. 23(2) (2020) 2050016, https://doi.org/10.1142/S0219025720500162.
  4. [4]B. V. R. Bhat, A completely entangled subspace of maximal dimension, Int. J. Quantum Inf. 4(2) (2006) 325–330, https://doi.org/10.1142/S0219749906001797.
  5. [5]M.-D. Choi, Completely positive linear maps on complex matrices, Linear Algebra Appl. 10 (1975) 285–290, https://doi.org/10.1016/0024-3795(75)90075-0.
  6. [6]T. Cubitt, A. Montanaro and A. Winter, On the dimension of subspaces with bounded Schmidt rank, J. Math. Phys. 49 (2008) 022107.DOI
  7. [7]R. L. Dobrushin, Central limit theorem for nonstationary Markov chains I, II, Theory Probab. Appl. 1 (1956) 65–80, 329–383.DOI
  8. [8]D. E. Evans and R. Høegh-Krohn, Spectral properties of positive maps on C*-algebras, J. London Math. Soc. (2) 17 (1978) 345–355.DOI
  9. [9]M. Fannes, B. Nachtergaele and R. F. Werner, Finitely correlated states on quantum spin chains, Comm. Math. Phys. 144 (1992) 443–490, https://doi.org/10.1007/BF02099178.
  10. [10]D. R. Farenick, Irreducible positive linear maps on operator algebras, Proc. Amer. Math. Soc. 124 (1996) 3381–3390.DOI
  11. [11]W. G. Faris, Invariant cones and uniqueness of the ground state for fermion systems, J. Math. Phys. 13 (1972) 1285–1290.DOI
  12. [12]S. Gaubert and Z. Qu, Dobrushin’s ergodicity coefficient for Markov operators on cones, Integral Equations Operator Theory 81 (2015) 127–150.DOI
  13. [13]S. Gaubert and Z. Qu, Checking strict positivity of Kraus maps is NP-hard, Inform. Process. Lett. 118 (2017) 35–43.DOI
  14. [14]L. Gross, Existence and uniqueness of physical ground states, J. Funct. Anal. 10 (1972) 52–109.DOI
  15. [15]K. Kraus, States, Effects, and Operations, Lecture Notes in Physics, Vol. 190 (Springer, Berlin, 1983).
  16. [16]F. Mukhamedov, The Dobrushin ergodicity coefficient and the ergodicity of noncommutative Markov chains, J. Math. Anal. Appl. 408 (2013) 364–373.DOI
  17. [17]F. Mukhamedov and A. Souissi, Quantum Markov states on Cayley trees, J. Math. Anal. Appl. 473 (2019) 313–333.DOI
  18. [18]K. R. Parthasarathy, On the maximal dimension of a completely entangled subspace for finite level quantum systems, Proc. Indian Acad. Sci. Math. Sci. 114 (2004) 365–374.DOI
  19. [19]M. Raginsky, Strictly contractive quantum channels and physically realizable quantum computers, Phys. Rev. A 65 (2002) 032306.DOI
  20. [20]M. Reed and B. Simon, Methods of Modern Mathematical Physics IV: Analysis of Operators (Academic Press, New York, 1978).
  21. [21]D. Reeb, M. J. Kastoryano and M. M. Wolf, Hilbert’s projective metric in quantum information theory, J. Math. Phys. 52 (2011) 082201, https://doi.org/10.1063/1.3615729.
  22. [22]M. B. Ruskai, Beyond strong subadditivity? Improved bounds on the contraction of generalized relative entropy, Rev. Math. Phys. 6 (1994) 1147–1161.
  23. [23]M. Sanz, D. Pérez-García, M. M. Wolf and J. I. Cirac, A quantum version of Wielandt’s inequality, IEEE Trans. Inform. Theory 56(9) (2010) 4668–4673.DOI
  24. [24]E. Seneta, Non-negative Matrices and Markov Chains, revised printing (Springer, New York, 2006).
  25. [25]A. Souissi, Ergodic theory of inhomogeneous quantum processes, Quantum Inf. Process. 25 (2026) 220, https://doi.org/10.1007/s11128-026-05228-8.
  26. [26]L. Accardi, Y. G. Lu and A. Souissi, A Markov–Dobrushin inequality for quantum channels, Open Syst. Inf. Dyn. 28(4) (2021) 2150018, https://doi.org/10.1142/S1230161221500189.
  27. [27]A. Souissi and A. Barhoumi, An exponential mixing condition for quantum channels: Application to matrix product states, Quantum Inf. Process. 24(5) (2025) 150.DOI
  28. [28]Movassagh, R., Schenker, J.: Theory of ergodic quantum processes. Phys. Rev. X 11, 041001 (2021)DOI
  29. [29]W. F. Stinespring, Positive functions on C*-algebras, Proc. Amer. Math. Soc. 6 (1955) 211–216.
  30. [30]N. R. Wallach, An unentangled Gleason’s theorem, in Quantum Computation and Information, Contemp. Math., Vol. 305 (Amer. Math. Soc., Providence, 2002), pp. 291–298.
  31. [31]R. F. Werner and A. S. Holevo, Counterexample to an additivity conjecture for output purity of quantum channels, J. Math. Phys. 43 (2002) 4353–4357.DOI
  32. [32]M. M. Wolf, Quantum channels & operations: Guided tour, lecture notes (2012), available at the author’s web page, Technische Universität München.

Paper details

vlnc:W_2uh6hxad
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents