betaFeedback

WorksMathematics

Ergodicity, memory loss, and directional stability on von Neumann algebras

Abstract

For a deterministic sequence of normal unital completely positive maps on a von Neumann algebra, convergence of normal states, loss of memory, and convergence of Cesàro averages are different properties. This paper separates these properties in the predual norm and identifies the precise role of composition order. Uniform loss of memory for right-appended products on the predual implies uniform convergence to a normal state without any compactness or trace assumption; the convergence rate is bounded by the image diameter. In contrast, on every infinite-dimensional σσ-finite von Neumann algebra there are right-appended products with pointwise loss of memory but no convergent normal-state trajectory, even though all nontrivial state images equal a fixed compact line segment. Consequently, finite dimensionality is characterized by the universal implication from pointwise backward memory loss to mixing. An attained normal-functional minorization coefficient extends the common-lower-bound construction for matrix channels to arbitrary von Neumann algebras and yields quantitative contraction and unique normal entrance laws. Further results distinguish pointwise from uniform relaxation, Cesàro convergence from mixing, fixed-start from time-uniform contraction, and stationary from moving asymptotic states. As an application, inhomogeneous transfer states on spin chains, including matrix product states, are shown to have a unique thermodynamic limit, uniformly in the boundary condition, with an arbitrary von Neumann algebra as bond algebra; boundary dependence and correlations are controlled by the same product coefficients. All counterexamples are explicit and extend to algebras of types I, II, and III.

Introduction

The ergodic theory of quantum dynamics has developed along two largely separate lines. The first is operator algebraic. Here the dynamics is a normal positive map, an automorphism, or a semigroup on a von Neumann algebra, usually with a distinguished invariant normal state. Mean and individual ergodic theorems were obtained by Lance for automorphisms and convex sets [23], and by Yeadon for semifinite algebras in the noncommutative L1L^1 setting [37]. Junge and Xu later proved maximal ergodic inequalities in noncommutative LpL^p-spaces [20]. The long-time behaviour of dynamical semigroups on W∗W^*-algebras was analysed by Frigerio and Verri [16]. Kümmerer’s theory of Markov dilations embeds completely positive dynamics with a faithful invariant state into reversible dynamics [21]. Fidaleo studied strong ergodic properties, including unique ergodicity and uniform convergence of Cesàro averages [15]. More recent work on the structure of ergodic dynamics includes the existence of ergodic states on type III1_1 factors, due to Marrakchi and Vaes [24], and a multiplicative ergodic theorem for cocycles with values in a finite von Neumann algebra, due to Bowen, Hayes and Lin [7]. In this tradition the dynamics is typically homogeneous, and convergence is studied in weak-star, strong, or noncommutative LpL^p topologies relative to a fixed state.

The second line comes from quantum information theory. A quantum channel on a matrix algebra is ergodic or mixing according to spectral properties of its invariant states and its peripheral spectrum [9]. The peripheral structure, period, and cyclic decomposition of a channel were analysed by Carbone and Jenčová [10]. Albert related the asymptotic projection of a channel to its conserved quantities and to matrix product states [5]. Quantitative contraction is central in this line. Strict contractivity of noise channels limits fault-tolerant computation [34]. Quantum Doeblin coefficients give computable upper bounds on trace-distance contraction coefficients, with applications to time-varying channels [19, 18]. Within quantum probability, Accardi, Lu and Souissi proved a Markov–Dobrushin inequality for channels [3]. Souissi and Barhoumi showed that a positive Markov–Dobrushin constant implies exponential mixing, and applied this to matrix product states [35].

When the channel changes in time, the iterates of a single map are replaced by ordered products, and no invariant state need exist. In classical probability this situation is governed by Dobrushin’s ergodicity coefficient. Its extension to abstract state spaces and to von Neumann algebra preduals is due to Mukhamedov [27, 28]. Mukhamedov and Al-Rawashdeh treated weak and uniform ergodicity of nonhomogeneous chains relative to a projection [29]. Gaubert and Qu established the duality between contraction and oscillation for Markov operators on cones [17].

A second group of results concerns random products. Bruneau, Joye and Merkli studied random repeated interaction systems [8]. Movassagh and Schenker obtained an ergodic theorem for processes driven by an ergodic dynamical system, with exponential convergence to an ergodic sequence of states that does not depend on the initial state [25, 26]. Ekblad and Schenker developed a reducibility theory without the decoherence hypothesis of this framework [12]. Nelson and Roon extended the stationary random theory to finite von Neumann algebras, using a projective metric on the cone of positive elements [30]. Pathirana and Schenker studied continuous-time inhomogeneous Markovian dynamics in a stationary random environment [32]. Ergodic theorems for quantum trajectories form a related direction [22, 6, 11].

For deterministic inhomogeneous channels on matrix algebras, Souissi [36] separated the forward and backward composition orders. That paper showed that weak and strong mixing coincide for backward products, which are automatically nested, and derived convergence rates from common positive lower bounds of channel outputs. Pathirana [31] developed a product-level trace-Dobrushin theory for finite-dimensional channel products. In the deterministic case, decay of the product coefficient is equivalent to asymptotic replacement by a moving rank-one channel; for two-sided products, pullback forgetting gives a unique boundary state.

Quantum Markov chains originate in Accardi’s noncommutative Markov property [1]. Later constructions include entangled Markov chains [2] and a unifying framework for the existing definitions [4]. Finitely correlated states were introduced by Fannes, Nachtergaele and Werner [13], and their pure case was analysed in [14]. Matrix product states are their pure vector counterparts, and their representation theory was systematised in [33].

In all these constructions the state of a spin chain is produced by iterating a transfer map on an auxiliary, or bond, algebra. Consequently, the thermodynamic limit, boundary independence, and decay of correlations are statements about products of channels. For translation-invariant or ergodic chains this connection appears in [26, 35]. For non-translation-invariant chains, [36] expresses the infinite-volume state through a backward sequence of boundary states, and [31] obtains infinite-volume limits and correlation bounds from product coefficients. Each of these works assumes a finite-dimensional bond space.

The transfer construction itself does not require this. It defines states for an arbitrary auxiliary von Neumann algebra, for instance B(K)B(K) with KK infinite dimensional, or a II1\mathrm{II}_1 factor. Finite dimensionality is what makes such a state “finitely” correlated in the sense of [13], but it is not needed for the construction. The theory developed below applies directly to this setting.

This paper asks which parts of the finite-dimensional, deterministic theory of [36] remain valid on an arbitrary von Neumann algebra M\mathcal{M}, and which fail. Three features of the infinite-dimensional setting are decisive. The normal state space is in general not norm compact. There may be no trace, no density operators, and no normal pure states. Finally, in applications to spin chains, the bond algebra itself can be infinite dimensional.

The predual Dobrushin coefficient and its submultiplicativity are known [27, 28] and are not presented as new. The contribution of this paper is to determine exactly where directional stability holds for arbitrary deterministic sequences of normal channels. All composition directions refer to maps on states, not to maps on observables. Throughout, SN(M)\mathcal{S}_{N}(\mathcal{M}) denotes the set of normal states of M\mathcal{M}; the subscript n stands for “normal” and is not an integer index (see § 2).

The first two main results describe this boundary.

Theorem 1.1 (Uniform backward stability). Let M\mathcal{M} be a von Neumann algebra, let Tn=(Φn)∗T_n=(\Phi_n)_* be the preduals of normal unital completely positive maps, and put Bn=T0⋯Tn−1B_n=T_0\cdots T_{n-1}. Then

sup⁡φ,ψ∈SN(M)∥Bnφ−Bnψ∥⟶0\sup_{\varphi,\psi\in\mathcal{S}_{N}(\mathcal{M})}\lVert B_n\varphi-B_n\psi\rVert\longrightarrow0

if and only if there exists ω∈SN(M)\omega\in\mathcal{S}_{N}(\mathcal{M}) such that

sup⁡φ∈SN(M)∥Bnφ−ω∥⟶0.\sup_{\varphi\in\mathcal{S}_{N}(\mathcal{M})}\lVert B_n\varphi-\omega\rVert\longrightarrow0.

In that case, for every nn,

sup⁡φ∈SN(M)∥Bnφ−ω∥≤sup⁡φ,ψ∈SN(M)∥Bnφ−Bnψ∥.\sup_{\varphi\in\mathcal{S}_{N}(\mathcal{M})}\lVert B_n\varphi-\omega\rVert\leq\sup_{\varphi,\psi\in\mathcal{S}_{N}(\mathcal{M})}\lVert B_n\varphi-B_n\psi\rVert.

No norm compactness of the state space or of a channel image is required.

Theorem 1.2 (A finite-dimensionality characterization). For a nonzero σ\sigma-finite von Neumann algebra M\mathcal{M}, the following are equivalent:

  1. M\mathcal{M} is finite dimensional.

  1. For every sequence of normal unital completely positive maps, pointwise loss of memory of the backward state products BnB_n implies mixing to a normal state.

  1. The implication in (2) holds for every such sequence for which ⋃n≥1Bn(S(M))\bigcup_{n\ge1} B_n(\mathcal{S}(\mathcal{M})) is norm relatively compact.

If M\mathcal{M} is infinite dimensional, the counterexample can be chosen so that Bn(S(M))B_n(\mathcal{S}(\mathcal{M})) is the same compact line segment for every n≥1n \ge1, while the process has pointwise loss of memory and is not even Cesàro ergodic.

Theorem 1.1 is a dimension-free positive result. Theorem 1.2 shows that uniformity cannot be replaced by pointwise convergence: the matrix theorem of [36] holds pointwise only because the state space of a matrix algebra is compact.

The third main result applies the backward theory to spin chains with an infinite-dimensional bond algebra. Let B\mathcal{B} be a von Neumann algebra, let d≥1d \ge1, and let

Ek:Md(C)⊗B⟶B,k≥1,\mathcal{E}_k : M_d(\mathbb{C}) \otimes\mathcal{B} \longrightarrow\mathcal{B}, \qquad k \ge1,

be normal unital completely positive maps. Put Φk(b)=Ek(1⊗b)\Phi_k(b) = \mathcal{E}_k(1 \otimes b) and Tk=(Φk)∗T_k = (\Phi_k)_*. The maps Ek\mathcal{E}_k define, for every normal boundary state σ\sigma at site NN, a finite-volume state φN\varphi_N on A[1,N]=Md(C)⊗N\mathcal{A}_{[1,N]} = M_d(\mathbb{C})^{\otimes N}; see § 11.

Theorem 1.3 (Inhomogeneous transfer states). Assume that δ(Tm⋯Tn−1)→0\delta(T_m \cdots T_{n-1}) \to0 as n→∞n \to\infty for every m≥1m \ge1. Then there is a unique state φ∞\varphi_\infty on the quasi-local algebra A=⋃NA[1,N]\mathcal{A} = \bigcup_N \mathcal{A}_{[1,N]} such that, for all m<Nm < N and all normal states σ\sigma of B\mathcal{B},

∥φNσ∣A[1,m]−φ∞∣A[1,m]∥≤2δ(Tm+1⋯TN).\left\|\varphi_N^\sigma\big|_{\mathcal{A}_{[1,m]}}-\varphi_\infty\big|_{\mathcal{A}_{[1,m]}}\right\| \le2\delta(T_{m+1}\cdots T_N).

Its local restrictions are expressed through the unique normal entrance law of the backward products. For local observables xx on [1,m][1,m] and zz on [m+ℓ+1,m+ℓ+r][m+\ell+1,m+\ell+r],

∣φ∞(x⊗1ℓ⊗z)−φ∞(x)φ∞(z)∣≤4δ(Tm+1⋯Tm+ℓ)∥x∥∥z∥.\left|\varphi_\infty(x \otimes1_\ell\otimes z)-\varphi_\infty(x)\varphi_\infty(z)\right| \le4\delta(T_{m+1}\cdots T_{m+\ell})\|x\|\|z\|.

This theorem extends the boundary-state representation of [36] and the finite-dimensional replacement estimates of [31] to an arbitrary bond algebra. The estimate is uniform in the boundary state, and it cannot be obtained from pointwise memory loss (Remark 11.5).

Further results identify the precise role of each hypothesis. An attained normal minorization coefficient replaces the rank-one construction of [36] and controls δ\delta in any von Neumann algebra (§ 5). Unique normal entrance laws and forward tracking estimates are proved in § 6 and § 7. Explicit counterexamples separate pointwise from uniform convergence, Cesàro convergence from mixing, convergence from a fixed starting time from uniformity over all starting times, and a single strict contraction of a prefix from recurrent contraction (§ 8 and § 9). A Hilbert-space criterion for processes that preserve a faithful normal state appears in § 10.

§ 2 fixes notation and composition orders, and § 3 states the hierarchy of convergence notions. § 4 and § 5 treat the contraction and minorization coefficients. Theorem 1.1 is proved in § 6, and quantitative estimates follow in § 7. Theorem 1.2 and the universal counterexamples are in § 8, with homogeneous comparisons in § 9. § 10 gives the Hilbert-space criterion, § 11 proves Theorem 1.3, and § 12 summarises the conclusions.

Preduals, channels, and composition order

Let M\mathcal{M} be a nonzero von Neumann algebra, with unit 11, predual M∗\mathcal{M}_*, self-adjoint part M∗,sa\mathcal{M}_{*,\mathrm{sa}}, and positive cone M∗,+\mathcal{M}_{*,+}. The normal state space of M\mathcal{M} is

SN(M)={φ∈M∗,+:φ(1)=1},S_N(\mathcal{M})=\{\varphi\in\mathcal{M}_{*,+}:\varphi(1)=1\},

that is, the set of states of M\mathcal{M} that are normal (equivalently, σ\sigma-weakly continuous, or completely additive on orthogonal families of projections). The subscript n abbreviates “normal”. When M=B(H)\mathcal{M}=B(H), normal states are exactly the functionals x↦Tr⁡(ρx)x\mapsto\operatorname{Tr}(\rho x) with ρ\rho a density operator, so SN(M)S_N(\mathcal{M}) is the von Neumann algebra counterpart of the set S(H)S(H) of density operators. Also write

E0={f∈M∗,sa:f(1)=0}.E_0=\{f\in\mathcal{M}_{*,\mathrm{sa}}:f(1)=0\}.

All norms on functionals are predual norms. In particular, ∥φ∥=φ(1)\lVert\varphi\rVert=\varphi(1) for φ≥0\varphi\ge0. When M\mathcal{M} is σ\sigma-finite, a faithful normal state may be denoted by χ\chi; no trace is assumed. The general positive results below do not require σ\sigma-finiteness. That assumption is convenient for the universal counterexamples.

A channel on observables is a normal unital completely positive map Φ:M→M\Phi:\mathcal{M}\to\mathcal{M}. Its predual T=Φ∗T=\Phi_* is defined by (Tf)(x)=f(Φ(x))(Tf)(x)=f(\Phi(x)). It is a contraction and preserves SN(M)S_N(\mathcal{M}). In a general von Neumann algebra the intrinsic normalization statement is

(Tf)(1)=f(1),(Tf)(1)=f(1),

not preservation of an unspecified trace. Density operators and a trace become available only after an appropriate additional representation or semifinite hypothesis.

Fix a sequence (Φn)n≥0(\Phi_n)_{n\ge0} and put Tn=(Φn)∗T_n=(\Phi_n)_*. For m<nm<n define

Fm,n=Tn−1⋯Tm,Bm,n=Tm⋯Tn−1,Fm,m=Bm,m=id⁡.(1)F_{m,n}=T_{n-1}\cdots T_m,\qquad B_{m,n}=T_m\cdots T_{n-1},\qquad F_{m,m}=B_{m,m}=\operatorname{id}. \tag*{(1)}

Thus Fm,nF_{m,n} is the chronological state evolution, and Bm,nB_{m,n} is the reverse-ordered state product. The terms forward and backward will always mean these two state-space orders. Their observable duals are, respectively,

Fm,n∗=Φm⋯Φn−1,Bm,n∗=Φn−1⋯Φm.(2)F^*_{m,n}=\Phi_m\cdots\Phi_{n-1},\qquad B^*_{m,n}=\Phi_{n-1}\cdots\Phi_m. \tag*{(2)}

Here the superscript ∗* denotes the Banach adjoint, with M=(M∗)∗\mathcal{M}=(\mathcal{M}_*)^*.

The reversal is essential: (Φ∘Ψ)∗=Ψ∗∘Φ∗(\Phi\circ\Psi)_*=\Psi_*\circ\Phi_*. In particular, the observable product Φn−1⋯Φ0\Phi_{n-1}\cdots\Phi_0, which is chronological in the Heisenberg picture, has the backward state product B0,nB_{0,n} as its predual. Formula (2) removes any ambiguity in the terms forward and backward. Neither product is an inverse dynamics, a Petz reversal, or a modular adjoint. Such notions would require additional data and are not used here.

The composition identities are

Fm,n=Fk,nFm,k,Bm,n=Bm,kBk,n(m≤k≤n).F_{m,n}=F_{k,n}F_{m,k},\qquad B_{m,n}=B_{m,k}B_{k,n}\qquad(m\le k\le n).

For backward products the state images satisfy

Bm,n+1(SN(M))⊆Bm,n(SN(M)).(3)B_{m,n+1}(S_N(\mathcal{M}))\subseteq B_{m,n}(S_N(\mathcal{M})). \tag*{(3)}

For forward products one instead has Fm,n+1=TnFm,nF_{m,n+1}=T_nF_{m,n}, which need not give nested images. This is an order-of-composition distinction; it occurs already for classical noncommuting stochastic matrices and is not exclusively quantum.

Convergence notions and the correct hierarchy

Fix mm and one of the families Pn=Fm,nP_n = F_{m,n} or Pn=Bm,nP_n = B_{m,n}, and write

AN=1N∑j=0N−1Pm+j,Pm=id⁡.A_N = \frac{1}{N}\sum_{j=0}^{N-1} P_{m+j}, \qquad P_m = \operatorname{id}.

The initial index will be suppressed when m=0m = 0.

Definition 3.1. The family (Pn)(P_n) is mixing if there exists ω∈SN(M)\omega\in S_N(\mathcal{M}) with ∥Pnφ−ω∥→0\lVert P_n\varphi- \omega\rVert\to0 for every φ∈SN(M)\varphi\in S_N(\mathcal{M}). It is uniformly mixing if this convergence is uniform over SN(M)S_N(\mathcal{M}), and exponentially mixing if, for constants C<∞C < \infty and 0<r<10 < r < 1,

sup⁡φ∈SN(M)∥Pnφ−ω∥≤Crn−m.\sup_{\varphi\in S_N(\mathcal{M})} \lVert P_n\varphi- \omega\rVert\le Cr^{n-m}.

It is mean ergodic to one state if ∥ANφ−ω∥→0\lVert A_N\varphi- \omega\rVert\to0 for all normal states, and uniformly mean ergodic to one state if the convergence is uniform.

It has pointwise loss of memory if

∥Pnφ−Pnψ∥⟶0(φ,ψ∈SN(M)),\lVert P_n\varphi- P_n\psi\rVert\longrightarrow0 \qquad(\varphi,\psi\in S_N(\mathcal{M})),

and uniform loss of memory if dn:=sup⁡φ,ψ∈SN(M)∥Pnφ−Pnψ∥→0d_n := \sup_{\varphi,\psi\in S_N(\mathcal{M})} \lVert P_n\varphi- P_n\psi\rVert\to0. The exponential version is dn≤Crn−md_n \le Cr^{n-m}. Mean loss of memory and uniform mean loss of memory are defined by replacing PnP_n by ANA_N in these two conditions.

The phrases “mean ergodic” and “ergodic” below refer to convergence to a single state, not merely existence of an arbitrary mean-ergodic projection. The term “weak mixing” in [36] corresponds to the norm loss-of-memory property just defined. It does not mean convergence in a weak topology, and is not the usual ergodic-theoretic weak mixing defined by Cesàro decay of absolute correlations. Keeping these notions separate prevents an otherwise substantial terminological ambiguity.

Proposition 3.2 (Hierarchy). For either composition order,

exponential mixing⟹uniform mixing⟹mixing⟹mean ergodicity.\mathit{exponential\ mixing} \Longrightarrow\mathit{uniform\ mixing} \Longrightarrow\mathit{mixing} \Longrightarrow\mathit{mean\ ergodicity}.

and

uniform mixing⟹uniform mean ergodicity⟹mean ergodicity.\mathit{uniform\ mixing} \Longrightarrow\mathit{uniform\ mean\ ergodicity} \Longrightarrow\mathit{mean\ ergodicity}.

The analogous implications hold for loss of memory. Each convergence-to-one-state property implies its corresponding loss-of-memory property. In finite dimension, pointwise and uniform versions coincide, both for products and for averages. In infinite dimension, mixing need not imply uniform mean ergodicity.

Proof. Convergence to a common state gives loss of memory by the triangle inequality. Cesàro summation preserves pointwise norm convergence, and also preserves convergence of the relevant suprema when these converge to zero. To prove the finite-dimensional assertion, use that all channels and their averages are contractions. If Pnφ→ωP_n\varphi\to\omega pointwise but not uniformly, choose φn\varphi_n violating uniform convergence and pass to a norm-convergent subsequence. Contractivity gives

∥Pnφn−ω∥≤∥φn−φ∥+∥Pnφ−ω∥⟶0,\lVert P_n\varphi_n-\omega\rVert\le\lVert\varphi_n-\varphi\rVert+\lVert P_n\varphi-\omega\rVert\longrightarrow0,

a contradiction. Apply the same argument to pairs of states and to averages. The final assertion follows from Example 8.2.

Thus the correct infinite-dimensional diagram branches: mixing implies mean ergodicity, but there is no general arrow from mixing to uniform mean ergodicity. In finite dimension the appropriate chain is

exponential mixing⟹uniform mixing⟺mixing\text{exponential mixing}\Longrightarrow\text{uniform mixing}\Longleftrightarrow\text{mixing}
⟹uniform mean ergodicity⟺mean ergodicity.\Longrightarrow\text{uniform mean ergodicity}\Longleftrightarrow\text{mean ergodicity}.

The two remaining reverse arrows fail for inhomogeneous processes; explicit examples appear below. For a homogeneous process, uniform mixing always has an exponential rate; see Theorem 9.3.

The predual contraction coefficient

For a channel predual TT define

δ(T)=∥T∣E0∥=sup⁡0≠f∈E0∥Tf∥∥f∥,(4)\delta(T)=\lVert T|_{E_0}\rVert=\sup_{0\ne f\in E_0}\frac{\lVert Tf\rVert}{\lVert f\rVert}, \tag*{(4)}

with δ(T)=0\delta(T)=0 if E0={0}E_0=\{0\}. This is the usual predual Dobrushin coefficient, as in [27, 28]. The following elementary facts are included to make all subsequent bounds self-contained.

Lemma 4.1. For channel preduals TT, SS,

0≤δ(T)≤1,δ(TS)≤δ(T)δ(S),2δ(T)=sup⁡φ,ψ∈SN(M)∥Tφ−Tψ∥.(5)0\le\delta(T)\le1,\qquad\delta(TS)\le\delta(T)\delta(S),\qquad2\delta(T)=\sup_{\varphi,\psi\in S_{\mathcal N}(\mathcal M)}\lVert T\varphi-T\psi\rVert. \tag*{(5)}

For x=x∗∈Mx=x^*\in\mathcal M put osc⁡(x)=sup⁡spec⁡(x)−inf⁡spec⁡(x)\operatorname{osc}(x)=\sup\operatorname{spec}(x)-\inf\operatorname{spec}(x). Then

δ(T)=sup⁡x=x∗, osc⁡(x)>0osc⁡(T∗x)osc⁡(x).(6)\delta(T)=\sup_{\substack{x=x^*,\ \operatorname{osc}(x)>0}}\frac{\operatorname{osc}(T^*x)}{\operatorname{osc}(x)}. \tag*{(6)}

Proof. A self-adjoint normal functional ff has a normal Jordan decomposition f=f+−f−f=f_+-f_- with ∥f∥=f+(1)+f−(1)\lVert f\rVert=f_+(1)+f_-(1). If f∈E0f\in E_0, then f+(1)=f−(1)=∥f∥/2f_+(1)=f_-(1)=\lVert f\rVert/2. For f≠0f\ne0, therefore,

f=∥f∥2(φ−ψ)f=\frac{\lVert f\rVert}{2}(\varphi-\psi)

for some φ,ψ∈SN(M)\varphi,\psi\in S_{\mathcal N}(\mathcal M).

This gives δ(T)≤12sup⁡∥Tφ−Tψ∥\delta(T)\le\frac12\sup\lVert T\varphi-T\psi\rVert. Conversely, ∥Tφ−Tψ∥≤δ(T)∥φ−ψ∥≤2δ(T)\lVert T\varphi-T\psi\rVert\le\delta(T)\lVert\varphi-\psi\rVert\le2\delta(T). The coefficient is at most one by contractivity; submultiplicativity follows because S(E0)⊆E0S(E_0)\subseteq E_0.

For the last assertion, the real Banach dual of E0E_0 is Msa/R1\mathcal{M}_{\mathrm{sa}}/\mathbb{R}1, with quotient norm

inf⁡c∈R∥x−c1∥=12osc⁡(x).\inf_{c\in\mathbb{R}}\lVert x-c1\rVert= \frac{1}{2}\operatorname{osc}(x).

Indeed the annihilator of E0E_0 is R1\mathbb{R}1, and centering the spectrum gives the displayed equality. The adjoint of T∣E0T|_{E_0} is the induced map of this quotient. Equality of the norms of an operator and its adjoint proves (6).

All operator norms in statements using E0E_0 are real operator norms. Uniform convergence over states is equivalent to operator norm convergence on M∗,sa\mathcal{M}_{*,\mathrm{sa}} to the rank-one map

Rωf=f(1)ω.R_\omega f=f(1)\omega.

In fact the operator norm of T−RωT-R_\omega on this real space equals its supremum on states, by Jordan decomposition. On the complex predual these two norms are equivalent with a factor at most two, which is sufficient for convergence but should not be silently replaced by an exact identity.

An attained normal minorization coefficient

Definition 5.1. For a channel predual TT let

K(T)={κ∈M∗,+:κ≤Tφ for every φ∈SN(M)},β(T)=sup⁡κ∈K(T)κ(1).\mathcal{K}(T)=\{\kappa\in\mathcal{M}_{*,+}:\kappa\leq T\varphi\text{ for every }\varphi\in\mathcal{S}_{N}(\mathcal{M})\},\qquad \beta(T)=\sup_{\kappa\in\mathcal{K}(T)}\kappa(1).

The order is the ordinary order of normal positive functionals. Thus κ≤λ\kappa\leq\lambda means κ(x)≤λ(x)\kappa(x)\leq\lambda(x) for all x∈M+x\in\mathcal{M}_+.

Theorem 5.2 (Normal common-lower-bound theorem). For every von Neumann algebra and every normal channel predual TT, the maximum defining β(T)\beta(T) is attained. Moreover,

0≤β(T)≤1,δ(T)≤1−β(T).(7)0\leq\beta(T)\leq1,\qquad\delta(T)\leq1-\beta(T). \tag*{(7)}

A normal positive functional κ\kappa belongs to K(T)\mathcal{K}(T) if and only if

T∗(x)≥κ(x)1(x∈M+).(8)T^*(x)\geq\kappa(x)1\qquad(x\in\mathcal{M}_+). \tag*{(8)}

When M=B(H)\mathcal{M}=B(H) with dim⁡H<∞\dim H<\infty, writing κ(x)=Tr⁡(kx)\kappa(x)=\operatorname{Tr}(kx) identifies β(T)\beta(T) with the maximal trace of a common positive lower bound of the outputs of all rank-one input states. Thus it recovers the variational common-lower-bound coefficient of [36], with ηMD=1−β(T)\eta_{\mathrm{MD}}=1-\beta(T).

Proof. The set K(T)\mathcal{K}(T) contains zero. Fix φ0∈SN(M)\varphi_0\in\mathcal{S}_{N}(\mathcal{M}) and put λ=Tφ0\lambda=T\varphi_0. Regard the order interval [0,λ][0,\lambda] as a subset of the full Banach dual M∗\mathcal{M}^*. It is σ(M∗,M)\sigma(\mathcal{M}^*,\mathcal{M}) closed and contained in its closed unit ball, hence compact by Banach–Alaoglu. Every ff in this interval is normal: if xi↑xx_i\uparrow x in M+\mathcal{M}_+, then

0≤f(x−xi)≤λ(x−xi)⟶0.0\leq f(x-x_i)\leq\lambda(x-x_i)\longrightarrow0.

For every normal state φ\varphi, the constraint f≤Tφf\leq T\varphi is weak-star closed, since it is the intersection over x∈M+x\in\mathcal{M}_+ of the scalar closed inequalities f(x)≤(Tφ)(x)f(x)\leq(T\varphi)(x). Consequently K(T)\mathcal{K}(T) is a compact subset of [0,λ][0,\lambda] in this topology. Evaluation at 11 is continuous and therefore has a maximum. Its value lies in [0,1][0,1].

For admissible κ\kappa, let b=κ(1)b=\kappa(1). The map

S=T−Rκ,Rκ(f)=f(1)κ,S=T-R_{\kappa},\qquad R_{\kappa}(f)=f(1)\kappa,

is positive: for f≥0f\geq0, normalize ff to a state when it is nonzero. For such ff, (Sf)(1)=(1−b)f(1)(Sf)(1)=(1-b)f(1). A Jordan decomposition therefore gives ∥Sf∥≤(1−b)∥f∥\lVert Sf\rVert\leq(1-b)\lVert f\rVert on M∗,sa\mathcal{M}_{*,\mathrm{sa}}. On E0E_0, S=TS=T, so δ(T)≤1−b\delta(T)\leq1-b. Take a maximizing κ\kappa.

The order condition is equivalent to φ(T∗(x))≥κ(x)\varphi(T^*(x))\geq\kappa(x) for every normal state and every positive xx. Normal states determine the order in a von Neumann algebra: a nonzero negative spectral projection of a self-adjoint element supports a normal state detecting the negative part. This proves (8).

In finite dimension, normal states are convex combinations of rank-one states. A lower bound for all rank-one outputs is consequently a lower bound for all state outputs, and the converse is immediate. The density-functional correspondence preserves order and identifies κ(1)\kappa(1) with Tr⁡(k)\operatorname{Tr}(k).

Remark 5.3. The compactness used in the proof is the weak-star compactness of an order interval inside M∗\mathcal{M}^*, not norm compactness of SN(M)\mathcal{S}_N(\mathcal{M}). Domination by one fixed normal functional prevents a singular functional from appearing in that interval. The maximizing functional need not be unique, but its maximal mass is a scalar and is unambiguous. No lattice infimum in the noncommutative order is asserted.

Example 5.4 (Strict contraction without a common lower bound). On ℓ∞({1,2,3})\ell^\infty(\{1,2,3\}), take a classical channel with output laws

p1=(1/2,1/2,0),p2=(0,1/2,1/2),p3=(1/2,0,1/2).p_1=(1/2,1/2,0),\qquad p_2=(0,1/2,1/2),\qquad p_3=(1/2,0,1/2).

The smallest value of each output coordinate over the three inputs is zero, so β(T)=0\beta(T)=0. Pairwise distances are one, and convexity shows that the diameter of the output simplex is one. Lemma 4.1 gives δ(T)=1/2\delta(T)=1/2. Thus 1−β(T)1-\beta(T) is a sufficient contraction bound, not the optimal coefficient in general.

Example 5.5 (Positive subtraction is not CP subtraction). On M2(C)M_2(\mathbb{C}), in the Schrödinger picture, let

T(ρ)=Tr⁡(ρ)I+ρT3.T(\rho)=\frac{\operatorname{Tr}(\rho)I+\rho^{\mathsf{T}}}{3}.

Its Choi matrix is (I⊗I+F)/3≥0(I\otimes I+F)/3\geq0, where FF is the flip; also Tr⁡(Tρ)=Tr⁡ρ\operatorname{Tr}(T\rho)=\operatorname{Tr}\rho. Every state output dominates I/3I/3. Conversely, given a unit vector vv, choose a pure input ρ\rho with ρv=0\rho v=0; any common lower bound kk then satisfies ⟨v,kv⟩≤1/3\langle v,kv\rangle\leq1/3. Hence the maximal lower bound is I/3I/3 and β(T)=2/3\beta(T)=2/3. But

T(ρ)−Tr⁡(ρ)I/3=ρT/3T(\rho)-\operatorname{Tr}(\rho)I/3=\rho^{\mathsf{T}}/3

is not completely positive: its Choi matrix F/3F/3 has a negative eigenvalue. Thus Theorem 5.2 only uses positivity of the residual. It must not be read as a CP-order Doeblin decomposition.

Directional convergence and normal entrance laws

Theorem 6.1 (Nested images with shrinking diameter). Let PnP_n be channel preduals on a von Neumann algebra. Suppose that Pn+1(SN(M))⊆Pn(SN(M))P_{n+1}(\mathcal{S}_N(\mathcal{M})) \subseteq P_n(\mathcal{S}_N(\mathcal{M})) and dn=diam⁡(Pn(SN(M)))→0d_n = \operatorname{diam}(P_n(\mathcal{S}_N(\mathcal{M}))) \to0. Then there exists a unique ω∈SN(M)\omega\in\mathcal{S}_N(\mathcal{M}) such that

sup⁡φ∈SN(M)∥Pnφ−ω∥≤dn.\sup_{\varphi\in\mathcal{S}_N(\mathcal{M})} \lVert P_n\varphi-\omega\rVert\leq d_n.

The sets Pn(SN(M))P_n(\mathcal{S}_N(\mathcal{M})) need not be closed. The state ω\omega belongs to the norm closure of each set Pn(SN(M))P_n(\mathcal{S}_N(\mathcal{M})).

Proof. Choose φ0∈SN(M)\varphi_0 \in\mathcal{S}_N(\mathcal{M}) and set zn=Pnφ0z_n=P_n\varphi_0. If k≥nk\geq n, nesting puts zk,znz_k,z_n in Pn(SN(M))P_n(\mathcal{S}_N(\mathcal{M})), so ∥zk−zn∥≤dn\lVert z_k-z_n\rVert\leq d_n. Thus (zn)(z_n) is norm Cauchy. The predual is complete, and its positive cone and normalization hyperplane are norm closed, hence its limit ω\omega is a normal state. For any φ∈SN(M)\varphi\in\mathcal{S}_N(\mathcal{M}) and k≥nk\geq n, the two states Pnφ,zkP_n\varphi,z_k lie in Pn(SN(M))P_n(\mathcal{S}_N(\mathcal{M})), so ∥Pnφ−zk∥≤dn\lVert P_n\varphi-z_k\rVert\leq d_n. Pass to the limit in kk. The bound proves both uniqueness and membership in each image closure.

Apply this theorem to (3), and use (5). It proves Theorem 1.1. In particular, for every fixed mm, uniform backward mixing is equivalent to δ(Bm,n)→0\delta(B_{m,n})\to0, and exponential uniform backward memory loss implies exponential backward mixing with the same diameter bound. This dimension-free conclusion does not extend to pointwise memory loss, as § 8 will show.

Corollary 6.2 (Forward nesting). If F0,n+1(SN(M))⊆F0,n(SN(M))F_{0,n+1}(\mathcal{S}_N(\mathcal{M}))\subseteq F_{0,n}(\mathcal{S}_N(\mathcal{M})) for all nn and the forward process has uniform loss of memory, then it is uniformly mixing, with sup⁡φ∥F0,nφ−ω∥≤2δ(F0,n)\sup_\varphi\lVert F_{0,n}\varphi-\omega\rVert\leq2\delta(F_{0,n}). In particular exponential uniform loss of memory and nesting give exponential mixing with the same rate, in either direction.

This is the infinite-dimensional form of the nesting principle of [36]. In finite dimension pointwise memory loss suffices there, by Proposition 3.2. In infinite dimension the hypothesis must be uniform: the backward products of Theorem 8.4 below are nested and lose memory pointwise, yet do not converge.

Theorem 6.3 (Entrance laws). Suppose that, for every fixed m≥0m\geq0, δ(Bm,n)→0\delta(B_{m,n})\to0 as n→∞n\to\infty. Then there is a unique sequence of normal states (ωm)m≥0(\omega_m)_{m\geq0} satisfying

ωm=Tmωm+1(m≥0).(9)\omega_m=T_m\omega_{m+1}\qquad(m\geq0). \tag*{(9)}

For this sequence,

sup⁡φ∈SN(M)∥Bm,nφ−ωm∥≤2δ(Bm,n).(10)\sup_{\varphi\in\mathcal{S}_N(\mathcal{M})}\lVert B_{m,n}\varphi-\omega_m\rVert\leq2\delta(B_{m,n}). \tag*{(10)}

More generally, the same estimate holds with an arbitrarily chosen boundary state φn\varphi_n at time nn in place of a fixed φ\varphi.

Proof. For each mm, Theorem 6.1 gives a limiting normal state ωm\omega_m and the bound. Since Bm,n=TmBm+1,nB_{m,n}=T_mB_{m+1,n}, norm continuity gives (9). If (νm)(\nu_m) is another normal entrance law, iteration gives νm=Bm,nνn\nu_m=B_{m,n}\nu_n, hence ∥νm−ωm∥≤2δ(Bm,n)→0\lVert\nu_m-\omega_m\rVert\leq2\delta(B_{m,n})\to0. The uniform bound also proves the assertion about arbitrary boundary sequences.

A single fixed starting time does not suffice for uniqueness of a full entrance law. For example, T0=RωT_0=R_\omega and Tn=id⁡T_n=\operatorname{id} for n≥1n\ge1 give B0,n=RωB_{0,n}=R_\omega for all n≥1n\ge1, but (ω,ν,ν,…)(\omega,\nu,\nu,\ldots) is an entrance law for every normal state ν\nu.

Proposition 6.4 (Synchronization and convergence). Let PnP_n have pointwise loss of memory. Then (Pn)(P_n) is mixing if and only if Pnφ0P_n\varphi_0 converges in predual norm for one normal state φ0\varphi_0. If loss of memory is uniform and zn=Pnφ0z_n=P_n\varphi_0, then

sup⁡φ∈SN(M)∥Pnφ−ω∥≤2δ(Pn)+∥zn−ω∥.\sup_{\varphi\in S_N(\mathcal{M})}\lVert P_n\varphi-\omega\rVert\le2\delta(P_n)+\lVert z_n-\omega\rVert.

The same statements hold for Cesàro averages. In particular, a common invariant normal state ω\omega converts pointwise or uniform memory loss into the corresponding mixing property, in either direction; it also converts mean memory loss into mean ergodicity.

Proof. Use ∥Pnφ−ω∥≤∥Pnφ−Pnφ0∥+∥Pnφ0−ω∥\lVert P_n\varphi-\omega\rVert\le\lVert P_n\varphi-P_n\varphi_0\rVert+\lVert P_n\varphi_0-\omega\rVert. The necessity is immediate. If every TnT_n fixes ω\omega, then every product and every average fixes it, so it is an admissible constant anchor.

Example 6.5 (Complete memory loss with a moving output). Choose distinct normal states α,γ\alpha,\gamma and let Tn=RσnT_n=R_{\sigma_n}, where σn=α\sigma_n=\alpha for even nn and σn=γ\sigma_n=\gamma for odd nn. Since RαRγ=RαR_\alpha R_\gamma=R_\alpha, one has

F0,n=Rσn−1,B0,n=Rσ0(n≥1).F_{0,n}=R_{\sigma_{n-1}},\qquad B_{0,n}=R_{\sigma_0}\quad(n\ge1).

Both image diameters vanish after one step. The backward process is stationary after that step; the forward process does not converge. Its Cesàro averages converge uniformly to R(α+γ)/2R_{(\alpha+\gamma)/2}. This example works already on a two-state classical algebra. It separates synchronization from convergence, not finite from infinite dimension.

If the sequence σn\sigma_n instead alternates between α\alpha and γ\gamma on blocks whose successive lengths dominate the sum of all previous lengths, its Cesàro means fail to converge. The forward process then has even exponential loss of memory, and uniform mean loss of memory, without mean ergodicity.

Nesting is sufficient for convergence under uniform memory loss; it is not necessary. For instance, take distinct states α,ω\alpha,\omega and σn=(1−1/(n+2))ω+(1/(n+2))α\sigma_n=(1-1/(n+2))\omega+(1/(n+2))\alpha. The forward replacement products converge uniformly to ω\omega, but their singleton images are distinct and hence are not nested.

Quantitative contraction, windows, and tracking

Theorem 7.1 (Deterministic contraction estimates). For either order Pm,n=Fm,nP_{m,n}=F_{m,n} or Pm,n=Bm,nP_{m,n}=B_{m,n},

δ(Pm,n)≤∏j=mn−1δ(Tj)≤Qm,n:=∏j=mn−1(1−β(Tj)).(11)\delta(P_{m,n})\le\prod_{j=m}^{n-1}\delta(T_j)\le Q_{m,n}:=\prod_{j=m}^{n-1}(1-\beta(T_j)). \tag*{(11)}

Consequently, ∑j=0∞β(Tj)=∞\sum_{j=0}^{\infty}\beta(T_j)=\infty implies uniform loss of memory in both orders from every fixed starting time. Backward normal entrance laws then exist uniquely and satisfy

sup⁡φ∈SN(M)∥Bm,nφ−ωm∥≤2Qm,n.\sup_{\varphi\in S_N(\mathcal{M})}\lVert B_{m,n}\varphi-\omega_m\rVert\le2Q_{m,n}.

If every TjT_j fixes a common normal state ω\omega, then

sup⁡φ∈SN(M)∥Pm,nφ−ω∥≤2Qm,n\sup_{\varphi\in S_N(\mathcal{M})}\|P_{m,n}\varphi-\omega\|\le2Q_{m,n}

in both orders.

Proof. Apply Lemma 4.1 repeatedly and then Theorem 5.2. The estimate 1−u≤e−u1-u\le e^{-u} for 0≤u≤10\le u\le1 gives Qm,n≤exp⁡(−∑j=mn−1β(Tj))Q_{m,n}\le\exp\left(-\sum_{j=m}^{n-1}\beta(T_j)\right); a factor equal to zero only strengthens the bound. Lemma 4.1, Theorem 6.3, and Proposition 6.4 give the remaining assertions.

If β(Tj)≥ε>0\beta(T_j)\ge\varepsilon>0 on a set GG of indices and NG(m,n)=#(G∩[m,n))N_G(m,n)=\#(G\cap[m,n)), the bound is

2Qm,n≤2(1−ε)NG(m,n).(12)2Q_{m,n}\le2(1-\varepsilon)^{N_G(m,n)}. \tag*{(12)}

Positive lower density, namely NG(0,n)≥cn−C0N_G(0,n)\ge cn-C_0 with c>0c>0, gives an exponential rate in physical time. Merely having NG(0,n)=O(n)N_G(0,n)=O(n) does not: that upper bound is true for every subset of indices. If NG(0,n)≍log⁡nN_G(0,n)\asymp\log n, the guaranteed bound is polynomial. Furthermore, divergence of ∑jβ(Tj)\sum_j\beta(T_j) can hold when β(Tj)→0\beta(T_j)\to0, so a positive accumulation point is not needed.

Contraction may also occur only in blocks. For a partition m=n0<n1<⋯m=n_0<n_1<\cdots and either chosen direction, put qk=δ(Pnk,nk+1)q_k=\delta(P_{n_k,n_{k+1}}). Whenever n≥nKn\ge n_K, composition yields

δ(Pm,n)≤∏k=0K−1qk.\delta(P_{m,n})\le\prod_{k=0}^{K-1}q_k.

This condition is strictly more flexible than a one-step minorization condition. The order of each block must agree with the order of the process; a forward block estimate cannot simply be transferred to a backward block.

Proposition 7.2 (Uniformity over starting times). For either direction let

h(ℓ)=sup⁡m≥0δ(Pm,m+ℓ).h(\ell)=\sup_{m\ge0}\delta(P_{m,m+\ell}).

Then the following conditions are equivalent:

  1. h(ℓ)→0h(\ell)\to0 as ℓ→∞\ell\to\infty.

  2. There are L≥1L\ge1 and q<1q<1 such that h(L)≤qh(L)\le q.

Under (2), h(ℓ)≤q⌊ℓ/L⌋h(\ell)\le q^{\lfloor\ell/L\rfloor}; if q=0q=0, the coefficient is zero for all ℓ≥L\ell\ge L.

Proof. The composition identities and submultiplicativity give h(a+b)≤h(a)h(b)h(a+b)\le h(a)h(b). If h(ℓ)→0h(\ell)\to0, choose LL with h(L)<1h(L)<1. Conversely, write ℓ=kL+r\ell=kL+r, 0≤r<L0\le r<L, and use h(r)≤1h(r)\le1.

The statement “for each fixed mm, δ(Pm,n)→0\delta(P_{m,n})\to0” is weaker than uniformity over mm. For example, take Tj=id⁡T_j=\operatorname{id} except at j=2kj=2^k, where Tj=12id⁡+12RωT_j=\frac{1}{2}\operatorname{id}+\frac{1}{2}R_\omega on a nontrivial algebra. Every fixed tail contains infinitely many contraction steps, so it loses memory uniformly over states. There are nevertheless arbitrarily long intervals containing only identity maps, and h(ℓ)=1h(\ell)=1 for every finite ℓ\ell.

Theorem 7.3 (Tracking an approximate forward trajectory). Suppose δ(Fj,n)≤qj,n≤1\delta(F_{j,n}) \le q_{j,n} \le1 for j<nj<n, and put qn,n=1q_{n,n}=1. For a sequence νj∈SN(M)\nu_j \in S_{\mathrm{N}}(\mathcal{M}), set εj=∥Tjνj−νj+1∥\varepsilon_j=\lVert T_j\nu_j-\nu_{j+1}\rVert. Then

∥Fm,nφ−νn∥≤2qm,n+∑j=mn−1qj+1,nεj.(13)\lVert F_{m,n}\varphi-\nu_n\rVert\le2q_{m,n}+\sum_{j=m}^{n-1}q_{j+1,n}\varepsilon_j. \tag*{(13)}

If qj,n→0q_{j,n}\to0 for each fixed jj and ∑jεj<∞\sum_j\varepsilon_j<\infty, the left-hand side tends to zero uniformly over initial states at any fixed mm. The same conclusion holds when qj,n≤Crn−jq_{j,n}\le Cr^{n-j} for some r<1r<1 and εj→0\varepsilon_j\to0.

Proof. A telescoping expansion gives

Fm,nφ−νn=Fm,n(φ−νm)+∑j=mn−1Fj+1,n(Tjνj−νj+1).F_{m,n}\varphi-\nu_n=F_{m,n}(\varphi-\nu_m)+\sum_{j=m}^{n-1}F_{j+1,n}(T_j\nu_j-\nu_{j+1}).

Every functional in parentheses belongs to E0E_0. Apply the restriction norm bound and ∥φ−νm∥≤2\lVert\varphi-\nu_m\rVert\le2. In the summable case, split the sum into a finite initial part and a tail with small total mass. In the exponential case, first make the late errors small and use the geometric sum bound, then let the finitely many early contributions decay.

In particular, contraction gives a synchronized moving state, not necessarily a stationary limit. If νn→ω\nu_n\to\omega, tracking yields mixing to ω\omega; if νn\nu_n oscillates, there need not be any limiting state. This is the forward counterpart of the backward entrance-law statement.

Infinite-dimensional counterexamples and universality

The following retraction transfers classical counterexamples without imposing a trace or an algebra type.

Lemma 8.1 (A countable classical retract). Every infinite-dimensional σ\sigma-finite von Neumann algebra admits normal unital completely positive maps

J:ℓ∞(N0)⟶M,E:M⟶ℓ∞(N0),EJ=id⁡,J:\ell^\infty(\mathbb{N}_0)\longrightarrow\mathcal{M},\qquad E:\mathcal{M}\longrightarrow\ell^\infty(\mathbb{N}_0),\qquad EJ=\operatorname{id},

where JJ is a unital injective ∗*-homomorphism. Its predual E∗:ℓ1(N0)→M∗E_*:\ell^1(\mathbb{N}_0)\to\mathcal{M}_* is an isometry, J∗E∗=id⁡J_*E_*=\operatorname{id}, and both preduals preserve normal states. If QnQ_n are classical stochastic operators on ℓ1(N0)\ell^1(\mathbb{N}_0), then

T^n=E∗QnJ∗(14)\widehat{T}_n=E_*Q_nJ_* \tag*{(14)}

are normal quantum-channel preduals on M\mathcal{M}. Every nonempty product of the lifted maps equals E∗E_* followed by the corresponding classical product followed by J∗J_*. In particular, output distances and suprema over states for nonempty products are exactly the classical ones.

Proof. We first show that M\mathcal{M} contains an infinite family of pairwise orthogonal nonzero projections. Suppose not. Then every nonzero projection ee dominates a minimal projection of M\mathcal{M}: otherwise one can choose a strictly decreasing sequence e=e1>e2>⋯e=e_1>e_2>\cdots of nonzero projections, and the differences ek−ek+1e_k-e_{k+1} form an infinite orthogonal family. Let {q1,…,qr}\{q_1,\ldots,q_r\} be a maximal orthogonal family of minimal projections; it is finite by assumption. If 1−∑iqi≠01-\sum_i q_i\ne0, it dominates a further minimal projection, contradicting maximality. Hence ∑iqi=1\sum_i q_i=1. For a minimal projection qq, the von Neumann algebra qMqq\mathcal{M}q has no projections other than 00 and qq, so the spectral theorem gives qMq=Cqq\mathcal{M}q=\mathbb{C}q. Let x,y∈qiMqjx,y\in q_i\mathcal{M}q_j with x≠0x\ne0. Then x∗x∈qjMqjx^*x\in q_j\mathcal{M}q_j and xx∗∈qiMqixx^*\in q_i\mathcal{M}q_i, so x∗x=cqjx^*x=cq_j and xx∗=cqixx^*=cq_i with c=∥x∥2>0c=\lVert x\rVert^2>0. Also x∗y=μqjx^*y=\mu q_j for some μ∈C\mu\in\mathbb{C}. Therefore y=qiy=c−1xx∗y=c−1μxqj=c−1μxy=q_iy=c^{-1}xx^*y=c^{-1}\mu xq_j=c^{-1}\mu x. Thus dim⁡qiMqj≤1\dim q_i\mathcal{M}q_j\le1, and dim⁡M=dim⁡∑i,jqiMqj≤r2\dim\mathcal{M}=\dim\sum_{i,j}q_i\mathcal{M}q_j\le r^2, a contradiction.

Let (qj)j≥0(q_j)_{j\ge0} be such an infinite orthogonal family. Put p0=q0+(1−∑jqj)p_0=q_0+(1-\sum_jq_j) and pj=qjp_j=q_j for j≥1j\ge1, the sum being strongly convergent. These are nonzero pairwise orthogonal projections with ∑jpj=1\sum_jp_j=1.

Choose a faithful normal state χ\chi and set

σj(x)=χ(pjxpj)χ(pj),J(a)=∑j≥0ajpj,E(x)=(σj(x))j≥0.\sigma_j(x)=\frac{\chi(p_jxp_j)}{\chi(p_j)},\qquad J(a)=\sum_{j\ge0}a_jp_j,\qquad E(x)=(\sigma_j(x))_{j\ge0}.

These maps are normal and unital. The map JJ is a ∗*-homomorphism; EE is completely positive because each coordinate is a positive scalar functional. Since σj(pk)=δjk\sigma_j(p_k)=\delta_{jk}, EJ=id⁡EJ=\operatorname{id}.

For a∈ℓ1a\in\ell^1, E∗a=∑jajσjE_*a=\sum_j a_j\sigma_j in norm. The triangle inequality gives an upper bound ∑∣aj∣\sum|a_j|, and testing on J(b)J(b) with ∥b∥∞≤1\lVert b\rVert_\infty\le1 and ajbj=∣aj∣a_jb_j=|a_j| gives the reverse bound. Thus E∗E_* is isometric. The predual identity follows from EJ=id⁡EJ=\operatorname{id}. Every classical stochastic map has a normal unital positive, hence completely positive, adjoint on the commutative algebra. Therefore the adjoint of (14) is JQn∗EJQ_n^*E, a normal unital CP map. Cancellation of J∗E∗J_*E_* proves the product formula. Finally J∗J_* maps states onto all probability vectors, since J∗E∗=id⁡J_*E_*=\operatorname{id}, proving equality of suprema.

Example 8.2 (Pointwise mixing without uniformity). On ℓ1(N0)\ell^1(\mathbb{N}_0), define the stochastic operator

Ce0=e0,Cej=ej−1(j≥1).Ce_0=e_0,\qquad Ce_j=e_{j-1}\quad(j\ge1).

where eje_j is the point mass at jj. Its adjoint on ℓ∞\ell^\infty is the normal unital ∗*-homomorphism C∗a=(a0,a0,a1,a2,…)C^*a=(a_0,a_0,a_1,a_2,\ldots). For every probability vector pp,

∥Cnp−e0∥1=2∑j>npj⟶0.(15)\lVert C^np-e_0\rVert_1=2\sum_{j>n}p_j\longrightarrow0. \tag*{(15)}

The convergence is not uniform: take p=en+1p=e_{n+1} to obtain distance two. In fact δ(Cn)=1\delta(C^n)=1 for all nn.

Let AN=N−1∑k=0N−1CkA_N=N^{-1}\sum_{k=0}^{N-1}C^k. Pointwise mixing gives ANp→e0A_Np\to e_0. But for p=eNp=e_N all states CkpC^kp, 0≤k<N0\le k<N, are supported outside zero, so ∥ANp−e0∥1=2\lVert A_Np-e_0\rVert_1=2. Thus the homogeneous process is mixing and mean ergodic, but is neither uniformly mixing nor uniformly mean ergodic. Lifting CC by Lemma 8.1 gives the same homogeneous counterexample on every infinite-dimensional σ\sigma-finite M\mathcal{M}, in both directions.

This example simultaneously proves that pointwise mixing does not imply uniform mean ergodicity and that mean ergodicity does not imply its uniform version. A trace on the algebra does not by itself repair either implication: the lifted construction works on any II1\mathrm{II}_1 factor, although the chosen channels need not preserve its trace.

Example 8.3 (A strict prefix bound does not accumulate). Fix 0<q<10<q<1, take T0=q id+(1−q)Re0T_0=q\,\mathrm{id}+(1-q)R_{e_0}, and Tj=CT_j=C for j≥1j\geq1, with CC as above. These maps commute because Ce0=e0Ce_0=e_0. For n≥1n\geq1, both ordered products are

Pn=qCn−1+(1−q)Re0.P_n=qC^{n-1}+(1-q)R_{e_0}.

Hence every initial state converges to e0e_0, while

sup⁡p∥Pnp−e0∥1=2q,δ(Pn)=q.\sup_p\lVert P_np-e_0\rVert_1=2q,\qquad\delta(P_n)=q.

For the coefficient identity use the upper bound and the pair e0,ene_0,e_n, whose outputs have distance 2q2q. Thus even mixing together with lim sup⁡nδ(Pn)<1\limsup_n\delta(P_n)<1 does not imply uniform mixing. Hence a strict bound on the coefficient of the whole prefix cannot replace a recurrent block-contraction hypothesis. Again the example lifts to every infinite-dimensional σ\sigma-finite algebra.

Theorem 8.4 (Compact backward images without stability). Let (an)n≥1(a_n)_{n\geq1} be any sequence in {0,1}\{0,1\}. There are stochastic operators QnQ_n on ℓ1(N0)\ell^1(\mathbb{N}_0) such that, for Rn=Q1⋯QnR_n=Q_1\cdots Q_n,

Rn(P(N0))=conv⁡{e0,e1},(16)R_n\bigl(\mathcal{P}(\mathbb{N}_0)\bigr)=\operatorname{conv}\{e_0,e_1\}, \tag*{(16)}
∥Rnp−ean∥1≤2∑j>npj,(17)\lVert R_np-e_{a_n}\rVert_1\leq2\sum_{j>n}p_j, \tag*{(17)}
∥Rnp−Rnq∥1≤2∑j>n(pj+qj).(18)\lVert R_np-R_nq\rVert_1\leq2\sum_{j>n}(p_j+q_j). \tag*{(18)}

Consequently the backward products have pointwise loss of memory for every choice of (an)(a_n), but they converge to one state if and only if (ean)(e_{a_n}) converges. Their Cesàro averages converge to one state if and only if the scalar means N−1∑n=1NanN^{-1}\sum_{n=1}^N a_n converge. Uniform loss of memory always fails: δ(Rn)=1\delta(R_n)=1.

Proof. For each n≥1n\geq1 define a deterministic function rn:N0→{0,1}r_n:\mathbb{N}_0\to\{0,1\} by

rn(j)={an,j≤n,j mod 2,j>n.r_n(j)= \begin{cases} a_n, & j\leq n,\\ j\bmod2, & j>n. \end{cases}

Let s1=r1s_1=r_1. Given n≥1n\geq1, put un=2n+2u_n=2n+2, vn=2n+3v_n=2n+3, and define

sn+1(j)={un,rn+1(j)=0,vn,rn+1(j)=1.s_{n+1}(j)= \begin{cases} u_n, & r_{n+1}(j)=0,\\ v_n, & r_{n+1}(j)=1. \end{cases}

Since rn(un)=0r_n(u_n)=0 and rn(vn)=1r_n(v_n)=1, one has rn∘sn+1=rn+1r_n\circ s_{n+1}=r_{n+1}. Thus induction gives s1∘⋯∘sn=rns_1\circ\cdots\circ s_n=r_n. Define Qnej=esn(j)Q_ne_j=e_{s_n(j)}. Each QnQ_n is stochastic, with normal unital ∗*-homomorphic adjoint a↦a∘sna\mapsto a\circ s_n. The backward product is therefore

Rnej=ern(j).R_ne_j=e_{r_n(j)}.

Both parities occur among j>nj>n, so the image of all probability vectors is exactly the segment in (16). Its diameter is two, giving δ(Rn)=1\delta(R_n)=1. All coordinates j≤nj\leq n are sent to ana_n. Only the tail mass can differ from eane_{a_n}, and its contribution to an ℓ1\ell^1 difference is at most twice that mass. This proves (17); the triangle inequality proves (18).

For each fixed pp, the right-hand side of (17) tends to zero. Taking p=e0p=e_0 gives Rne0=eanR_ne_0=e_{a_n} exactly, proving the assertion about ordinary limits. For averages,

∥1N∑n=1NRnp−1N∑n=1Nean∥1≤2N∑n=1N∑j>npj⟶0\left\lVert\frac{1}{N}\sum_{n=1}^{N}R_np-\frac{1}{N}\sum_{n=1}^{N}e_{a_n}\right\rVert_1\leq\frac{2}{N}\sum_{n=1}^{N}\sum_{j>n}p_j\longrightarrow0

by Cesàro summation. The last vector average is (1−aˉN)e0+aˉNe1(1-\bar{a}_N)e_0+\bar{a}_Ne_1, where aˉN=N−1∑n=1Nan\bar{a}_N=N^{-1}\sum_{n=1}^{N}a_n. This proves the Cesàro assertion. Including an initial identity term or shifting the average by one term changes the average by at most 2/N2/N and has no effect on the conclusions.

For an=nmod⁡2a_n=n\mathbin{\operatorname{mod}}2, this gives pointwise backward memory loss and mean ergodicity to (e0+e1)/2(e_0+e_1)/2, but not mixing. Uniform mean ergodicity also fails: for j>Nj>N, all RnejR_ne_j, 1≤n≤N1\leq n\leq N, equal ejmod⁡2e_{j\mathbin{\operatorname{mod}}2}, at distance one from the mean limit.

For a failure of mean ergodicity, let N0=0N_0=0, Nk=22kN_k=2^{2^k} for k≥1k\geq1, and prescribe an=kmod⁡2a_n=k\mathbin{\operatorname{mod}}2 for Nk−1<n≤NkN_{k-1}<n\leq N_k. Then Nk−1/Nk→0N_{k-1}/N_k\to0, and the means at times NkN_k approach the alternating block values. Hence the scalar Cesàro means do not converge. Pointwise mean loss of memory still holds by (18) and Cesàro summation. This example therefore separates weak mean memory loss from mean ergodicity even for backward products.

Proof of Theorem 1.2. If MM is finite dimensional, its normal state space is norm compact. Proposition 3.2 upgrades pointwise backward memory loss to uniform memory loss, and Theorem 1.1 gives mixing. Hence (1) implies (2), which trivially implies (3).

If MM is infinite dimensional, lift the stochastic sequence in Theorem 8.4 using Lemma 8.1. With σi=E∗ei\sigma_i=E_*e_i, every nonempty backward product has state image exactly conv⁡{σ0,σ1}\operatorname{conv}\{\sigma_0,\sigma_1\}, a norm-compact segment. Equations (17)–(18) and the isometry give pointwise loss of memory on all normal states of MM. Choose the non-Cesàro-convergent block sequence above. The state σ0\sigma_0 has output σan\sigma_{a_n}, so its trajectory and its Cesàro averages do not converge. Thus (3) fails. This proves (3) implies (1) and the additional assertion.

Remark 8.5 (What compactness does and does not repair). The reservoir example has finite-rank products, a collectively norm-compact set of all nontrivial state outputs, and nested images. None of these assumptions converts pointwise backward memory loss into uniform memory loss. The missing compactness is on the input-state space: optimizing input states may depend on nn and may put all their mass in an unexplored tail. In contrast, once the image diameters actually tend to zero, completeness alone proves Theorem 6.1.

Rates, periodicity, and homogeneous comparisons

Example 9.1 (Arbitrarily slow uniform mixing). On any nontrivial von Neumann algebra choose ω∈SN(M)\omega\in\mathcal{S}_N(M) and a positive nonincreasing sequence bn↓0b_n\downarrow0, with b0=1b_0=1. Define

Tn=bn+1bnid⁡+(1−bn+1bn)Rω.T_n=\frac{b_{n+1}}{b_n}\operatorname{id}+\left(1-\frac{b_{n+1}}{b_n}\right)R_\omega.

The maps commute, and both ordered products satisfy

P0,n=bn id+(1−bn)Rω,δ(P0,n)=bn,P_{0,n}=b_n\,\mathrm{id}+(1-b_n)R_\omega,\qquad\delta(P_{0,n})=b_n,

while

sup⁡φ∈SN(M)∥P0,nφ−ω∥=bncω,cω:=sup⁡φ∈SN(M)∥φ−ω∥>0.\sup_{\varphi\in S_{\mathsf{N}}(\mathcal{M})}\lVert P_{0,n}\varphi-\omega\rVert=b_n c_\omega,\qquad c_\omega:=\sup_{\varphi\in S_{\mathsf{N}}(\mathcal{M})}\lVert\varphi-\omega\rVert>0.

The coefficient equality follows from P0,n∣E0=bn idP_{0,n}|_{E_0}=b_n\,\mathrm{id}. The channel can therefore have any prescribed decreasing positive relaxation profile. Choosing bn=1/(n+1)b_n=1/(n+1) proves that uniform mixing need not be exponential for inhomogeneous processes. Choosing bn=1/log⁡(n+e)b_n=1/\log(n+e), for which b0=1b_0=1, gives a still slower example.

For Md(C)M_d(\mathbb{C}) and ω=I/d\omega=I/d, the correct constant is cω=2(1−1/d)c_\omega=2(1-1/d), not two. Indeed, diagonalize ρ\rho and write ρ=∑jλjPj\rho=\sum_j\lambda_jP_j; convexity bounds ∥ρ−I/d∥1\lVert\rho-I/d\rVert_1 by the pure-state value 2(1−1/d)2(1-1/d), and every pure state attains that value. In particular no state can be orthogonal to the full support of I/dI/d.

Example 9.2 (Uniform mean ergodicity without mixing). On the two-state algebra C2\mathbb{C}^2, let Ue0=e1Ue_0=e_1, Ue1=e0Ue_1=e_0. Its invariant state is ω=(e0+e1)/2\omega=(e_0+e_1)/2 and

sup⁡p∥1N∑k=0N−1Ukp−ω∥1≤1N.\sup_p\left\lVert\frac{1}{N}\sum_{k=0}^{N-1}U^kp-\omega\right\rVert_1\le\frac{1}{N}.

For even NN the average is exactly ω\omega; for odd NN it differs by (p−ω)/N(p-\omega)/N. Nevertheless Une0U^ne_0 alternates and does not converge, and ∥Une0−Une1∥1=2\lVert U^ne_0-U^ne_1\rVert_1=2 for every nn. Thus uniform mean ergodicity, and even uniform mean loss of memory, do not imply mixing or pointwise memory loss.

On M2(C)M_2(\mathbb{C}) the channel

T(ρ)=(ρ1100ρ00)T(\rho)=\begin{pmatrix}\rho_{11}&0\\0&\rho_{00}\end{pmatrix}

(with indices 0,10,1) gives the same example after deleting off-diagonal entries at the first step. It is completely positive: use Kraus operators ∣1⟩⟨0∣|1\rangle\langle0| and ∣0⟩⟨1∣|0\rangle\langle1|. It has invariant density I/2I/2. A two-projection version of Lemma 8.1, with normal states σj\sigma_j satisfying σj(pj)=1\sigma_j(p_j)=1, lifts the classical example to every von Neumann algebra with two nonzero orthogonal projections summing to one.

Theorem 9.3 (Homogeneous benchmark). Let T=Φ∗T=\Phi_* be a single normal channel predual.

  1. Uniform mixing of (Tn)(T^n), uniform loss of memory, and the existence of k≥1k\ge1 with δ(Tk)<1\delta(T^k)<1 are equivalent. Whenever they hold there is a unique invariant normal state ω\omega, and the convergence is exponential.

  1. If an invariant normal state ω\omega is given, pointwise loss of memory is equivalent to mixing to ω\omega; mean loss of memory is equivalent to mean ergodicity to ω\omega.

  1. Mean ergodicity to one normal state is equivalent to the conjunction

SN(M)∩Fix⁡(T)≠∅,Fix⁡(Φ)=C1.S_{\mathsf{N}}(\mathcal{M})\cap\operatorname{Fix}(T)\ne\varnothing,\qquad\operatorname{Fix}(\Phi)=\mathbb{C}1.

Proof. Uniform mixing implies uniform memory loss. If δ(Tn)→0\delta(T^n) \to0, some power has coefficient less than one. Conversely, if q=δ(Tk)<1q = \delta(T^k) < 1, then δ(Tn)≤q⌊n/k⌋\delta(T^n) \le q^{\lfloor n/k \rfloor}. The state images Tn(SN(M))T^n(\mathcal{S}_{\mathcal{N}}(\mathcal{M})) are nested, so Theorem 6.1 gives a normal state ω\omega with

sup⁡φ∈SN(M)∥Tnφ−ω∥≤2q⌊n/k⌋.\sup_{\varphi\in\mathcal{S}_{\mathcal{N}}(\mathcal{M})} \|T^n\varphi-\omega\| \le2q^{\lfloor n/k \rfloor}.

If q>0q > 0, this is an exponential bound with rate q1/kq^{1/k} and prefactor 2/q2/q; if q=0q = 0, the products are rank one from time kk onward and an exponential estimate with any chosen rate follows after adjusting the finite prefactor. Passing to the limit in Tn+1φ=T(Tnφ)T^{n+1}\varphi= T(T^n\varphi) proves Tω=ωT\omega= \omega. Any invariant normal state must equal this limit. This proves (1). Part (2) is Proposition 6.4.

For (3), suppose ANf→f(1)ωA_N f \to f(1)\omega on states and hence on the complex linear span of states, which is all of M∗\mathcal{M}_*. Since TAN−AN=(TN−id)/NTA_N-A_N=(T^N-\mathrm{id})/N, Tω=ωT\omega=\omega. If Φx=x\Phi x=x, then φ(x)=(ANφ)(x)→ω(x)\varphi(x)=(A_N\varphi)(x)\to\omega(x) for every normal state, so x=ω(x)1x=\omega(x)1.

Conversely, let Tω=ωT\omega=\omega and Fix⁡(Φ)=C1\operatorname{Fix}(\Phi)=\mathbb{C}1. Hahn–Banach annihilator duality gives

ran⁡(id−T)‾∥⋅∥={f∈M∗:f(x)=0 for all x∈Fix⁡(Φ)}={f∈M∗:f(1)=0}.\overline{\operatorname{ran}(\mathrm{id}-T)}^{\|\cdot\|}=\{f\in\mathcal{M}_*:f(x)=0\text{ for all }x\in\operatorname{Fix}(\Phi)\}=\{f\in\mathcal{M}_*:f(1)=0\}.

On ran⁡(id−T)\operatorname{ran}(\mathrm{id}-T), telescoping gives AN(id−T)g=(g−TNg)/N→0A_N(\mathrm{id}-T)g=(g-T^Ng)/N\to0. Since ∥AN∥≤1\|A_N\|\le1, this extends to the closure. Decomposing f=f(1)ω+(f−f(1)ω)f=f(1)\omega+(f-f(1)\omega) gives ANf→f(1)ωA_Nf\to f(1)\omega.

The uniform/exponential part is a standard Dobrushin consequence, recorded here for comparison with [27, 28]. It cannot be claimed as a new homogeneous theorem. Its relevance is that an inhomogeneous prefix bound less than one has no corresponding submultiplicative-power mechanism, as Example 8.3 demonstrates.

Example 9.4 (Unique invariant normal state is not enough). On ℓ1(N0)\ell^1(\mathbb{N}_0) define Ve0=e0Ve_0=e_0 and Vej=ej+1Ve_j=e_{j+1} for j≥1j\ge1. The only invariant probability vector is e0e_0: invariance first forces the mass at coordinate one to be zero, and then recursively forces all positive coordinates to have zero mass. But

1N∑k=0N−1Vke1=1N∑j=1Nej\frac{1}{N}\sum_{k=0}^{N-1}V^ke_1=\frac{1}{N}\sum_{j=1}^{N}e_j

has no ℓ1\ell^1 limit. Every coordinate tends to zero while its norm remains one. Thus uniqueness of a normal invariant state does not imply mean ergodicity in infinite dimension. Here Fix⁡(V∗)\operatorname{Fix}(V^*) consists of bounded functions constant on {1,2,…}\{1,2,\ldots\} with an independent value at zero; it is two dimensional, in agreement with Theorem 9.3(3). This example also lifts to any infinite-dimensional σ\sigma-finite M\mathcal{M}: every invariant state of the lifted predual lies in E∗ℓ1E_*^{\ell^1}, where the assertion reduces to the classical one.

State-preserving Hilbert-space criteria

The counterexamples above do not claim that all distinctions persist under every additional state-preserving hypothesis. A useful positive criterion is available without a trace, and it illustrates what a common faithful normal state adds.

Proposition 10.1 (GNS contraction criterion). Suppose ω\omega is a faithful normal state with ωΦn=ω\omega\Phi_n=\omega for every nn. On the GNS Hilbert space L2(M,ω)L^2(\mathcal{M},\omega), with cyclic vector Ω\Omega, define

Vn(xΩ)=Φn(x)Ω.V_n(x\Omega)=\Phi_n(x)\Omega.

Then VnV_n extends to a contraction, fixes Ω\Omega, and preserves Ω⊥\Omega^\perp. If there are 0≤cn≤10\le c_n\le1 such that ∥Vn∣Ω⊥∥≤cn\lVert V_n|_{\Omega^\perp}\rVert\le c_n, then for either observable product Ψm,n\Psi_{m,n}, in either order,

∥(Ψm,n(x)−ω(x)1)Ω∥≤(∏j=mn−1cj)∥(x−ω(x)1)Ω∥.(19)\lVert(\Psi_{m,n}(x)-\omega(x)1)\Omega\rVert\le\left(\prod_{j=m}^{n-1}c_j\right)\lVert(x-\omega(x)1)\Omega\rVert. \tag*{(19)}

If ∏j=mn−1cj→0\prod_{j=m}^{n-1}c_j\to0 for every fixed mm, both corresponding state products are mixing in predual norm to ω\omega. The conclusion need not be uniform over all normal initial states.

Proof. Kadison–Schwarz and invariance give

∥Φn(x)Ω∥2=ω(Φn(x)∗Φn(x))≤ω(Φn(x∗x))=ω(x∗x).\lVert\Phi_n(x)\Omega\rVert^2=\omega(\Phi_n(x)^*\Phi_n(x))\le\omega(\Phi_n(x^*x))=\omega(x^*x).

Also ⟨Ω,VnxΩ⟩=ω(x)\langle\Omega,V_nx\Omega\rangle=\omega(x), so VnV_n fixes Ω\Omega and preserves its orthogonal complement. Iteration proves (19) for either order.

It remains to justify the predual conclusion, which is stronger than mere convergence of each fixed correlation. In this faithful normal GNS representation, Ω\Omega is cyclic and separating for M\mathcal{M} and cyclic for its commutant M′\mathcal{M}'. The linear span of functionals

fa(x)=⟨a′Ω,xΩ⟩,a′∈M′,f_a(x)=\langle a'\Omega,x\Omega\rangle,\qquad a'\in\mathcal{M}',

is norm dense in M∗\mathcal{M}_*. Here is a direct verification. Vector functionals span a norm-dense subspace of the predual of a faithful normal representation. Approximate each pair of vectors by b′Ωb'\Omega, c′Ωc'\Omega, using cyclicity for M′\mathcal{M}'. The corresponding coefficient is ⟨b′Ω,xc′Ω⟩=⟨c′∗b′Ω,xΩ⟩\langle b'\Omega,xc'\Omega\rangle=\langle c'^*b'\Omega,x\Omega\rangle, with the inner product taken linear in the second variable. Approximation in vector norm implies approximation in functional norm, proving the density.

For faf_a, Cauchy–Schwarz and (19) give

∣fa(Ψm,n(x))−fa(1)ω(x)∣≤∥a′Ω∥(∏j=mn−1cj)∥(x−ω(x)1)Ω∥≤2∥a′Ω∥∥x∥∏j=mn−1cj.\lvert f_a(\Psi_{m,n}(x))-f_a(1)\omega(x)\rvert\le\lVert a'\Omega\rVert\left(\prod_{j=m}^{n-1}c_j\right)\lVert(x-\omega(x)1)\Omega\rVert\le2\lVert a'\Omega\rVert\lVert x\rVert\prod_{j=m}^{n-1}c_j.

Taking the supremum over ∥x∥≤1\lVert x\rVert\le1 proves predual convergence for these functionals. The channel preduals and RωR_\omega are contractions, so norm density extends convergence to every normal functional, in particular every normal state.

This proposition requires a quantitative contraction of the GNS implementation on the centered space. Faithfulness or existence of a trace alone does not give such a contraction. No identification M≃M∗\mathcal{M}\simeq\mathcal{M}_*, no tracial cycling of factors, and no use of a density relative to ω\omega has entered the proof. The estimate in (19) is an L2L^2 estimate on observables; its conversion to a predual estimate uses a dense class of functionals whose constants are not uniform over the whole state space.

Inhomogeneous transfer states over infinite bond algebras

This section proves Theorem 1.3. The physical algebra at each site is Md=Md(C)M_d = M_d(\mathbb{C}), and A[p,q]=⨂k=pqMd\mathcal{A}_{[p,q]} = \bigotimes_{k=p}^{q} M_d for p≤qp \le q. For p≤q<q′p \le q < q', the algebra A[p,q]\mathcal{A}_{[p,q]} is identified with A[p,q]⊗1⊆A[p,q′]\mathcal{A}_{[p,q]} \otimes1 \subseteq\mathcal{A}_{[p,q']}. The quasi-local algebra A\mathcal{A} is the norm closure of ⋃NA[1,N]\bigcup_N \mathcal{A}_{[1,N]}. The bond algebra B\mathcal{B} is an arbitrary von Neumann algebra. Since MdM_d is finite dimensional, Md⊗B=Md(B)M_d \otimes\mathcal{B} = M_d(\mathcal{B}) is a von Neumann algebra.

Transfer maps and finite-volume states

Fix normal unital completely positive maps Ek:Md⊗B→B\mathcal{E}_k : M_d \otimes\mathcal{B} \to\mathcal{B}, k≥1k \ge1, and put

Φk(b)=Ek(1⊗b),Tk=(Φk)∗,Bm,n=Tm⋯Tn−1(1≤m≤n),\Phi_k(b) = \mathcal{E}_k(1 \otimes b), \qquad T_k = (\Phi_k)_*, \qquad B_{m,n} = T_m \cdots T_{n-1} \quad(1 \le m \le n),

with Bm,m=id⁡B_{m,m} = \operatorname{id}. Each Φk\Phi_k is a normal unital completely positive map on B\mathcal{B}. For p≤qp \le q define E[p,q]:A[p,q]⊗B→B\mathcal{E}_{[p,q]} : \mathcal{A}_{[p,q]} \otimes\mathcal{B} \to\mathcal{B} recursively by E[p,p]=Ep\mathcal{E}_{[p,p]} = \mathcal{E}_p and

E[p,q](x⊗y⊗b)=Eq(y⊗E[p,q−1](x⊗b)),x∈A[p,q−1], y∈Md, b∈B,(20)\mathcal{E}_{[p,q]}(x \otimes y \otimes b) = \mathcal{E}_q\left(y \otimes\mathcal{E}_{[p,q-1]}(x \otimes b)\right), \qquad x \in\mathcal{A}_{[p,q-1]},\ y \in M_d,\ b \in\mathcal{B}, \tag*{(20)}

extended by linearity. Thus E[p,q]=Eq∘(id⁡Md⊗E[p,q−1])∘θ\mathcal{E}_{[p,q]} = \mathcal{E}_q \circ(\operatorname{id}_{M_d} \otimes\mathcal{E}_{[p,q-1]}) \circ\theta, where θ:A[p,q−1]⊗Md⊗B→Md⊗A[p,q−1]⊗B\theta: \mathcal{A}_{[p,q-1]} \otimes M_d \otimes\mathcal{B} \to M_d \otimes\mathcal{A}_{[p,q-1]} \otimes\mathcal{B} is the normal ∗*-isomorphism that moves the factor MdM_d to the left. The amplification id⁡Md⊗E[p,q−1]\operatorname{id}_{M_d} \otimes\mathcal{E}_{[p,q-1]} of a normal unital completely positive map to Md(⋅)M_d(\cdot) is again normal, unital and completely positive. By induction, every E[p,q]\mathcal{E}_{[p,q]} has these three properties. In this recursion site qq is outermost, and the bond variable bb enters at site pp.

For x∈A[p,q]x \in\mathcal{A}_{[p,q]} write

E[p,q]x(b)=E[p,q](x⊗b),yx=E[1,m](x⊗1)(x∈A[1,m]).\mathcal{E}_{[p,q]}^x(b) = \mathcal{E}_{[p,q]}(x \otimes b), \qquad y_x = \mathcal{E}_{[1,m]}(x \otimes1) \quad(x \in\mathcal{A}_{[1,m]}).

A unital completely positive map is contractive and ∥x⊗b∥=∥x∥∥b∥\lVert x \otimes b\rVert= \lVert x\rVert\lVert b\rVert. Hence E[p,q]x\mathcal{E}_{[p,q]}^x is a normal map with ∥E[p,q]x∥≤∥x∥\lVert\mathcal{E}_{[p,q]}^x\rVert\le\lVert x\rVert, and ∥yx∥≤∥x∥\lVert y_x\rVert\le\lVert x\rVert. If x=x∗x = x^*, then E[p,q]x\mathcal{E}_{[p,q]}^x preserves adjoints.

Definition 11.1. For N≥1N \ge1 and σ∈Sn(B)\sigma\in\mathcal{S}_n(\mathcal{B}), the finite-volume state with right boundary condition σ\sigma is

φNσ(x)=σ(E[1,N](x⊗1)),x∈A[1,N].\varphi_N^\sigma(x) = \sigma\left(\mathcal{E}_{[1,N]}(x \otimes1)\right), \qquad x \in\mathcal{A}_{[1,N]}.

The functional φNσ\varphi_N^\sigma is a state on A[1,N]\mathcal{A}_{[1,N]}, since x↦E[1,N](x⊗1)x \mapsto\mathcal{E}_{[1,N]}(x \otimes1) is unital and positive. The left end of the chain is closed by the unit of B\mathcal{B}, and the boundary state acts after the transfer map of site NN. This is the convention of [36], in which channels are trace preserving in the Schrödinger picture and therefore unital in the Heisenberg picture.

Lemma 11.2 (Block factorisation). For p≤s<qp \le s < q, x∈A[p,s]x \in\mathcal{A}_{[p,s]}, w∈A[s+1,q]w \in\mathcal{A}_{[s+1,q]} and b∈Bb \in\mathcal{B},

E[p,q](x⊗w⊗b)=E[s+1,q]w(E[p,s](x⊗b)),E[s+1,q]1=Φq∘⋯∘Φs+1.(21)\mathcal{E}_{[p,q]}(x \otimes w \otimes b) = \mathcal{E}_{[s+1,q]}^w\left(\mathcal{E}_{[p,s]}(x \otimes b)\right), \qquad\mathcal{E}_{[s+1,q]}^1 = \Phi_q \circ\cdots\circ\Phi_{s+1}. \tag*{(21)}

Consequently, for 1≤m<N1 \le m < N, x∈A[1,m]x \in\mathcal{A}_{[1,m]} and σ∈Sn(B)\sigma\in\mathcal{S}_n(\mathcal{B}),

φNσ(x)=(Bm+1,N+1σ)(yx),yx⊗1=Φm+1(yx).(22)\varphi_N^\sigma(x) = \left(B_{m+1,N+1}\sigma\right)(y_x), \qquad y_{x \otimes1} = \Phi_{m+1}(y_x). \tag*{(22)}

If moreover z∈A[m+ℓ+1,n]z \in\mathcal{A}_{[m+\ell+1,n]} with ℓ≥1\ell\ge1 and n>m+ℓn > m+\ell, then for every ν∈Sn(B)\nu\in\mathcal{S}_{n}(\mathcal{B}),

ν(E[1,n](x⊗1ℓ⊗z⊗1))=(Bm+1,m+ℓ+1g)(yx),g=ν∘E[m+ℓ+1,n]z,(23)\nu\left(\mathcal{E}_{[1,n]}(x \otimes1_{\ell} \otimes z \otimes1)\right)=\left(B_{m+1,m+\ell+1}g\right)(y_x), \qquad g=\nu\circ\mathcal{E}^{z}_{[m+\ell+1,n]}, \tag*{(23)}

where 1ℓ1_{\ell} denotes the unit of A[m+1,m+ℓ]\mathcal{A}_{[m+1,m+\ell]}.

Proof. By linearity it suffices to take w=w′⊗yw=w^{\prime}\otimes y with w′∈A[s+1,q−1]w^{\prime}\in\mathcal{A}_{[s+1,q-1]} and y∈Mdy\in M_d when q>s+1q>s+1. For q=s+1q=s+1, the first identity in (21) is (20). Assume it for q−1q-1. Then

E[p,q](x⊗w′⊗y⊗b)=Eq(y⊗E[p,q−1](x⊗w′⊗b))=Eq(y⊗E[s+1,q−1]w′(E[p,s](x⊗b))).\mathcal{E}_{[p,q]}(x\otimes w^{\prime}\otimes y\otimes b)=\mathcal{E}_{q}\left(y\otimes\mathcal{E}_{[p,q-1]}(x\otimes w^{\prime}\otimes b)\right)=\mathcal{E}_{q}\left(y\otimes\mathcal{E}^{w^{\prime}}_{[s+1,q-1]}\left(\mathcal{E}_{[p,s]}(x\otimes b)\right)\right).

and the right-hand side is E[s+1,q]w′⊗y(E[p,s](x⊗b))\mathcal{E}^{w^{\prime}\otimes y}_{[s+1,q]}\left(\mathcal{E}_{[p,s]}(x\otimes b)\right) by (20) applied on [s+1,q][s+1,q]. The formula for E[s+1,q]1\mathcal{E}^{1}_{[s+1,q]} follows by the same induction, since Eq(1⊗b)=Φq(b)\mathcal{E}_{q}(1\otimes b)=\Phi_{q}(b).

For (22), apply (21) with p=1p=1, s=ms=m, q=Nq=N, w=1w=1 and b=1b=1. This gives

E[1,N](x⊗1⊗1)=ΦN∘⋯∘Φm+1(yx).\mathcal{E}_{[1,N]}(x\otimes1\otimes1)=\Phi_{N}\circ\cdots\circ\Phi_{m+1}(y_x).

The predual of ΦN∘⋯∘Φm+1\Phi_{N}\circ\cdots\circ\Phi_{m+1} is Tm+1⋯TN=Bm+1,N+1T_{m+1}\cdots T_{N}=B_{m+1,N+1}. The case N=m+1N=m+1 gives yx⊗1=Φm+1(yx)y_{x\otimes1}=\Phi_{m+1}(y_x). For (23), apply (21) first with s=m+ℓs=m+\ell, w=zw=z, and then with s=ms=m, w=1ℓw=1_{\ell}. This gives

E[1,n](x⊗1ℓ⊗z⊗1)=E[m+ℓ+1,n]z(Φm+ℓ∘⋯∘Φm+1(yx)).\mathcal{E}_{[1,n]}(x\otimes1_{\ell}\otimes z\otimes1)=\mathcal{E}^{z}_{[m+\ell+1,n]}\left(\Phi_{m+\ell}\circ\cdots\circ\Phi_{m+1}(y_x)\right).

Applying ν\nu and passing to preduals completes the proof.

Thermodynamic limit and decay of correlations

Proof of Theorem 1.3. By Theorem 6.3, applied to the sequence (Tk)k≥1(T_k)_{k\ge1}, there is a unique family (ωk)k≥1(\omega_k)_{k\ge1} of normal states of B\mathcal{B} with ωk=Tkωk+1\omega_k=T_k\omega_{k+1}, and

sup⁡σ∈Sn(B)∥Bk,N+1σ−ωk∥≤2δ(Bk,N+1).(24)\sup_{\sigma\in\mathcal{S}_{n}(\mathcal{B})}\left\|B_{k,N+1}\sigma-\omega_k\right\|\le2\delta(B_{k,N+1}). \tag*{(24)}

For x∈A[1,m]x\in\mathcal{A}_{[1,m]} define φ∞(x)=ωm+1(yx)\varphi_{\infty}(x)=\omega_{m+1}(y_x). This is consistent: if x∈A[1,m]x\in\mathcal{A}_{[1,m]} is regarded as x⊗1∈A[1,m+1]x\otimes1\in\mathcal{A}_{[1,m+1]}, then yx⊗1=Φm+1(yx)y_{x\otimes1}=\Phi_{m+1}(y_x) by Lemma 11.2. Hence

ωm+2(Φm+1(yx))=(Tm+1ωm+2)(yx)=ωm+1(yx).\omega_{m+2}\left(\Phi_{m+1}(y_x)\right)=(T_{m+1}\omega_{m+2})(y_x)=\omega_{m+1}(y_x).

Thus φ∞\varphi_{\infty} is a well-defined linear functional on ⋃mA[1,m]\bigcup_m\mathcal{A}_{[1,m]}. It is positive and unital on each A[1,m]\mathcal{A}_{[1,m]}, because x↦yxx\mapsto y_x is unital and positive. It therefore has norm one and extends uniquely to a state on A\mathcal{A}.

By (22) and (24), for x∈A[1,m]x\in\mathcal{A}_{[1,m]},

∣φNσ(x)−φ∞(x)∣=∣(Bm+1,N+1σ−ωm+1)(yx)∣≤2δ(Bm+1,N+1)∥x∥.\left|\varphi_N^{\sigma}(x)-\varphi_{\infty}(x)\right|=\left|\left(B_{m+1,N+1}\sigma-\omega_{m+1}\right)(y_x)\right|\le2\delta(B_{m+1,N+1})\|x\|.

This is the first estimate, since Bm+1,N+1=Tm+1⋯TNB_{m+1,N+1}=T_{m+1}\cdots T_N. Uniqueness follows because the right-hand side tends to zero for every fixed mm: a state with the same property has the same restriction to every A[1,m]\mathcal{A}_{[1,m]}, and these restrictions determine a state on A\mathcal{A}.

For the clustering bound, let n=m+ℓ+rn=m+\ell+r and put g=ωn+1∘E[m+ℓ+1,n]zg=\omega_{n+1}\circ\mathcal{E}^{z}_{[m+\ell+1,n]}, a normal functional with ∥g∥≤∥z∥\|g\|\le\|z\|. By definition of φ∞\varphi_{\infty} and (23),

φ∞(x⊗1ℓ⊗z)=(Bm+1,m+ℓ+1g)(yx),φ∞(z)=g(1),\varphi_{\infty}(x\otimes1_{\ell}\otimes z)=\left(B_{m+1,m+\ell+1}g\right)(y_x),\qquad\varphi_{\infty}(z)=g(1),

where the second identity is (23) with x=1x=1, since y1=1y_1=1 and (Bm+1,m+ℓ+1g)(1)=g(1)(B_{m+1,m+\ell+1}g)(1)=g(1). Also φ∞(x)=ωm+1(yx)=(Bm+1,m+ℓ+1ωm+ℓ+1)(yx)\varphi_\infty(x)=\omega_{m+1}(y_x)=(B_{m+1,m+\ell+1}\omega_{m+\ell+1})(y_x). Hence

φ∞(x⊗1ℓ⊗z)−φ∞(x)φ∞(z)=(Bm+1,m+ℓ+1h)(yx),h=g−g(1)ωm+ℓ+1.\varphi_\infty(x\otimes1_\ell\otimes z)-\varphi_\infty(x)\varphi_\infty(z)=(B_{m+1,m+\ell+1}h)(y_x),\qquad h=g-g(1)\omega_{m+\ell+1}.

Suppose first that z=z∗z=z^*. Then gg is self-adjoint, h(1)=0h(1)=0, and ∥h∥≤2∥g∥≤2∥z∥\lVert h\rVert\leq2\lVert g\rVert\leq2\lVert z\rVert. Therefore h∈E0(B)h\in E_0(\mathcal{B}) and

∣(Bm+1,m+ℓ+1h)(yx)∣≤δ(Bm+1,m+ℓ+1)∥h∥∥yx∥≤2δ(Bm+1,m+ℓ+1)∥x∥∥z∥.\left|(B_{m+1,m+\ell+1}h)(y_x)\right|\leq\delta(B_{m+1,m+\ell+1})\lVert h\rVert\lVert y_x\rVert\leq2\delta(B_{m+1,m+\ell+1})\lVert x\rVert\lVert z\rVert.

For general zz, write z=z1+iz2z=z_1+iz_2 with zjz_j self-adjoint and ∥zj∥≤∥z∥\lVert z_j\rVert\leq\lVert z\rVert, and use linearity in zz. This gives the factor 4. Finally, Bm+1,m+ℓ+1=Tm+1⋯Tm+ℓB_{m+1,m+\ell+1}=T_{m+1}\cdots T_{m+\ell}.

Corollary 11.3 (Quantitative regimes). Under the notation of Theorem 1.3, put βk=β(Tk)\beta_k=\beta(T_k).

  1. If ∑kβk=∞\sum_k\beta_k=\infty, the hypothesis of Theorem 1.3 holds, with

∥φNσ∣A[1,m]−φ∞∣A[1,m]∥≤2∏k=m+1N(1−βk).\left\|\varphi_N^\sigma|_{\mathcal{A}_{[1,m]}}-\varphi_\infty|_{\mathcal{A}_{[1,m]}}\right\|\leq2\prod_{k=m+1}^{N}(1-\beta_k).
  1. If there are L≥1L\geq1 and q<1q<1 with sup⁡m≥0δ(Tm+1⋯Tm+L)≤q\sup_{m\geq0}\delta(T_{m+1}\cdots T_{m+L})\leq q, then boundary dependence and correlations decay exponentially, uniformly in position:

∥φNσ∣A[1,m]−φ∞∣A[1,m]∥≤2q⌊(N−m)/L⌋,\left\|\varphi_N^\sigma|_{\mathcal{A}_{[1,m]}}-\varphi_\infty|_{\mathcal{A}_{[1,m]}}\right\|\leq2q^{\lfloor(N-m)/L\rfloor},
∣φ∞(x⊗1ℓ⊗z)−φ∞(x)φ∞(z)∣≤4q⌊ℓ/L⌋∥x∥∥z∥.\left|\varphi_\infty(x\otimes1_\ell\otimes z)-\varphi_\infty(x)\varphi_\infty(z)\right|\leq4q^{\lfloor\ell/L\rfloor}\lVert x\rVert\lVert z\rVert.

Proof. Part (1) follows from Theorem 1.3 and (11), since δ(Tm+1⋯TN)≤∏k=m+1N(1−βk)\delta(T_{m+1}\cdots T_N)\leq\prod_{k=m+1}^{N}(1-\beta_k) and the product tends to zero for every fixed mm when ∑kβk=∞\sum_k\beta_k=\infty. For part (2), apply Proposition 7.2 to the backward products of (Tk)k≥1(T_k)_{k\geq1}: the window coefficient satisfies h(ℓ)≤q⌊ℓ/L⌋h(\ell)\leq q^{\lfloor\ell/L\rfloor}, and both estimates of Theorem 1.3 are bounded by hh evaluated at the length of the relevant product.

Connection with matrix product states

Let B=MD\mathcal{B}=M_D and let K1[k],…,Kd[k]∈MDK^{[k]}_1,\ldots,K^{[k]}_d\in M_D satisfy the left-canonical condition ∑iKi[k]†Ki[k]=1\sum_i K_i^{[k]\dagger}K_i^{[k]}=1. Define the isometry Vk:CD→Cd⊗CDV_k:\mathbb{C}^D\to\mathbb{C}^d\otimes\mathbb{C}^D by Vkξ=∑j∣j⟩⊗Kj[k]ξV_k\xi=\sum_j|j\rangle\otimes K_j^{[k]}\xi, and set

Ek(y⊗b)=Vk†(y⊗b)Vk=∑i,j=1d⟨i∣y∣j⟩Ki[k]†bKj[k].\mathcal{E}_k(y\otimes b)=V_k^\dagger(y\otimes b)V_k=\sum_{i,j=1}^{d}\langle i|y|j\rangle K_i^{[k]\dagger}bK_j^{[k]}.

Then Ek\mathcal{E}_k is unital and completely positive, and Φk(b)=∑iKi[k]†bKi[k]\Phi_k(b)=\sum_i K_i^{[k]\dagger}bK_i^{[k]} is the Heisenberg dual of the trace-preserving channel ρ↦∑iKi[k]ρKi[k]†\rho\mapsto\sum_i K_i^{[k]}\rho K_i^{[k]\dagger}. Iterating (20) gives, for b=1b=1,

E[1,N](∣i1⋯iN⟩⟨j1⋯jN∣⊗1)=KiN[N]†⋯Ki1[1]†Kj1[1]⋯KjN[N].\mathcal{E}_{[1,N]}\left(|i_1\cdots i_N\rangle\langle j_1\cdots j_N|\otimes1\right)=K_{i_N}^{[N]\dagger}\cdots K_{i_1}^{[1]\dagger}K_{j_1}^{[1]}\cdots K_{j_N}^{[N]}.

For a boundary density σ\sigma, therefore,

φNσ(∣i1⋯iN⟩⟨j1⋯jN∣)=Tr⁡(Kj1[1]⋯KjN[N] σ KiN[N]†⋯Ki1[1]†).\varphi_N^\sigma(|i_1\cdots i_N\rangle\langle j_1\cdots j_N|)=\operatorname{Tr}\left(K_{j_1}^{[1]}\cdots K_{j_N}^{[N]}\,\sigma\,K_{i_N}^{[N]\dagger}\cdots K_{i_1}^{[1]\dagger}\right).

This is the finite-volume matrix product expression with right boundary σ\sigma. The limiting formula φ∞(x)=ωm+1(yx)\varphi_\infty(x)=\omega_{m+1}(y_x) is the boundary-state representation of [36 Theorem 1.3], with (ωk)(\omega_k) the backward entrance law. In this finite-dimensional case, Theorem 5.2 reduces the minorization hypothesis to the Markov–Dobrushin condition of [36, 35], and the coefficients δ(Tm+1⋯TN)\delta(T_{m+1}\cdots T_N) are the product trace-Dobrushin coefficients of [31].

Example 11.4 (An infinite bond algebra). Let KK be an infinite-dimensional separable Hilbert space and B=B(K)\mathcal{B}=B(K). Fix unitaries Uk:K→Cd⊗KU_k:K\to\mathbb{C}^d\otimes K, a normal state ν\nu of B\mathcal{B}, states ψk\psi_k of MdM_d, and numbers pk∈[0,1]p_k\in[0,1]. Define

Ek(y⊗b)=(1−pk)Uk∗(y⊗b)Uk+pkψk(y)ν(b)1.\mathcal{E}_k(y\otimes b)=(1-p_k)U_k^*(y\otimes b)U_k+p_k\psi_k(y)\nu(b)1.

Each Ek\mathcal{E}_k is normal, unital and completely positive. Moreover

Tk=(1−pk)Tk′+pkRν,Tk′=(Uk∗(1⊗⋅)Uk)∗,T_k=(1-p_k)T'_k+p_kR_\nu,\qquad T'_k=(U_k^*(1\otimes\mathord{\cdot})U_k)_*,

so Tkφ≥pkνT_k\varphi\ge p_k\nu for every normal state φ\varphi, and β(Tk)≥pk\beta(T_k)\ge p_k. If ∑kpk=∞\sum_k p_k=\infty, Corollary 11.3(1) applies. If pk≥p>0p_k\ge p>0 for all kk, it gives exponential decay of boundary dependence and of correlations, with L=1L=1 and q=1−pq=1-p. Here the bond algebra is a type I∞_\infty factor. More generally, replace Uk∗(⋅)UkU_k^*(\mathord{\cdot})U_k by a normal ∗*-isomorphism πk:Md⊗B→B\pi_k:M_d\otimes\mathcal{B}\to\mathcal{B}. Such isomorphisms exist whenever B\mathcal{B} is properly infinite, in particular for every type III factor, and when B\mathcal{B} is the hyperfinite II1_1 factor, whose fundamental group is R+∗\mathbb{R}_+^*. The same conclusions hold verbatim.

Remark 11.5 (Uniformity is necessary). The estimate in Theorem 1.3 is uniform in the boundary state. The following example shows that pointwise mixing of the backward products does not suffice for boundary independence of the local states. Let d≥2d\ge2, B=ℓ∞(N0)\mathcal{B}=\ell^\infty(\mathbb{N}_0), and let QkQ_k, sks_k, rkr_k be the maps of Theorem 8.4 with an≡0a_n\equiv0. Write ϵi(y)=⟨i∣y∣i⟩\epsilon_i(y)=\langle i|y|i\rangle for i∈{0,1}i\in\{0,1\}. Define

(E1(y⊗b))(j)=ϵj mod 2(y)b(s1(j)),Ek(y⊗b)=ϵ0(y)Qk∗b(k≥2).\left(\mathcal{E}_1(y\otimes b)\right)(j)=\epsilon_{j\bmod2}(y)b(s_1(j)),\qquad\mathcal{E}_k(y\otimes b)=\epsilon_0(y)Q_k^*b\quad(k\ge2).

Every coordinate of E1\mathcal{E}_1 is the normal product state ϵj mod 2⊗ev⁡s1(j)\epsilon_{j\bmod2}\otimes\operatorname{ev}_{s_1(j)}. A unital positive map into a commutative von Neumann algebra whose coordinates are normal states is normal and completely positive. Thus E1\mathcal{E}_1 is admissible, and so is each Ek=ϵ0⊗Qk∗\mathcal{E}_k=\epsilon_0\otimes Q_k^*. Moreover Φk=Qk∗\Phi_k=Q_k^* for all kk.

For x∈Mdx\in M_d at site 1 and σ=ej\sigma=e_j, Lemma 11.2 gives

φNej(x)=(Q2⋯QNej)(yx)=ϵt mod 2(x),t=s2∘⋯∘sN(j),\varphi_N^{e_j}(x)=\left(Q_2\cdots Q_Ne_j\right)(y_x)=\epsilon_{t\bmod2}(x),\qquad t=s_2\circ\cdots\circ s_N(j),

because yx(t)=ϵt mod 2(x)y_x(t)=\epsilon_{t\bmod2}(x). Since t∈{u1,v1}={4,5}t\in\{u_1,v_1\}=\{4,5\} and r1(t)=t mod 2r_1(t)=t\bmod2 for t>1t>1, we get t mod 2=r1(t)=rN(j)t\bmod2=r_1(t)=r_N(j). With an≡0a_n\equiv0, rN(j)=0r_N(j)=0 for j≤Nj\le N and rN(j)=j mod 2r_N(j)=j\bmod2 for j>Nj>N. Hence, for every fixed probability vector pp,

φNp(x)⟶ϵ0(x).\varphi_N^p(x)\longrightarrow\epsilon_0(x).

The relevant backward products are mixing: Q1⋯QNp→e0Q_1\cdots Q_Np \to e_0 by Theorem 8.4, and Q2⋯QNp→e4Q_2\cdots Q_Np \to e_4, because t=4t=4 whenever rN(j)=0r_N(j)=0, in particular for all j≤Nj\le N. But for the boundary sequence σN=e2N+3\sigma_N=e_{2N+3} one has φNσ(x)=ϵ1(x)\varphi_N^\sigma(x)=\epsilon_1(x) for every NN, so the local state at site 1 depends on the boundary in the limit.

The same example lives on every infinite-dimensional σ\sigma-finite bond algebra. With J,EJ,E as in Lemma 8.1, replace Ek\mathcal{E}_k by J∘Ek∘(id⁡Md⊗E)J\circ\mathcal{E}_k\circ(\operatorname{id}_{M_d}\otimes E). Since EJ=id⁡EJ=\operatorname{id}, the block maps become J∘E[1,N]∘(id⁡⊗E)J\circ\mathcal{E}_{[1,N]}\circ(\operatorname{id}\otimes E), and the finite-volume states become φNJ∗∘σ\varphi_N^{J^*\circ\sigma}.

Consequences and limits of the extension

The preceding results give the following precise conclusions. First, uniform backward memory loss implies uniform convergence to a normal state on every von Neumann algebra, with no compactness hypothesis. Applying this from every starting time yields a unique normal entrance law and quantitative independence of terminal boundary conditions. This extends the backward contraction mechanism of the matrix setting at the uniform level.

Second, the corresponding pointwise assertion is universally valid exactly in finite dimension. In infinite dimension, nested, collectively compact state images and even finite-rank products do not suffice: the reservoir construction in Theorem 8.4 gives a counterexample on every infinite-dimensional σ\sigma-finite algebra. Thus the obstruction is not the distinction between types I, II, and III, nor merely the lack of compact output sets. It is the failure to control moving initial states when passing from pointwise to uniform convergence.

Third, the attained coefficient β(T)\beta(T) makes the common-lower-bound criterion meaningful in the absence of rank-one projections, traces, or normal pure states. It provides an upper bound on the optimal contraction coefficient δ(T)\delta(T), not an equality in general and not a CP-order decomposition. The construction avoids a nonexistent noncommutative lattice infimum and retains normality by domination.

Fourth, the backward theory applies directly to quantum spin chains. For transfer maps with values in an arbitrary von Neumann algebra, uniform backward memory loss gives a unique thermodynamic limit, uniformly in the boundary condition. The same product coefficients bound the decay of correlations, which recovers the finite-dimensional matrix product state picture of [36, 31] for an infinite-dimensional bond algebra. Remark 11.5 shows that this uniformity cannot be relaxed to pointwise mixing.

Finally, mixing, ergodicity, and memory loss address three separate questions: whether a trajectory converges, whether its time average converges, and whether initial preparations become indistinguishable. Composition order supplies a fourth distinction, and uniformity over starting times supplies a fifth. The homogeneous benchmark removes some of these distinctions through the power structure; time-dependent channels do not have that mechanism. The examples above identify exactly which additional assumptions are needed in the stated theorems, rather than attributing convergence to an algebra type or to unproved compactness intuitions.

References

  1. [1]L. Accardi, On the noncommutative Markov property, Functional Analysis and Its Applications 9 (1975), 1–8.
  2. [2]L. Accardi, F. Fidaleo, Entangled Markov chains, Annali di Matematica Pura ed Applicata 184 (2005), 327–346. https://doi.org/10.1007/s10231-004-0118-4.
  3. [3]L. Accardi, Y. G. Lu, A. Souissi, A Markov–Dobrushin inequality for quantum channels, Open Systems & Information Dynamics 28 (2021), 2150018. https://doi.org/10.1142/S1230161221500189.
  4. [4]L. Accardi, A. Souissi, E. G. Soueidy, Quantum Markov chains: a unification approach, Infinite Dimensional Analysis, Quantum Probability and Related Topics 23 (2020), 2050016. https://doi.org/10.1142/S0219025720500162.
  5. [5]V. V. Albert, Asymptotics of quantum channels: conserved quantities, an adiabatic limit, and matrix product states, Quantum 3 (2019), 151. https://doi.org/10.22331/q-2019-06-06-151.
  6. [6]T. Benoist, M. Fraas, Y. Pautrat, C. Pellegrini, Invariant measure for quantum trajectories, Probability Theory and Related Fields 174 (2019), 307–334. https://doi.org/10.1007/s00440-018-0862-9.
  7. [7]L. Bowen, B. Hayes, Y. F. Lin, A multiplicative ergodic theorem for von Neumann algebra valued cocycles, Communications in Mathematical Physics 384 (2021), 1291–1350. https://doi.org/10.1007/s00220-021-04043-9.
  8. [8]L. Bruneau, A. Joye, M. Merkli, Random repeated interaction quantum systems, Communications in Mathematical Physics 284 (2008), 553–581. https://doi.org/10.1007/s00220-008-0580-8.
  9. [9]D. Burgarth, G. Chiribella, V. Giovannetti, P. Perinotti, K. Yuasa, Ergodic and mixing quantum channels in finite dimensions, New Journal of Physics 15 (2013), 073045. https://doi.org/10.1088/1367-2630/15/7/073045.
  10. [10]R. Carbone, A. Jenčová, On period, cycles and fixed points of a quantum channel, Annales Henri Poincaré 21 (2020), 155–188. Preprint: https://arxiv.org/abs/1905.00857.
  11. [11]O. Ekblad, E. Moreno-Nadales, L. Pathirana, Ergodic theorems for quantum trajectories under disordered generalized measurements, arXiv:2501.18014 (2025). https://arxiv.org/abs/2501.18014.
  12. [12]O. Ekblad, J. Schenker, Reducibility theory and ergodic theorems for ergodic quantum processes, arXiv:2406.10982 (2024); first version entitled “Ergodic repeated interaction quantum systems: steady states and reducibility theory”. https://arxiv.org/abs/2406.10982.
  13. [13]M. Fannes, B. Nachtergaele, R. F. Werner, Finitely correlated states on quantum spin chains, Communications in Mathematical Physics 144 (1992), 443–490. https://doi.org/10.1007/BF02099178.
  14. [14]M. Fannes, B. Nachtergaele, R. F. Werner, Finitely correlated pure states, Journal of Functional Analysis 120 (1994), 511–534. https://doi.org/10.1006/jfan.1994.1041.
  15. [15]F. Fidaleo, On strong ergodic properties of quantum dynamical systems, Infinite Dimensional Analysis, Quantum Probability and Related Topics 12 (2009), 551–564. Preprint: https://arxiv.org/abs/0802.2076.
  16. [16]A. Frigerio, M. Verri, Long-time asymptotic properties of dynamical semigroups on W*-algebras, Mathematische Zeitschrift 180 (1982), 275–286. https://doi.org/10.1007/BF01318911.
  17. [17]S. Gaubert, Z. Qu, Dobrushin’s ergodicity coefficient for Markov operators on cones, Integral Equations and Operator Theory 81 (2015), 127–150. https://doi.org/10.1007/s00020-014-2193-2.
  18. [18]I. George, C. Hirche, T. Nuradha, M. M. Wilde, Quantum Doeblin coefficients: interpretations and applications, arXiv:2503.22823 (2025). https://arxiv.org/abs/2503.22823.
  19. [19]C. Hirche, Quantum Doeblin coefficients: a simple upper bound on contraction coefficients, in: Proceedings of the 2024 IEEE International Symposium on Information Theory (ISIT), 2024. Preprint: https://arxiv.org/abs/2405.00105.
  20. [20]M. Junge, Q. Xu, Noncommutative maximal ergodic theorems, Journal of the American Mathematical Society 20 (2007), 385–439. https://doi.org/10.1090/S0894-0347-06-00533-9.
  21. [21]B. Kümmerer, Markov dilations on W*-algebras, Journal of Functional Analysis 63 (1985), 139–177.DOI
  22. [22]B. Kümmerer, H. Maassen, A pathwise ergodic theorem for quantum trajectories, Journal of Physics A: Mathematical and General 37 (2004), 11889–11896. https://doi.org/10.1088/0305-4470/37/49/008.
  23. [23]E. C. Lance, Ergodic theorems for convex sets and operator algebras, Inventiones Mathematicae 37 (1976), 201–214. https://doi.org/10.1007/BF01390319.
  24. [24]A. Marrakchi, S. Vaes, Ergodic states on type III₁ factors and ergodic actions, Journal für die reine und angewandte Mathematik 809 (2024), 247–260. Preprint: https://arxiv.org/abs/2305.14217.
  25. [25]R. Movassagh, J. Schenker, Theory of ergodic quantum processes, Physical Review X 11 (2021), 041001. https://doi.org/10.1103/PhysRevX.11.041001.
  26. [26]R. Movassagh, J. Schenker, An ergodic theorem for quantum processes with applications to matrix product states, Communications in Mathematical Physics 395 (2022), 1175–1196. https://doi.org/10.1007/s00220-022-04448-0.
  27. [27]F. Mukhamedov, The Dobrushin ergodicity coefficient and the ergodicity of noncommutative Markov chains, Journal of Mathematical Analysis and Applications 408 (2013), 364–373. https://doi.org/10.1016/j.jmaa.2013.06.022.
  28. [28]F. Mukhamedov, Ergodic properties of nonhomogeneous Markov chains defined on ordered Banach spaces with a base, Acta Mathematica Hungarica 147 (2015), 294–323. Preprint: https://arxiv.org/abs/1311.0624.
  29. [29]F. Mukhamedov, A. Al-Rawashdeh, Generalized Dobrushin ergodicity coefficient and ergodicities of non-homogeneous Markov chains, Banach Journal of Mathematical Analysis 16 (2022), no. 1. Preprint: https://arxiv.org/abs/2001.07703.
  30. [30]B. Nelson, E. B. Roon, Ergodic quantum processes on finite von Neumann algebras, Journal of Functional Analysis 287 (2024), 110485. https://doi.org/10.1016/j.jfa.2024.110485.
  31. [31]L. Pathirana, Asymptotic replacement for quantum channel products with applications to inhomogeneous matrix product states, arXiv:2605.00157 (2026). https://arxiv.org/abs/2605.00157.
  32. [32]L. Pathirana, J. Schenker, Asymptotic behavior of random time-inhomogeneous Markovian quantum dynamics, arXiv:2509.08924 (2025). https://arxiv.org/abs/2509.08924.
  33. [33]D. Pérez-García, F. Verstraete, M. M. Wolf, J. I. Cirac, Matrix product state representations, Quantum Information and Computation 7 (2007), 401–430. Preprint: https://arxiv.org/abs/quant-ph/0608197.
  34. [34]M. Raginsky, Strictly contractive quantum channels and physically realizable quantum computers, Physical Review A 65 (2002), 032306. https://doi.org/10.1103/PhysRevA.65.032306.
  35. [35]A. Souissi, A. Barhoumi, An exponential mixing condition for quantum channels: application to matrix product states, Quantum Information Processing 24 (2025), 150. https://doi.org/10.1007/s11128-025-04762-1.
  36. [36]A. Souissi, Ergodic theory of inhomogeneous quantum processes, Quantum Information Processing 25 (2026), 220. https://doi.org/10.1007/s11128-026-05228-8.
  37. [37]F. J. Yeadon, Ergodic theorems for semifinite von Neumann algebras II, Mathematical Proceedings of the Cambridge Philosophical Society 88 (1980), 135–147.DOI

Paper details

vlnc:W_ng6bj7ms
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 1

downloads · 0

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

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents