Generalized Poincaré inequality for quantum Markov semigroups Marius Junge, Jia Wang Abstract We prove a noncommutative ( p , p ) (p,p) -Poincaré inequality for trace-symmetric quantum Markov semigroups on tracial von Neumann algebras, assuming only the existence of a spectral gap. Extending semi-commutative results of Huang and Tropp huang2021poincare , our argument uses Markov dilations to obtain chain-rule estimates for Dirichlet forms and employs amalgamated free products to define an appropriate noncommutative derivation. We further generalize the argument to non-tracial σ \sigma -finite von Neumann algebras under the weaker assumption of GNS-detailed balance, using Haagerup’s reduction and Kosaki’s interpolation theorem. As applications, we recover noncommutative Khintchine and sub-exponential concentration inequalities. keywords Poincaré inequalities, quantum Markov semigroups, Markov dilations, amalgamated free products, Haagerup L^{p} spaces, Kosaki L^{p} interpolation, Haagerup reduction, concentration inequality. 1 Introduction The Poincaré inequality, first introduced in poincare1890 , is a central tool in analysis, probability, and geometry. In its classical Euclidean form, let \Omega\subseteq\mathbb{R}^{d} be a bounded, connected open set with Lipschitz boundary. Then there exists a constant C>0 , depending only on \Omega , such that for every smooth function with compact support, \int_{\Omega}\bigl|f(x)-f_{\Omega}\bigr|^{2}dx\leq C\int_{\Omega}\|\nabla f(x)\|_{2}^{2}dx, (1) where f_{\Omega}=\frac{1}{|\Omega|}\int_{\Omega}f(x)dx denotes the mean of f . More generally, let \mu be a probability measure on a space X , and let \Gamma be the carré du champ/gradient form associated with a symmetric Markov generator L . The Poincaré inequality takes the invariant-measure form \mathrm{Var}_{\mu}(f):=\int_{X}|f-\mu(f)|^{2}d\mu\leq\frac{1}{\alpha}\int_{X}\Gamma(f,f)d\mu, (2) where \alpha>0 is the spectral gap of L on L^{2}(X,\mu) ; see, for example, bakry-emery . An equivalent semigroup formulation is as follows. Let (T_{t})_{t\geq 0}=(e^{-tL})_{t\geq 0} be a conservative, \mu -symmetric Markov semigroup on L^{2}(\mu) with Dirichlet form \mathscr{E}(f,g)=\int_{X}\Gamma(f,g)d\mu . Then ( 2 ) is equivalent to the existence of a spectral gap \alpha>0 , namely \|T_{t}f-\mu(f)\|_{L^{2}(\mu)}\leq e^{-\alpha t}\|f-\mu(f)\|_{L^{2}(\mu)},\qquad t\geq 0, (3) or, equivalently, to the infinitesimal form \mathrm{Var}_{\mu}(f)\leq\frac{1}{\alpha}\mathscr{E}(f,f),\qquad f\in\mathrm{Dom}(L). (4) The equivalences ( 2 )–( 4 ) are standard (see bakry-emery ) and play a fundamental role in quantitative mixing, functional inequalities, and concentration phenomena. In noncommutative analysis, the role of L^{\infty}(X,\mu) is played by a von Neumann algebra (vNa) \mathcal{M} . Let (\mathcal{M},\tau) be a tracial von Neumann algebra with faithful normal tracial state \tau . A (uniformly continuous) quantum Markov semigroup (QMS) on \mathcal{M} is a family \{T_{t}\}_{t\geq 0}\subset\mathrm{CP-}\mathrm{unital} maps satisfying the semigroup property and \tau\circ T_{t}=\tau for all t\geq 0 . Writing T_{t}=e^{-tL} with generator L , one defines the quantum carré du champ (gradient form) \Gamma(x,y)=\tfrac{1}{2}\bigl(L(x^{*})y+x^{*}L(y)-L(x^{*}y)\bigr),\quad x,y\in\mathrm{Dom}(L), (5) and the fixed-point subalgebra \mathcal{N}=\{x\in\mathcal{M}:T_{t}(x)=x\ \text{ for all }t\geq 0\}. (6) Let E:\mathcal{M}\to\mathcal{N} denote the canonical \tau -preserving conditional expectation. Then the (tracial) quantum Poincaré inequality asserts \|x-E(x)\|_{L^{2}(\mathcal{M},\tau)}^{2}\leq\frac{1}{\alpha}\tau\bigl(\Gamma(x,x)\bigr),\qquad\text{for all }x=x^{*}\in\mathrm{Dom}(L), (7) and is equivalent to the existence of an L^{2}(\mathcal{M},\tau) spectral gap \alpha>0 for T_{t} on the orthogonal complement of \mathcal{N} , i.e., \|T_{t}(x)-E(x)\|_{L^{2}(\mathcal{M},\tau)}\leq e^{-\alpha t}\|x-E(x)\|_{L^{2}(\mathcal{M},\tau)},\qquad t\geq 0. (8) Generalized tracial Poincaré inequalities. Motivated by applications beyond the L^{2} setting junge2015noncommutative , we introduce the following family of inequalities. Definition 1.1 (Tracial Poincaré inequality \mathrm{PI}(p,q) ) . Let (\mathcal{M},\tau) be a tracial von Neumann algebra with faithful normal trace \tau , and let \{T_{t}\}_{t\geq 0}:\mathcal{M}\to\mathcal{M} be a QMS with generator L (so T_{t}=e^{-tL} ). Let \Gamma be the gradient form ( 5 ), and let E:\mathcal{M}\to\mathcal{N}:=\{x\in\mathcal{M}:T_{t}(x)=x\ \ \forall t\geq 0\} be the \tau –preserving conditional expectation onto the fixed-point algebra \mathcal{N} . For p,q\in[1,\infty] , we say that the QMS satisfies the Poincaré inequality \mathrm{PI}(p,q) with constant C>0 if, for every self–adjoint x\in\mathrm{Dom}(L) , \|x-E(x)\|_{L^{p}(\mathcal{M},\tau)}\leq C\|\Gamma(x,x)^{1/2}\|_{L^{q}(\mathcal{M},\tau)}. (9) Here we denote by \|\cdot\|_{p} the L^{p} –norm on the tracial von Neumann algebra (\mathcal{M},\tau) . Remark 1.2 . The case (p,q)=(2,2) recovers the standard quantum Poincaré inequality ( 7 ) and is equivalent to an L^{2} spectral gap. In this work, we extend the semi-commutative results of Huang and Tropp huang2021poincare to a fully noncommutative setting for a quantum Markov semigroup with appropriate detailed balance conditions using operator algebraic methods. The first main result we establish in this paper is the following homogeneous (p,p) -type Poincaré inequality with constant O(p) : Theorem 1.3 ( \mathrm{PI}(p,p) ) . Let (\mathcal{M},\tau) be a tracial von Neumann algebra and (T_{t})_{t\geq 0}=e^{-tL} a \tau –symmetric quantum Markov semigroup with fixed-point algebra \mathcal{N} and conditional expectation E:\mathcal{M}\to\mathcal{N} . If \{T_{t}\}_{t\geq 0} has a spectral gap \alpha>0 , then for every self-adjoint x in the domain of L and p=2 or p\geq 3 , it satisfies the Poincaré inequality \mathrm{PI}(p,p) with constant \tfrac{p}{\sqrt{2\alpha}} , i.e., \|x-E(x)\|_{L^{p}(\mathcal{M},\tau)}\leq\frac{p}{\sqrt{2\alpha}}\|\Gamma(x,x)^{1/2}\|_{L^{p}(\mathcal{M},\tau)}. The proof uses the argument of Huang and Tropp as a stepping stone. In particular, they established a chain rule estimate relating the Dirichlet form to the associated gradient form. A further ingredient in their argument is a symmetrization procedure, based on doubling variables in the product space of a Markov process, which leads to key cancellations. Our proof, in addition, identifies the appropriate dilation framework for quantum Markov semigroups, which allows us to derive chain rule estimates directly in the noncommutative setting. Since there are obstacles to generalizing their produce measure symmetrization procedure, we instead formulate it in terms of amalgamated free products, providing a natural setting for noncommutative algebras. Beyond the tracial setting: GNS-detailed balance via Haagerup reduction. Many physically relevant QMS are not tracial but are reversible with respect to a faithful normal state \phi (GNS-symmetry/detailed balance). In that case one works on Haagerup L^{p} spaces haagerup1979lpspaces L^{p}(\mathcal{M}) and uses the GNS inner product, see Section 4.1 for details. Our strategy reduces the non-tracial problem to the tracial one via Haagerup’s reduction (crossed-product) technique. We embed (\mathcal{M},\phi) into a semifinite crossed product (\widehat{\mathcal{M}},\widehat{\phi}) , which admits an increasing sequence of tracial subalgebras (\mathcal{M}_{k},\tau_{k}) . The semigroup is lifted to be \widehat{\phi} -detailed balance on \widehat{\mathcal{M}} and \tau_{k} -symmetric on \mathcal{M}_{k} , where a \mathrm{PI}(p,q) inequality on \widehat{\mathcal{M}} is obtained by tracial \mathrm{PI}(p,p) on each slice (\mathcal{M}_{k},\tau_{k}) . The result is then descended to \mathcal{M} via conditional expectations compatible with the modular action, giving a GNS- \phi -detailed balance version of \mathrm{PI}(p,q) : \|x-E^{\phi}(x)\|_{L^{p}(\mathcal{M})}\leq C\bigl\|\Gamma_{p}(x,x)^{1/2}\bigr\|_{L^{q}(\mathcal{M})}, (10) where E^{\phi} denotes the \phi -preserving conditional expectation onto the fixed-point algebra, \Gamma_{p} the L^{p} version of the gradient form, and \|\cdot\|_{L^{p}(\mathcal{M})} the norm in Haagerup’s L^{p} space. A key advantage of Haagerup’s L^{p} construction is its functoriality: every normal * -homomorphism between von Neumann algebras induces canonically bounded maps on the associated L^{p} -spaces, compatible with composition, duality, and interpolation. Our main result is stated in this framework as follows: Theorem 1.4 (cf Theorem 4.17 and 4.18 ) . Let (\mathcal{M},\phi) be a \sigma -finite von Neumann algebra with faithful normal state \phi , and let (T_{t})_{t\geq 0} be a GNS- \phi -detailed balanced QMS with fixed-point algebra \mathcal{N} . Suppose (T_{t}) has an L^{2}(\mathcal{M},\phi) spectral gap \alpha>0 . Let E_{p}^{\phi}:L^{p}(\mathcal{M})\to L^{p}(\mathcal{N}) be the induced conditional expectation. Then for every p=2 and every p\geq 3 , and for all self-adjoint a\in L^{p}(\mathcal{M})_{\mathrm{sa}} in the domain, we have \|a-E_{p}^{\phi}(a)\|_{L^{p}(\mathcal{M})}\leq\frac{p}{\sqrt{2\alpha}}\|(\Gamma_{p}(a,a))^{1/2}\|_{L^{{p}}(\mathcal{M})}. Furthermore, for general a\in L^{p}(\mathcal{M}) in the domain, \|a-E_{p}^{\phi}(a)\|_{L^{p}(\mathcal{M})}\leq\frac{p}{\sqrt{2\alpha}}\bigl(\|\Gamma_{p}(a,a)^{1/2}\|_{L^{p}(\mathcal{M})}+\|\Gamma_{p}(a^{*},a^{*})^{1/2}\|_{L^{p}(\mathcal{M})}\bigr). In particular, for one-sided dense elements a=xh_{\phi}^{1/p}\in L^{p}(\mathcal{M}) with x\in\mathcal{M} and h_{\phi} the density of \phi in the crossed-product realization (iv) Proposition 4.2 , one has \|(x-E^{\phi}(x))h_{\phi}^{\frac{1}{p}}\|_{L^{p}(\mathcal{M})}\leq\frac{p}{\sqrt{2\alpha}}(\|h_{\phi}^{\frac{1}{p}}\Gamma(x,x)h_{\phi}^{\frac{1}{p}}\|_{L^{p/2}(\mathcal{M})}^{1/2}+\|h_{\phi}^{\frac{1}{p}}\Gamma(\sigma_{\frac{-i}{p}}(x^{*}),\sigma_{\frac{-i}{p}}(x^{*}))h_{\phi}^{\frac{1}{p}}\|_{L^{p/2}(\mathcal{M})}^{1/2}), where \sigma_{t}(x)=h_{\phi}^{it}xh_{\phi}^{-it} denotes the modular group passing to the crossed product representation. Sub-exponential concentration inequalities. Poincaré-type inequalities are closely connected to the concentration of measure. In the random matrix setting, such a principle was established by Huang and Tropp ( huang2021poincare , Thm 2.7) for Markov processes with a spectral gap, showing that \mathrm{PI}(p,p) implies sub-exponential concentration. Their approach relies on matrix inequalities and dilation techniques, and improves upon earlier results of Aoun, Banna, and Youssef Aoun_Banna_Youssef_2020 . In particular, Huang and Tropp introduce a symmetrization (or doubling-of-variables) argument that avoids technical difficulties present in the earlier approach. We show that the \mathrm{PI}(p,p) inequality proved in this work implies a sub-exponential concentration inequality for observables measured in the Lipschitz semi-norm induced by the gradient form as below. Corollary (cf 5.5 ) . Let T_{t}=e^{-tL} be a GNS– \phi -detailed balance QMS with spectral gap \alpha>0 and conditional expectation E onto its fixed-point algebra. Then for any x\in\mathcal{M} and large t>0 , {\mathbb{P}}_{\phi}(|x-E(x)|>t)\leq 2\exp\Bigl(-\frac{\sqrt{\alpha}t}{2\sqrt{2}e\|x\|_{\mathrm{Lip}_{\Gamma}}}\Bigr). Our result highlights a clear dichotomy: under the sole assumption of an L^{2} spectral gap, one cannot expect Gaussian or Talagrand-type concentration, but sub-exponential tails are still universally guaranteed. We further clarify the sharpness of this regime by contrasting it with the stronger concentration phenomena arising from the modified logarithmic Sobolev inequalities (MLSI), and by presenting a birth-death process that rules out any improvement of Gaussian decay under a uniform spectral gap assumption alone. Related works. To our knowledge, this is the first result establishing a noncommutative (p,p) -type Poincaré inequality under only an L^{2} spectral gap assumption. Earlier works obtain Poincaré inequalities with different index pairs and under stronger functional or curvature assumptions. In the noncommutative setting, such curvature bounds are often difficult to verify beyond specific model classes Wirth_Zhang_2023 ; münch2024intertwiningcurvatureboundsgraphs . In Gao_Junge_LaRacuente_Li_2025 , assuming a modified logarithmic Sobolev inequality (MLSI) with constant \alpha_{MLSI}>0 , the authors prove \mathrm{PI}(p,\infty) with optimal order O(\sqrt{p}) . The relation between our \mathrm{PI}(p,p) (or the weakened corollary of \mathrm{PI}(p,\infty) 5.1 with constant O(p) ) and the MLSI-based \mathrm{PI}(p,\infty) regime is not clear, and neither inequality evidently forces the other. MLSI further implies the geometric Talagrand inequality from ( gao_fisher_2020 , Remark 6.9) , while Talagrand-type bounds do not hold for all QMS with only an L^{2} spectral gap, see Example 5.7 . Under the noncommutative \Gamma_{2} -condition \tau\bigl(\Gamma(T_{t}x,T_{t}x)y\bigr)\leq Ce^{-2\alpha t}\tau\bigl(T_{t}\Gamma(x,x)y\bigr)\qquad(x\in\operatorname{Dom}(A^{1/2}),y\in\mathcal{N}_{+}), junge2015noncommutative obtains both (p,p) - and (p,\infty) -type inequalities with constants of order O(p) and O(\sqrt{p}) , respectively. Since both MLSI and \Gamma_{2} imply an L^{2} spectral gap, while the converse fails in general, our hypothesis is strictly weaker. As for applications, DE_palma_2023 uses a (2,\infty) -type Poincaré inequality to derive lower bounds on quantum circuit depth. Using our framework, we say (2,\infty) -type Poincaré inequality holds for a GNS- \phi detailed balanced QMS if \displaystyle\|x-E^{\phi}(x)\|_{L^{2}(\mathcal{M},\phi)_{L}}\leq C\|\Gamma(x,x)^{1/2}\|_{\infty},\quad\forall x=x^{*}, where \|x\|_{L^{2}(\mathcal{M},\phi)_{L}}=\tau(D_{\phi}x^{*}x) . We show (2,\infty) -type Poincaré inequality in 5.1 by choosing p=2,a=xD_{\phi}^{1/2} , where D_{\phi} is the density of \phi . The relationship between the (2,\infty) -type Poincaré inequality used in DE_palma_2023 and our formulation ( 9 ) (up to a dimension factor) is made explicit in ( araiza2023 , (3.9), (4.16)) . Further connections to circuit complexity appear in Bu_2024 ; Li_2025 ; Ding_2025 ; araiza2025 . When detailed balance (or symmetry) fails, \mathrm{PI}(p,q) generally breaks down; instead, space–time Poincaré inequalities under weaker symmetry assumptions are established in li2025 . Organization of this paper. In the remainder of this paper, we present the proof of the main results. Section 2 sets notation and collects technical preliminaries, including von Neumann algebras, quantum Markov semigroups, Markov dilations, and amalgamated free products. Section 3 establishes the appropriate swap operator and derivation in AFP, a generalized Klein inequality, the key input for the tracial Poincaré inequalities, and proves Theorem 1.3 . Section 4 introduces Haagerup’s and Kosaki’s L^{p} spaces, GNS-detailed balance, develops the Haagerup reduction technique, and proves Theorem 1.4 . Finally, Section 5 presents non-self-adjoint \mathrm{PI}(p,p) examples arising from degree-one Rademacher variables, together with applications to sub-exponential concentration inequalities and semigroup diameter in finite dimensions. Notations. We use script letters \mathcal{M},\mathcal{N} for von Neumann algebras, \tau for a trace and \phi for a non-tracial state. The identity operator is denoted by 1 and the identity map by \operatorname{Id} . For x\in\mathcal{M} , x^{*} denotes the adjoint and \Phi_{*} the pre-adjoint of a normal map \Phi . Let (T_{t})_{t\geq 0}=e^{-tL} be a QMS with generator L , and denote by \Gamma and \mathscr{E} the associated gradient and Dirichlet forms. 2 Preliminaries 2.1 Von Neumann algebras and conditional expectations Let (\mathcal{M},\phi) denote a von Neumann algebra \mathcal{M} equipped with a normal, faithful state \phi . Write (\pi_{\phi},H_{\phi},\xi_{\phi}) for the GNS (Gelfand–Naimark–Segal) representation of \mathcal{M} with respect to \phi , where \pi_{\phi}:\mathcal{M}\to B(H_{\phi}) is a *-isomorphism and \xi_{\phi}\in H_{\phi} is cyclic and separating for \pi_{\phi}(\mathcal{M}) , satisfying \phi(x)=\langle\xi_{\phi},\pi_{\phi}(x)\xi_{\phi}\rangle,\qquad x\in\mathcal{M}. Here \xi_{\phi} is cyclic for \mathcal{M} if \overline{\mathcal{M}\xi_{\phi}}=H_{\phi} , and separating if x\xi_{\phi}=0 implies x=0 , or equivalently \overline{\mathcal{M}^{\prime}\xi_{\phi}}=H_{\phi} . After identifying \mathcal{M} with \pi_{\phi}(\mathcal{M}) , we may regard \mathcal{M} as a von Neumann algebra on H_{\phi} with cyclic and separating vector \xi_{\phi} . Define the densely defined antilinear map S:H_{\phi}\to H_{\phi},\qquad S(x\xi_{\phi})=x^{*}\xi_{\phi},\qquad x\in\mathcal{M}. By Tomita–Takesaki theory takesaki_theory_2003-2 , the closure \overline{S} admits the polar decomposition \overline{S}=J\Delta^{1/2} , where J is a conjugate-linear unitary and \Delta is a positive self-adjoint operator. We call J the modular conjugation and \Delta:=\overline{S}^{*}\overline{S} the modular operator associated with \phi . The modular automorphism group (\sigma_{t}^{\phi})_{t\in\mathbb{R}} is then given by \sigma_{t}^{\phi}(x)=\Delta^{it}x\Delta^{-it},\qquad x\in\mathcal{M},t\in\mathbb{R}. (11) Let \mathcal{N}\subset\mathcal{M} be a von Neumann subalgebra. If \mathcal{N} is globally invariant under the modular automorphism group, \sigma_{t}^{\phi}(\mathcal{N})=\mathcal{N}\quad(t\in\mathbb{R}), then by Takesaki’s theorem Takesaki_1972 ; takesaki_theory_2003-2 there exists a unique \phi –preserving conditional expectation E_{\phi}:\mathcal{M}\to\mathcal{N} such that 1. E_{\phi}(y)=y for all y\in\mathcal{N} ; 2. E_{\phi}(y_{1}xy_{2})=y_{1}E_{\phi}(x)y_{2} for all x\in\mathcal{M} , y_{1},y_{2}\in\mathcal{N} ; 3. \phi(E_{\phi}(x))=\phi(x) for all x\in\mathcal{M} . It follows that E_{\phi} is normal, completely positive, contractive, and \mathcal{N} –bimodular. For more background on von Neumann algebras, see Takesaki_1979 ; takesaki_theory_2003-2 . In the remainder of this section, we discuss L^{p} spaces in the tracial case, postponing the non-tracial case to Section 4.1 . Let (\mathcal{M},\tau) be a von Neumann algebra equipped with a faithful normal tracial state \tau . In this setting, one can define noncommutative L^{p} spaces in complete analogy with the classical case. For 1\leq p<\infty , the L^{p} norm is \|x\|_{p}:=\tau(|x|^{p})^{1/p},\qquad|x|=(x^{*}x)^{1/2},\quad x\in\mathcal{M}, and \|x\|_{\infty}:=\|x\| is the operator norm. The completion of \mathcal{M} under this norm gives the noncommutative L^{p} space L^{p}(\mathcal{M},\tau) . We write \mathcal{M}_{\mathrm{sa}} for the self-adjoint part of \mathcal{M} and \mathcal{M}_{+} for the positive cone. The case p=2 is of particular importance: L^{2}(\mathcal{M},\tau) carries the inner product \langle x,y\rangle_{L^{2}(\mathcal{M})}:=\tau(x^{*}y),\qquad x,y\in\mathcal{M}, under which \mathcal{M} embeds densely and isometrically. This is exactly the GNS Hilbert space H_{\phi} associated with \tau . 2.2 Functional inequalities In this section, we introduce the Poincaré inequality for trace-symmetric quantum Markov semigroups. Let (T_{t})_{t\geq 0}:\mathcal{M}\to\mathcal{M} be a quantum Markov semigroup (QMS) on \mathcal{M} , i.e., a family of normal, unital, completely positive maps (UCP maps) satisfying 1. T_{0}=\mathrm{Id}_{\mathcal{M}} and T_{t+s}=T_{t}\circ T_{s} for all s,t\geq 0 , 2. for each x\in\mathcal{M} , the map t\mapsto T_{t}(x) is weak ∗ –continuous. We write formally T_{t}=e^{-tL}, where the semigroup generator (Lindbladian) L is given by L(x):=\mathrm{w}^{*}\text{-}\lim_{t\to 0}\frac{1}{t}\bigl(x-T_{t}(x)\bigr), (12) on the domain of those x\in\mathcal{M} for which the limit exists. In addition, we have a finite von Neumann algebra \mathcal{M} equipped with a normal faithful tracial state \tau and we write in shorthand as (\mathcal{M},\tau) . We define a QMS \tau -symmetric case if it is self-adjoint with respect to the trace, i.e., \displaystyle\tau(x^{*}T_{t}(y))=\tau(T_{t}(x)^{*}y),\quad\forall x,y\in\mathcal{M},t\geq 0. (13) This definition implies that L is positive and \tau -symmetric on L^{2}(\mathcal{M},\tau) . It’s fixed-point algebra \mathcal{N}:=\ker(L)=\{x\in\mathcal{M}:T_{t}(x)=x\text{ for all }t\geq 0\} is a von Neumann subalgebra of \mathcal{M} , and each T_{t} is an \mathcal{N} –bimodule map, T_{t}(axb)=aT_{t}(x)b,\quad\forall a,b\in\mathcal{N},\ x\in\mathcal{M}. In particular, T_{t}\circ E=E\circ T_{t}=E, where E:\mathcal{M}\to\mathcal{N}. is the \tau –preserving conditional expectation onto \mathcal{N} . Before introducing the Poincaré inequality, we recall the associated energy and gradient forms. Definition 2.1 (Dirichlet (energy) form) . The Dirichlet form (or energy functional) of the semigroup is the quadratic map \mathscr{E}:\mathcal{A}\to\mathbb{R}_{\geq 0},\quad\mathscr{E}(x):=\tau(x^{*}L(x)),\quad\forall x\in\mathcal{A}. Definition 2.2 (Gradient (Carré du champ) form) . The non-commutative carré du champ (or gradient) operator associated to the Lindbladian L is the bilinear map \Gamma:{\mathcal{A}}\times{\mathcal{A}}\to\mathcal{M} where \mathcal{A}:=\operatorname{Dom}(L)\cap\mathcal{M}, 1 1 1 {\mathcal{A}} is a dense * –subalgebra of \mathcal{M} and a core for L^{1/2} cipriani_derivations_2003 ; davies_non-commutative_1992 . is defined by ( gao_fisher_2020 , Section 2.1) \Gamma(x,y)=\tfrac{1}{2}\bigl(L(x^{*})y+x^{*}L(y)-L(x^{*}y)\bigr),\quad\forall x,y\in\mathcal{A}. Equivalently by the weak ∗ –limit \Gamma(x,y):=\mathrm{w}^{*}\text{-}\lim_{t\downarrow 0}\frac{1}{2t}\Bigl(T_{t}(x^{*}y)-T_{t}(x^{*})y-x^{*}T_{t}(y)+x^{*}y\Bigr),\quad\forall x,y\in\mathcal{A}, (14) or equivalently from ( cipriani_derivations_2003 , Lemma 9.1) , \Gamma(x,y):=\mathrm{w}^{*}\text{-}\lim_{\epsilon\downarrow 0}\frac{1}{2}\Bigl(x^{*}\frac{L}{1+\epsilon L}(y)+\frac{L}{1+\epsilon L}(x)^{*}y-\frac{L}{1+\epsilon L}(x^{*}y)\Bigr),\quad\forall x,y\in\mathcal{A}. (15) In particular, \Gamma(x,x)\in\mathcal{M}_{+} for every x\in\mathcal{A} . We can also re-write the energy form as \mathscr{E}(x)=\tau(\Gamma(x,x)),\qquad x\in{\mathcal{A}}. We now introduce the formal definitions of functional inequalities for QMSs. Definition 2.3 . Let \{T_{t}\}_{t\geq 0} a \tau –symmetric QMS with Lindbladian L , gradient form \Gamma , and E:\mathcal{M}\to\mathcal{N} be the \tau –preserving conditional expectation onto \mathcal{N} . Let p,r\in[0,\infty] , we say that • T_{t} admits the Poincaré inequality \mathrm{PI}(p,r) with constant c>0 , if \|x-E(x)\|_{L^{p}(\mathcal{M},\tau)}\leq c\|\Gamma(x,x)^{1/2}\|_{L^{r}(\mathcal{M},\tau)},\quad\forall x\in{\mathcal{A}} (16) • L admits the spectral gap \alpha>0 if \alpha\|x-E(x)\|_{L^{2}(\mathcal{M},\tau)}^{2}\leq\|\Gamma(x,x)^{1/2}\|^{2}_{L^{2}(\mathcal{M},\tau)}.\quad\forall x\in\mathcal{A}. (17) Note that the spectral gap is a special case of Poincaré inequalities \mathrm{PI}(2,2) , and the first main result we later derive is that given a spectral gap, one can derive a \mathrm{PI}(p,p) for every p=2 and p\geq 3 . Integrating Equation 17 gives the L^{2} decay estimate \|T_{t}(x)-E(x)\|_{L^{2}(\mathcal{M},\tau)}\leq e^{-\alpha t}\|x-E(x)\|_{L^{2}(\mathcal{M},\tau)},\qquad x\in{\mathcal{A}},\ t\geq 0. (18) 2.3 Construction of amalgamated free product (AFP) 2.3.1 Motivation: Fock spaces Starting from a one–particle Hilbert space \mathcal{H} with vacuum vector \xi , one defines three standard Fock space constructions. The symmetric Fock space arises as the GNS representation of the CCR (canonical commutation relations) algebra, the antisymmetric Fock space corresponds to the CAR (canonical anti–commutation relations) algebra, and the full (free) Fock space serves as the analogue without commutation restrictions. • Bosons (CCR). \mathcal{F}_{\mathrm{Bose}}(\mathcal{H})={\mathbb{C}}\xi\oplus\bigoplus_{n=1}^{\infty}\mathrm{Sym}^{n}(\mathcal{H}). • Fermions (CAR). \mathcal{F}_{\mathrm{Fermi}}(\mathcal{H})={\mathbb{C}}\xi\oplus\bigoplus_{n=1}^{\infty}\bigwedge^{n}\mathcal{H}. • Free particles (free product algebra). {\mathcal{F}}_{\mathrm{Free}}(\mathcal{H})={\mathbb{C}}\xi\oplus\bigoplus_{n=1}^{\infty}\mathcal{H}^{\otimes n}. (19) In the presence of multiple vacua encoded by a subalgebra, one passes to the amalgamated free product (AFP) Fock space, which may be viewed heuristically as obtained by gluing free Fock spaces over the different vacuum sectors. 2.3.2 Amalgamated free products (AFP) We recall the construction of amalgamated free products (AFP) of tracial von Neumann algebras, following speicher_combinatorial_1998 ; araki_operator_1985 ; ioana_amalgamated_2008 ; voiculescu_free_1992 ; brown_-algebras_2008 . Let (\mathcal{M}_{1},\tau_{1}) and (\mathcal{M}_{2},\tau_{2}) be tracial von Neumann algebras with a common subalgebra \mathcal{N}\subset\mathcal{M}_{i} , i=1,2 , such that \tau_{1}|_{\mathcal{N}}=\tau_{2}|_{\mathcal{N}} . There exist unique trace-preserving conditional expectations E_{i}\colon\mathcal{M}_{i}\to\mathcal{N} ( takesaki_theory_2003-2 , Theorem 4.2) , and we write \mathring{\mathcal{M}}_{i}:=\ker(E_{i}). Each \mathcal{M}_{i} is naturally an \mathcal{N} – \mathcal{N} bimodule with \mathcal{N} -valued inner product \langle x,y\rangle_{\mathcal{N}}=E_{i}(x^{*}y) for x,y\in\mathcal{M}_{i} , and hence a Hilbert \mathcal{N} – \mathcal{N} bimodule in the sense of lance1995hilbert . The orthogonal direct sum \bigoplus_{i}\mathcal{M}_{i} is again an \mathcal{N} – \mathcal{N} bimodule with inner product \langle\oplus_{i}x_{i},\oplus_{i}y_{i}\rangle_{\mathcal{N}}=\sum_{i}\langle x_{i},y_{i}\rangle_{\mathcal{N}} . On simple tensors, the interior tensor product \mathring{\mathcal{M}}_{i_{1}}\otimes_{\mathcal{N}}\cdots\otimes_{\mathcal{N}}\mathring{\mathcal{M}}_{i_{n}} carries the inner product \langle x_{1}\otimes\cdots\otimes x_{n},y_{1}\otimes\cdots\otimes y_{n}\rangle_{\mathcal{N}}=E_{i_{n}}\Big(x_{n}^{*}E_{i_{n-1}}\big(\cdots E_{i_{1}}(x_{1}^{*}y_{1})\cdots\big)y_{n}\Big). (20) The AFP \mathcal{M}_{1}*_{\mathcal{N}}\mathcal{M}_{2} can be realized as the completion of the * -algebra of reduced words. \mathcal{N}\oplus\bigoplus_{n=1}^{\infty}\bigoplus_{\begin{subarray}{c}i_{1},\dots,i_{n}\in\{1,2\}\\ i_{k}\neq i_{k+1}\end{subarray}}\mathring{\mathcal{M}}_{i_{1}}\otimes_{\mathcal{N}}\cdots\otimes_{\mathcal{N}}\mathring{\mathcal{M}}_{i_{n}}. (21) This decomposition inherits the structure of a Hilbert \mathcal{N} – \mathcal{N} bimodule by discussion around eq. 20 . There are natural embeddings \pi_{i}:\mathcal{M}_{i}\hookrightarrow\mathcal{M}_{1}*_{\mathcal{N}}\mathcal{M}_{2},\qquad i=1,2. (22) The canonical trace on the AFP, denoted by \widetilde{\tau}:=\tau_{1}*\tau_{2} , is defined on reduced words. By ( takesaki_theory_2003-2 , Theorem 4.2) , there exists a unique \widetilde{\tau} -preserving conditional expectation \widetilde{E}:\mathcal{M}_{1}*_{\mathcal{N}}\mathcal{M}_{2}\to\mathcal{N} such that: 1. \pi_{1}|_{\mathcal{N}}=\pi_{2}|_{\mathcal{N}}=\operatorname{Id}_{\mathcal{N}} ; 2. \widetilde{\tau}\circ\widetilde{E}=\widetilde{\tau} ; 3. \widetilde{E}\circ\pi_{i}=E_{i} , i=1,2 ; 4. \widetilde{\tau}|_{\mathcal{N}}=\tau_{1}|_{\mathcal{N}}=\tau_{2}|_{\mathcal{N}} ; 5. ( Freeness with amalgamation voiculescu_free_1992 ) if i_{1},\dots,i_{n}\in\{1,2\} with i_{k}\neq i_{k+1} and m_{k}\in\mathring{\mathcal{M}}_{i_{k}} , then \widetilde{\tau}(\pi_{i_{1}}(m_{1})\cdots\pi_{i_{n}}(m_{n}))=0,\qquad\widetilde{E}(\pi_{i_{1}}(m_{1})\cdots\pi_{i_{n}}(m_{n}))=0; 6. \mathcal{M}_{1}*_{\mathcal{N}}\mathcal{M}_{2} is generated by \pi_{1}(\mathcal{M}_{1})\cup\pi_{2}(\mathcal{M}_{2}) . The above properties give the abstract characterization of AFP. The AFP Fock space representation We also construct the AFP Fock space with non-scalar \mathcal{N} . In the scalar case where \mathcal{N}={\mathbb{C}}1 , this reduces to the free Fock space representation Equation 19 . Following the notations from above, the GNS Hilbert space of the AFP (\mathcal{M}_{1}*_{\mathcal{N}}\mathcal{M}_{2},\widetilde{\tau}) is The GNS Hilbert space of the AFP (\mathcal{M}_{1}*_{\mathcal{N}}\mathcal{M}_{2},\widetilde{\tau}) is \mathcal{H}_{12}:=\overline{\mathcal{M}_{1}*_{\mathcal{N}}\mathcal{M}_{2}}^{\|\cdot\|_{2}},\qquad\|x\|_{2}=\widetilde{\tau}(x^{*}x)^{1/2}. with scalar inner product induced from the \mathcal{N} -valued inner product: \langle x,y\rangle:=\widetilde{\tau}(\widetilde{E}(x^{*}y))=\widetilde{\tau}(\langle x,y\rangle_{\mathcal{N}})\in{\mathbb{C}}. (23) One can then write \mathcal{H}_{12} 2 2 2 This construction is equivalent to introducing Connes fusion takesaki_theory_2003-2 , but we avoid introducing this machinery here for simplicity. as brown_-algebras_2008 : \mathcal{H}_{12}=L^{2}(\mathcal{N})\oplus\bigoplus_{n=1}^{\infty}\bigoplus_{\begin{subarray}{c}i_{1},\dots,i_{n}\in\{1,2\}\\ i_{k}\neq i_{k+1}\end{subarray}}L^{2}(\mathring{\mathcal{M}}_{i_{1}}\otimes_{\mathcal{N}}\cdots\otimes_{\mathcal{N}}\mathring{\mathcal{M}}_{i_{n}}). (24) where \oplus denotes the orthogonal direct sum with respect to the inner product in ( 23 ). Each inclusion \pi_{i}:\mathcal{M}_{i}\hookrightarrow\mathcal{M}_{1}*_{\mathcal{N}}\mathcal{M}_{2}\subset B(\mathcal{H}_{12}) , i=1,2 , defines a left regular action on simple tensors, which form a dense set, given by ( brown_-algebras_2008 , Lemma 4.7.3) \pi_{i}(x)(n\xi_{\mathcal{N}})=E_{i}(x)n\xi_{\mathcal{N}},\qquad n\in\mathcal{N}, where \xi_{\mathcal{N}} denotes the GNS vector for L^{2}(\mathcal{N})=\overline{\mathcal{N}\xi_{\mathcal{N}}} . For \zeta_{1}\otimes\cdots\otimes\zeta_{n}\in\mathring{\mathcal{M}}_{i_{1}}\otimes_{\mathcal{N}}\cdots\otimes_{\mathcal{N}}\mathring{\mathcal{M}}_{i_{n}},i_{1},\dots,i_{n}\in\{1,2\} , \pi_{i}(x)(\zeta_{1}\otimes\cdots\otimes\zeta_{n})=\begin{cases}\bigl(x\zeta_{1}-E_{i}(x\zeta_{1})\bigr)\otimes\zeta_{2}\otimes\cdots\otimes\zeta_{n}+E_{i}(x\zeta_{1})\zeta_{2}\otimes\cdots\otimes\zeta_{n},&i_{1}=i,\\ (x-E_{i}(x))\otimes\zeta_{1}\otimes\cdots\otimes\zeta_{n}+E_{i}(x)\zeta_{1}\otimes\cdots\otimes\zeta_{n},&i_{1}\neq i,\end{cases} (25) In words, if the leading index equals i , \pi_{i}(x) acts by left multiplication on the first tensor leg, with the usual centering correction on \mathring{\mathcal{M}}_{i} ; otherwise, it creates a new tensor leg on the left up to centering correction 3 3 3 Under the identification obtained by regrouping the decomposition ( 21 ) of \mathcal{M}_{1}*_{\mathcal{N}}\mathcal{M}_{2} as \mathcal{M}_{i}\otimes_{\mathcal{N}}\cdots , the representations coincide with the left regular actions: \pi_{i}(x)=x\otimes\mathbf{1} . . Each \pi_{i} is a faithful normal * -homomorphism with \pi_{i}|_{\mathcal{N}}=\operatorname{Id}_{\mathcal{N}} . Thus eq. 22 extends to the GNS representation after completion. 3 Poincaré inequality for trace-symmetric QMS In this section we prove Theorem 1.3 . The argument proceeds as follows: We introduce the swap-operator construction as a noncommutative symmetrization, state auxiliary lemmas for the associated derivation, establish a generalized Klein inequality giving a convex chain rule, and finally prove the \mathrm{PI}(p,p) inequality and its non-self-adjoint extension. To ease notation, in this section, we write p-norms \|\cdot\|_{L^{p}(\mathcal{M},\tau)} as \|\cdot\|_{p} . 3.1 Swap operator construction Specifically, we are working in the case where (\mathcal{M}_{1},\tau_{1}) and (\mathcal{M}_{2},\tau_{2}) be two copies of (\mathcal{M},\tau) containing the von Neumann subalgebra \mathcal{N} . We use the shorthand \mathcal{H}_{ij}:=L^{2}(\mathcal{M}_{i}*_{\mathcal{N}}\mathcal{M}_{j},\widetilde{\tau}) and define a swap operator U\colon\mathcal{H}_{12}\to\mathcal{H}_{21} that exchanges the roles of the two legs 1\leftrightarrow 2 in the Hilbert space generated by the AFP \mathcal{M}_{1}*_{\mathcal{N}}\mathcal{M}_{2} . As a prelude, we first display the commuting diagram that organizes the construction. The swap unitary U implementing \mathrm{Ad}_{U} is constructed later in this subsection. {\lx@inpgf@ignorespaces\mathcal{M}_{1}} {\lx@inpgf@ignorespaces\mathcal{M}_{1}\ast_{\mathcal{N}}\mathcal{M}_{2}} {\lx@inpgf@ignorespaces\mathcal{N}} {\lx@inpgf@ignorespaces\mathcal{M}_{2}} \scriptstyle{\lx@inpgf@ignorespaces E_{1}} \scriptstyle{\lx@inpgf@ignorespaces\pi_{1}} \scriptstyle{\lx@inpgf@ignorespaces{\mathcal{E}}_{1}} \scriptstyle{\lx@inpgf@ignorespaces\exists!\widetilde{E}} \scriptstyle{\lx@inpgf@ignorespaces\mathrm{Ad}_{U}} \scriptstyle{\lx@inpgf@ignorespaces\mathrm{Ad}_{U}|_{\mathcal{N}}=\operatorname{Id}|_{\mathcal{N}}} \scriptstyle{\lx@inpgf@ignorespaces E_{2}} \scriptstyle{\lx@inpgf@ignorespaces\pi_{2}} \scriptstyle{\lx@inpgf@ignorespaces{\mathcal{E}}_{2}} Figure 1: Roadmap for the AFP construction and the swap automorphism. Here, each arrow denotes 1. \pi_{i}\colon\mathcal{M}_{i}\hookrightarrow\mathcal{M}_{1}\ast_{\mathcal{N}}\mathcal{M}_{2} is the canonical inclusion. 2. {\mathcal{E}}_{i}\colon\mathcal{M}_{1}\ast_{\mathcal{N}}\mathcal{M}_{2}\to\mathcal{M}_{i} is the canonical conditional expectation onto \mathcal{M}_{i} , characterized by {\mathcal{E}}_{i}\circ\pi_{i}=\operatorname{Id}_{\mathcal{M}_{i}},\qquad{\mathcal{E}}_{i}(\pi_{j}(x))=0\quad(j\neq i). 3. E_{i}\colon\mathcal{M}_{i}\to\mathcal{N} is the given \tau_{i} -preserving conditional expectation. 4. The two triangles commute, i.e. E_{i}\circ{\mathcal{E}}_{i}=\widetilde{E} . 5. \widetilde{E}\colon\mathcal{M}_{1}\ast_{\mathcal{N}}\mathcal{M}_{2}\to\mathcal{N} is the unique conditional expectation whose restriction to \pi_{i}(\mathcal{M}) equals E_{i} . 6. \mathrm{Ad}_{U}\colon\mathcal{M}_{1}\ast_{\mathcal{N}}\mathcal{M}_{2}\to\mathcal{M}_{1}\ast_{\mathcal{N}}\mathcal{M}_{2} is the spatial ∗ -automorphism implemented by the swap unitary U , satisfying \mathrm{Ad}_{U}\circ\pi_{1}=\pi_{2},\qquad\mathrm{Ad}_{U}\circ\pi_{2}=\pi_{1},\qquad\mathrm{Ad}_{U}|_{\mathcal{N}}=\operatorname{Id}|_{\mathcal{N}}. Expanding the canonical orthogonal decomposition eq. 24 , \displaystyle\mathcal{H}_{12} \displaystyle=L^{2}(\mathcal{N})\oplus\bigl(L^{2}(\mathring{\mathcal{M}}_{1})\oplus L^{2}(\mathring{\mathcal{M}}_{2})\bigr) (26) \displaystyle\oplus\bigl(L^{2}(\mathring{\mathcal{M}}_{1}\otimes_{\mathcal{N}}\mathring{\mathcal{M}}_{2})\oplus L^{2}(\mathring{\mathcal{M}}_{2}\otimes_{\mathcal{N}}\mathring{\mathcal{M}}_{1})\bigr) \displaystyle\oplus\bigl(L^{2}(\mathring{\mathcal{M}}_{1}\otimes_{\mathcal{N}}\mathring{\mathcal{M}}_{2}\otimes_{\mathcal{N}}\mathring{\mathcal{M}}_{1})\oplus\cdots\bigr). Similarly, the decomposition of \mathcal{H}_{21} swaps 1\leftrightarrow 2 . Both decompositions are canonically isometric. In the symmetric case \mathcal{M}_{1}=\mathcal{M}_{2}=\mathcal{M} , the swap operator is defined by \displaystyle U\colon L^{2}(\mathring{\mathcal{M}}_{i_{1}}\otimes_{\mathcal{N}}\cdots\otimes_{\mathcal{N}}\mathring{\mathcal{M}}_{i_{n}}) \displaystyle\longrightarrow L^{2}(\mathring{\mathcal{M}}_{3-i_{1}}\otimes_{\mathcal{N}}\cdots\otimes_{\mathcal{N}}\mathring{\mathcal{M}}_{3-i_{n}}), (27) \displaystyle\eta_{1}\otimes\cdots\otimes\eta_{n} \displaystyle\longmapsto\zeta_{1}\otimes\cdots\otimes\zeta_{n}, with \zeta_{k}=\eta_{k} regarded in the corresponding copy of \overset{\circ}{\mathcal{M}}_{3-i_{k}} . On the vacuum sector we set U|_{L^{2}(\mathcal{N})}=\operatorname{Id}|_{L^{2}(\mathcal{N})} . This implies a unitary isomorphism \mathcal{H}_{12}\simeq\mathcal{H}_{21} implementing the label exchange 1\leftrightarrow 2 . Proposition 3.1 . The swap operator U is unitary. Proof. It suffices to verify norm preservation on simple tensors. Let \eta=\eta_{1}\otimes\cdots\otimes\eta_{n}\in\mathring{\mathcal{M}}_{i_{1}}\otimes_{\mathcal{N}}\cdots\otimes_{\mathcal{N}}\mathring{\mathcal{M}}_{i_{n}}, and set \zeta_{k}=\eta_{k}\in\mathring{\mathcal{M}}_{3-i_{k}} . Using the L^{2} inner product eq. 23 and \widetilde{\tau}|_{\mathcal{N}}=\tau|_{\mathcal{N}} , \displaystyle\|U\eta\|_{2}^{2} \displaystyle=\langle U\eta,U\eta\rangle=\widetilde{\tau}(\widetilde{E}((U\eta)^{*}(U\eta))) \displaystyle=\tau(E_{3-i_{n}}(\zeta_{n}^{*}E_{3-i_{n-1}}(\cdots E_{3-i_{1}}(\zeta_{1}^{*}\zeta_{1})\cdots)\zeta_{n})) \displaystyle=\tau(E_{i_{n}}(\eta_{n}^{*}E_{i_{n-1}}(\cdots)\eta_{n}))=\|\eta\|_{2}^{2}, where the last equality follows by relabeling 3-i_{k}\mapsto i_{k} . Thus U extends to an isometry \mathcal{H}_{12}\to\mathcal{H}_{21} . The inverse construction gives an isometry \mathcal{H}_{21}\to\mathcal{H}_{12} , hence U is unitary. ∎ Proposition 3.2 . Let U be the swap unitary on \mathcal{H}_{12} . Then \mathrm{Ad}_{U}(T)=UTU^{*} restricts to a * -automorphism of \mathcal{M}_{1}*_{\mathcal{N}}\mathcal{M}_{2} satisfying: 1. \mathrm{Ad}_{U}(\pi_{i}(x))=\pi_{3-i}(x) for all x\in\mathcal{M} . 2. \mathrm{Ad}_{U} commutes with \widetilde{E} . 3. \mathrm{Ad}_{U} preserves \widetilde{\tau} . Proof. Since U is unitary, \mathrm{Ad}_{U} defines a spatial * -isomorphism between B(\mathcal{H}_{12}) and B(\mathcal{H}_{21}) ( Takesaki_1979 , Section 3) . Identifying \mathcal{H}_{21} with \mathcal{H}_{12} via U , we view \mathrm{Ad}_{U} as an automorphism of B(\mathcal{H}_{12}) . (i) It suffices to check the action on simple tensors. Let \eta=\eta_{1}\otimes\cdots\otimes\eta_{n} with i_{1}=i . Using eq. 25 and U|_{L^{2}(\mathcal{N})}=\operatorname{Id} , \displaystyle U\pi_{i}(x)\eta \displaystyle=(x\zeta_{1}-E_{3-i}(x\zeta_{1}))\otimes\cdots\otimes\zeta_{n}+E_{3-i}(x\zeta_{1})\zeta_{2}\otimes\cdots\otimes\zeta_{n} \displaystyle=\pi_{3-i}(x)(\zeta_{1}\otimes\cdots\otimes\zeta_{n})=\pi_{3-i}(x)U\eta, where \zeta_{k}=\eta_{k} viewed in \mathring{\mathcal{M}}_{3-i_{k}} . Thus \mathrm{Ad}_{U}(\pi_{i}(x))=\pi_{3-i}(x) . Since \mathcal{M}_{1}*_{\mathcal{N}}\mathcal{M}_{2} is generated by \pi_{1}(\mathcal{M}) and \pi_{2}(\mathcal{M}) , \mathrm{Ad}_{U} restricts to a * -automorphism. (ii) Following Figure 1 , \mathrm{Ad}_{U}\circ\widetilde{E}\circ\pi_{i}=\widetilde{E}\circ\pi_{i}=\widetilde{E}\circ\pi_{3-i}=\widetilde{E}\circ\mathrm{Ad}_{U}\circ\pi_{i}, since U|_{L^{2}(\mathcal{N})}=\operatorname{Id}|_{L^{2}(\mathcal{N})} and \mathcal{M}_{1}=\mathcal{M}_{2}=\mathcal{M} . As \mathcal{M}_{1}*_{\mathcal{N}}\mathcal{M}_{2} is generated by \pi_{1}(\mathcal{M}) and \pi_{2}(\mathcal{M}) and both \mathrm{Ad}_{U} and \widetilde{E} are normal, the relation extends to all of \mathcal{M}_{1}*_{\mathcal{N}}\mathcal{M}_{2} . Thus \mathrm{Ad}_{U}\circ\widetilde{E}=\widetilde{E}=\widetilde{E}\circ\mathrm{Ad}_{U} . (iii) Because \widetilde{\tau}=\widetilde{\tau}\circ\widetilde{E} , \widetilde{\tau}(\mathrm{Ad}_{U}(T))=\widetilde{\tau}(\widetilde{E}(\mathrm{Ad}_{U}(T)))=\widetilde{\tau}(\widetilde{E}(T))=\widetilde{\tau}(T),\qquad T\in\mathcal{M}_{1}*_{\mathcal{N}}\mathcal{M}_{2}. Hence \mathrm{Ad}_{U} swaps the two embeddings, commutes with \widetilde{E} , and preserves \widetilde{\tau} . ∎ We will see in Section 3.4 that the swap unitary U is the essential ingredient for the symmetrization technique, replacing the product measure in the commutative case ( huang2021poincare , Section 4.1) . 3.2 Auxiliary lemmas We first show several auxiliary lemmas used in the proof of Theorem 1.3 . They state the relevant estimates in terms of a noncommutative derivation, replacing the product-measure construction of doubling-up variables techniques from huang2021poincare in the semi-commutative case. We define the noncommutative derivation \delta(x):=\pi_{1}(x)-\pi_{2}(x)\in\mathcal{M}_{1}*_{\mathcal{N}}\mathcal{M}_{2}. (28) Lemma 3.3 . Let x\in\mathcal{M} . Then for any 1\leq p<\infty , \|x-E_{1}(x)\|_{p}^{p}\leq\|\delta(x)\|_{p}^{p}. Proof. Since the conditional expectation {\mathcal{E}}_{1}:L_{p}(\mathcal{M}_{1}*_{\mathcal{N}}\mathcal{M}_{2})\to L_{p}(\mathcal{M}_{1}) is contractive takesaki_theory_2003-2 , we have \|{\mathcal{E}}_{1}(\delta(x)-E(\delta(x)))\|_{p}\leq\|\delta(x)-E(\delta(x))\|_{p}. Moreover, {\mathcal{E}}_{1}(\delta(x))={\mathcal{E}}_{1}(\pi_{1}(x))-{\mathcal{E}}_{1}(\pi_{2}(x))=x-E_{1}(x), and since E\circ\pi_{i}=E_{i} with E_{1}=E_{2} , one has E(\delta(x))=0 . The claim follows. ∎ The following lemma gives the noncommutative version of the symmetrization technique: Lemma 3.4 . With \delta(x) as above and U the swap unitary satisfying \mathrm{Ad}_{U}\circ\pi_{1}=\pi_{2} and \mathrm{Ad}_{U}\circ\pi_{2}=\pi_{1} , let h:\mathbb{R}\to\mathbb{R} be an odd function. Then 1. \mathrm{Ad}_{U}(h(\delta(x)))=-h(\mathrm{Ad}_{U}(\delta(x))) , 2. \widetilde{\tau}(h(\delta(x)))=0 , 3. E(h(\delta(x)))=0 , where \widetilde{\tau} is the canonical trace on \mathcal{M}_{1}*_{\mathcal{N}}\mathcal{M}_{2} . Proof. By Proposition 3.2 , \mathrm{Ad}_{U}(\delta(x))=-\delta(x) . Since h is odd, h(-\delta(x))=-h(\delta(x)) . The U -invariance of \widetilde{\tau} and E then implies the stated identities. ∎ Lemma 3.5 . Let \widetilde{L} and \widetilde{\Gamma} be the Lindblad generator and gradient form on \mathcal{M}_{1}*_{\mathcal{N}}\mathcal{M}_{2} induced from \mathcal{M} as in boca_completely_1993 , characterized by \widetilde{L}\bigl(\pi_{i}(x)\bigr)=\pi_{i}\bigl(L_{i}(x)\bigr),\qquad\widetilde{\Gamma}\bigl(\pi_{i}(x),\pi_{i}(x)\bigr)=\pi_{i}\bigl(\Gamma_{i}(x,x)\bigr),\quad i=1,2, where L_{i}=L and \Gamma_{i}=\Gamma are the corresponding operators on \mathcal{M} . Then for every x\in\mathcal{M} and p\geq 2 , \|\widetilde{\Gamma}(\delta(x),\delta(x))\|_{p/2}\leq 2\|\Gamma(x,x)\|_{p/2}. Proof. A direct computation shows \widetilde{\Gamma}(\delta(x),\delta(x))=\pi_{1}(\Gamma(x,x))+\pi_{2}(\Gamma(x,x)). Since \pi_{i} preserves the L^{p/2} -norm and the trace, the inequality follows immediately. ∎ 3.3 Generalized Klein inequalities The key analytic input for the tracial \mathrm{PI}(p,p) inequality is a convexity estimate relating functional calculus and the gradient form. This is provided by Markov dilation ( Appendix A ) and a generalized Klein inequality, which is used to obtain a convex-chain rule for the Dirichlet form. We recall the tracial version below and derive the corresponding gradient estimate. Lemma 3.6 (Generalized Klein inequality for tracial vNa) . Let (\mathcal{M},\tau) be a von Neumann algebra as in the last section. Let x,y\in\mathcal{M}_{\mathrm{sa}} be bounded self-adjoint elements, and let \varphi:\mathbb{R}\to\mathbb{R} be a continuously differentiable function whose squared derivative \psi=(\varphi^{\prime})^{2} is convex on \mathbb{R} . Then \tau\bigl[(\varphi(x)-\varphi(y))^{2}\bigr]\leq\frac{1}{2}\tau\Bigl[(x-y)^{2}\bigl(\psi(x)+\psi(y)\bigr)\Bigr]. (29) Proof. By ( petz_survey_1994 , Prop. 3) , the claim holds in finite dimensions, and the general bounded case follows by spectral approximation and norm continuity. ∎ Proposition 3.7 (Convex-chain estimate for gradient form) . Assume the standing hypotheses of Section 2 and lemma 3.6 , for every self-adjoint element x\in\operatorname{Dom}(L)\subset L_{2}(\mathcal{M},\tau) , \mathscr{E}\bigl(\varphi(x)\bigr)\leq\tau\bigl(\Gamma(x,x)\psi(x)\bigr), where \mathscr{E}(f):=\tau\bigl(\Gamma(f,f)\bigr) is the Dirichlet form associated with L . We defer the proof to Appendix A , where the main technique is Markov dilation that generalizes the Markov processes in huang2021poincare . 3.4 Proof of trace-symmetric \mathrm{PI}(p,p) We now prove Theorem 1.3 . We begin by recalling its statement. Theorem . Let (\mathcal{M},\tau) be a tracial von Neumann algebra and (T_{t})_{t\geq 0}=e^{-tL} a \tau –symmetric quantum Markov semigroup with fixed-point algebra \mathcal{N} and conditional expectation E:\mathcal{M}\to\mathcal{N} . If \{T_{t}\}_{t\geq 0} has a spectral gap \alpha>0 , then for every self-adjoint x in the domain of L and p=2 or p\geq 3 , it satisfies the Poincaré inequality \mathrm{PI}(p,p) with constant \tfrac{p}{\sqrt{2\alpha}} , i.e., \|x-E(x)\|_{p}\leq\frac{p}{\sqrt{2\alpha}}\|\Gamma(x,x)^{1/2}\|_{p}. Proof. For x\in\mathcal{M}_{sa} , Lemma 3.3 gives \|x-E_{1}(x)\|_{p}^{p}\leq\|\delta(x)\|_{p}^{p},\qquad\delta(x):=\pi_{1}(x)-\pi_{2}(x)\in(\mathcal{M}_{1}*_{\mathcal{N}}\mathcal{M}_{2})_{sa}. (30) Define the odd function \varphi(s):=\operatorname{sgn}(s)|s|^{p/2} (cf. huang2021poincare ). Its squared derivative is \psi(s):=(\varphi^{\prime}(s))^{2}=\Big(\frac{p}{2}\Big)^{2}|s|^{p-2}, which is convex for p=2 and p\geq 3 . Since \varphi is odd, Lemma 3.4 implies E(\varphi(\delta(x)))=0 , and hence \|\varphi(\delta(x))-E(\varphi(\delta(x)))\|_{2}^{2}=\widetilde{\tau}(\varphi(\delta(x))^{*}\varphi(\delta(x)))=\|\delta(x)\|_{p}^{p}=\|\delta(x)-E(\delta(x))\|_{p}^{p}. With the spectral-gap assumption ( 17 ) (extended to \mathcal{M}_{1}*_{\mathcal{N}}\mathcal{M}_{2} by B.3 ), \alpha\|\delta(x)\|_{p}^{p}\leq\|\widetilde{\Gamma}(\varphi(\delta(x)),\varphi(\delta(x)))^{1/2}\|_{2}^{2}=\widetilde{\tau}(\varphi(\delta(x))^{*}\widetilde{L}(\varphi(\delta(x))))=:\mathscr{E}(\varphi(\delta(x))). By Proposition 3.7 (convex chain rule), \mathscr{E}(\varphi(\delta(x)))\leq\widetilde{\tau}(\widetilde{\Gamma}(\delta(x),\delta(x))\psi(\delta(x)))=\Big(\frac{p}{2}\Big)^{2}\widetilde{\tau}(\widetilde{\Gamma}(\delta(x),\delta(x))|\delta(x)|^{p-2}). Applying Hölder’s inequality, \alpha\|\delta(x)\|_{p}^{p}\leq\Big(\frac{p}{2}\Big)^{2}\widetilde{\tau}(\widetilde{\Gamma}(\delta(x),\delta(x))^{p/2})^{2/p}\|\delta(x)\|_{p}^{p-2}. Dividing by \|\delta(x)\|_{p}^{p-2} and applying Lemma 3.5 gives \|\delta(x)\|_{p}^{2}\leq\frac{p^{2}}{4\alpha}\widetilde{\tau}(\widetilde{\Gamma}(\delta(x),\delta(x))^{p/2})^{2/p}\leq\frac{p^{2}}{2\alpha}\tau(\Gamma(x,x)^{p/2})^{2/p}. Taking square roots and using ( 30 ) completes the proof: \|x-E(x)\|_{p}\leq\frac{p}{\sqrt{2\alpha}}\|\Gamma(x,x)^{1/2}\|_{p}. ∎ Remark 3.8 (Missing Parameters huang2021poincare ) . Theorem 1.3 also holds for p\in(2,3) with an extra factor of \sqrt{2} on the right hand side through applying a variant of Proposition 3.7 that only need \psi to be monotone. We can also generalize to non-self-adjoint elements in \mathcal{M} by the following corollary. We note that in the non-tracial setting, the non-self-adjoint estimate in Proposition 4.16 may be useful for state-preparation tasks where a non-self-adjoint form of the Poincaré inequality is required. Corollary 3.9 (Non self-adjoint \mathrm{PI}(p,p) ) . In the context of Theorem 1.3 , for every p=2 and p\geq 3 and x\in\mathcal{M} : \|x-E(x)\|_{p}\leq\frac{p}{\sqrt{2\alpha}}\bigl(\|\Gamma(x,x)^{1/2}\|_{p}+\|\Gamma(x^{*},x^{*})^{1/2}\|_{p}\bigr). Proof. Since x=\real x+\mathrm{i}\imaginary x with \real x,\imaginary x\in\mathcal{M}_{sa} , the triangle inequality in L_{p} gives \|x-E(x)\|_{p}\leq\|\real x-E(\real x)\|_{p}+\|\imaginary x-E(\imaginary x)\|_{p}. We apply Theorem 1.3 to \real x and \imaginary x , \|\real x-E(\real x)\|_{p}+\|\imaginary x-E(\imaginary x)\|_{p}\leq\frac{p}{\sqrt{2\alpha}}\bigl(\|\Gamma(\real x,\real x)\|_{p/2}^{1/2}+\|\Gamma(\imaginary x,\imaginary x)\|_{p/2}^{1/2}\bigr). By sesquilinearity of \Gamma , \Gamma(\real x,\real x)+\Gamma(\imaginary x,\imaginary x)=\frac{1}{2}\bigl[\Gamma(x,x)+\Gamma(x^{*},x^{*})\bigr]. Hence the Minkowski, inequality gives \sqrt{a+b}\leq\sqrt{a}+\sqrt{b} imply \|\Gamma(\real x,\real x)\|_{p/2}^{1/2}+\|\Gamma(\imaginary x,\imaginary x)\|_{p/2}^{1/2}\leq\|\Gamma(x,x)\|_{p/2}^{1/2}+\|\Gamma(x^{*},x^{*})\|_{p/2}^{1/2}. Combining the above estimates gives the stated corollary. ∎ 4 Poincaré inequality for GNS-detailed balance QMS Previously, we studied trace-symmetric quantum Markov semigroups and proved the \mathrm{PI}(p,p) inequality ( theorem 1.3 ) under the assumption of a spectral gap on a tracial von Neumann algebra (\mathcal{M},\tau) . In this section, we extend this result to von Neumann algebras that are not equipped with a trace, and to QMSs with a weaker symmetry condition called GNS-detailed balanced with respect to a non-tracial state. The main tools are Haagerup’s reduction method haagerup2009reductionmethodnoncommutativelpspaces and Kosaki’s L^{p} interpolation theorem kosaki_applications_1984 . We also formulate the results using Haagerup’s functorial L^{p} construction. To set notations, we first recall results from Haagerup’s L^{p} spaces and Kosaki’s L^{p} interpolation spaces. 4.1 Haagerup’s and Kosaki’s L^{p} spaces Haagerup’s L^{p} spaces We give a concise introduction to Haagerup’s L^{p} spaces ( haagerup2009reductionmethodnoncommutativelpspaces , Section 1.2) haagerup1979lpspaces for general vNa. Let \mathcal{M} be a von Neumann algebra and \phi a faithful normal semifinite weight on \mathcal{M} . Denote by \sigma_{t}^{\phi}=\sigma_{t}\ (t\in\mathbb{R}) its modular automorphism group, and form the crossed product ( takesaki_theory_2003-2 , Section X) , {\mathcal{R}}=\mathcal{M}\rtimes_{\sigma}\mathbb{R}=\bigl\{\pi_{\sigma}(\mathcal{M})\cup\lambda(\mathbb{R})\bigr\}^{\prime\prime} ,where \pi_{\sigma} and \lambda act on L^{2}(\mathcal{M},\phi)\otimes L^{2}(\mathbb{R},dt) . It is well known that {\mathcal{R}} is semifinite and admits a unique normal semifinite faithful trace \tau (the canonical trace ) characterized by the scaling \tau\circ\widehat{\sigma}_{t}=e^{-t}\tau,\quad t\in\mathbb{R}, (31) where \widehat{\sigma} is the dual action on {\mathcal{R}} . We define and summarize some useful properties of Haagerup’s L^{p} –spaces haagerup1979lpspaces . Definition 4.1 (Haagerup’s L^{p} Spaces) . With the notation above, for 0
0 and the fixed point algebra \mathcal{N} . Let E_{p}^{\phi} be the induced conditional expectation from L^{p}(\mathcal{M}) onto L^{p}(\mathcal{N}) , then for every p=2 and p\geq 3 and a\in L^{p}(\mathcal{M})_{sa} in the domain, the following holds in Haagerup’s L^{p} spaces: \|a-E_{p}^{\phi}(a)\|_{L^{p}(\mathcal{M})}\leq\frac{p}{\sqrt{2\alpha}}\|\Gamma_{p}(a,a)\|_{L^{\frac{p}{2}}(\mathcal{M})}^{1/2}. (48) Equivalently, using Kosaki’s L^{p} spaces notation, for x\in\mathcal{M}_{sa} in the domain \bigl\|\iota_{\eta}(x-E^{\phi}(x))\bigr\|_{p,\phi,\eta}\leq\frac{p}{\sqrt{2\alpha}}\Bigl(\bigl\|\iota_{\frac{1}{2}}\bigl(\Gamma^{(p)}_{\eta}(x,x)\bigr)\bigr\|_{\frac{p}{2},\phi,\frac{1}{2}}^{1/2}\Bigr). (49) Corollary 4.18 (GNS- \phi -detailed balance \mathrm{PI}(p,p) for non-self-adjoint elements) . Under the construction of Theorem 4.17 , for a\in L^{p}(\mathcal{M}) in the domain, \|a-E_{p}^{\phi}(a)\|_{L^{p}(\mathcal{M})}\leq\frac{p}{\sqrt{2\alpha}}\bigl(\|\Gamma_{p}(a,a)^{1/2}\|_{L^{p}(\mathcal{M})}+\|\Gamma_{p}(a^{*},a^{*})^{1/2}\|_{L^{p}(\mathcal{M})}\bigr). (50) Equivalently, using Kosaki’s L^{p} spaces notation, for \eta\in[0,1] , x\in\mathcal{M} in the domain \bigl\|\iota_{\eta}(x-E(x))\bigr\|_{p,\phi,\eta}\leq\frac{p}{\sqrt{2\alpha}}\Bigl(\bigl\|\iota_{\frac{1}{2}}\bigl(\Gamma^{(p)}_{\eta}(x,x)\bigr)\bigr\|_{\frac{p}{2},\phi,\frac{1}{2}}^{1/2}+\bigl\|\iota_{\frac{1}{2}}\bigl(\Gamma^{(p)}_{1-\eta}(x^{*},x^{*})\bigr)\bigr\|_{\frac{p}{2},\phi,\frac{1}{2}}^{1/2}\Bigr). (51) Proof. As in 3.9 , write a=\real(a)+i\imaginary(a) and apply the bound to \real(a) and \imaginary(a) . ∎ 4.5 Haagerup’s reduction To generalize our conjectures to the non-tracial cases, we introduce the machinery called Haagerup’s reduction haagerup2009reductionmethodnoncommutativelpspaces , which shows that all vNa admit finite approximations. We later show that the tracial \mathrm{PI}(p,p) can be transferred to general vNa via Haagerup reduction. Throughout this section, let G=\bigcup_{n\geq 1}2^{-n}\mathbb{Z}\subset\mathbb{R} and let \mathcal{M} be a \sigma -finite 8 8 8 Mutually orthogonal projections in \mathcal{M} are at most countable. von Neumann algebra equipped with a normal faithful state \phi . We denote by \sigma_{t}^{\phi} the modular automorphism group of \phi on \mathcal{M} , and form the crossed-product \widehat{\mathcal{M}}:=\mathcal{M}\rtimes_{\sigma^{\phi}}G . There is a canonical normal conditional expectation {\mathcal{E}}:\widehat{\mathcal{M}}\to\mathcal{M} 9 9 9 We avoid using the crossed product \mathcal{M}\rtimes\mathbb{R} because, by ( Takesaki_1979 , Sec. IX.4) , the existence of a normal conditional expectation onto \mathcal{M} forces the factor type to be preserved. In particular, even if \mathcal{M} is type III, its crossed product with \mathbb{R} is always semifinite (type II ∞ ) due to the presence of a faithful trace. . Finally, let \widehat{\phi}=\phi\circ{\mathcal{E}} be the dual weight on \widehat{\mathcal{M}} , which is also a normal faithful state. Theorem 4.19 (Haagerup reduction ( haagerup2009reductionmethodnoncommutativelpspaces , Theorem 2.1) ) . Following the notation above, there exists an increasing sequence (\mathcal{M}_{k})_{k\geq 1} of von Neumann subalgebras of \widehat{\mathcal{M}} satisfying: (i) each \mathcal{M}_{k} is finite with normal faithful trace \tau_{k} ; (ii) \bigcup_{k\geq 1}\mathcal{M}_{k} is w*-dense in \widehat{\mathcal{M}} ; (iii) there is a family of normal faithful conditional expectation ({\mathcal{E}}_{k})_{k\geq 1}\colon\widehat{\mathcal{M}}\to\mathcal{M}_{k} such that \widehat{\phi}\circ{\mathcal{E}}_{k}=\widehat{\phi},\quad\sigma_{t}^{\widehat{\phi}}\circ{\mathcal{E}}_{k}={\mathcal{E}}_{k}\circ\sigma_{t}^{\widehat{\phi}},\quad t\in\mathbb{R}; (iv) let \phi_{k}(\cdot):=\widehat{\phi}|_{\mathcal{M}_{k}}(\cdot)=\tau_{k}(d_{k}\cdot) for d_{k}\in L(G)\subset\mathcal{Z}(\widehat{\mathcal{M}}_{\widehat{\phi}}) so that \sigma_{t}^{\widehat{\phi}}(d_{k})=d_{k} ; Furthermore a_{k}\leq d_{k}\leq a_{k}^{-1} for some a_{k}\in\mathbb{R}^{+} ; (v) \lim_{k\to\infty}{\mathcal{E}}_{k}(x)=x in \sigma -strong topology for any x\in\widehat{\mathcal{M}} . (vi) \varphi_{k}\to\widehat{\varphi} in the weak topology on S(\mathcal{M}) , where \varphi\in S(\mathcal{M}) and \varphi_{k}:=\widehat{\varphi}|_{\mathcal{M}_{k}} . We apply Haagerup’s reduction Theorem 4.19 on GNS- \phi -detailed balance QMS (T_{t})_{t\geq 0} . We call {(\widehat{T}_{t})}_{t\geq 0} its canonical extension in \widehat{\mathcal{M}} as \widehat{T}_{t}:=T_{t}\otimes\operatorname{Id}_{B(l_{2}(G))}|_{\widehat{\mathcal{M}}} , and in ( Gao_Junge_LaRacuente_Li_2025 , Section 4.2) it is also shown to be also a GNS- \widehat{\phi} -detailed balanced QMS with fixed point algebra \widehat{\mathcal{N}}:=\mathcal{N}\rtimes_{\sigma^{\phi}}G . Moreover, the \widehat{\phi} -preseving conditional expectation \widehat{E}:\widehat{\mathcal{M}}\to\widehat{\mathcal{N}} is just the canonical extension of E:\mathcal{M}\to\mathcal{N} . It turns out that (\widehat{T_{t}})_{t\geq 0} is compatible with the finite approximations in Haagerup reduction, summarized in the theorem below. Theorem 4.20 ( ( Gao_Junge_LaRacuente_Li_2025 , Lemma 4.5) ) . Let (T_{t})_{t\geq 0}\colon\mathcal{M}\to\mathcal{M} be a GNS - \phi - symmetric quantum Markov semigroup. With the notation above and t\geq 0 , the following hold: (i) \widehat{T}_{t} commutes with {\mathcal{E}} , \widehat{\mathcal{E}} and {\mathcal{E}}_{k} . In particular, \widehat{T}_{t}(\mathcal{M}_{k})\subseteq\mathcal{M}_{k} . (ii) The restriction T_{t,k}:=\widehat{T}_{t}\bigm|_{\mathcal{M}_{k}}:\mathcal{M}_{k}\longrightarrow\mathcal{M}_{k} is a \tau_{k} - symmetric normal unital completely positive map. (iii) Let \mathcal{N}_{k}:=\mathcal{M}_{k}\cap\widehat{\mathcal{N}}\subseteq\mathcal{M}_{k} be the fixed - point algebra of T_{t,k} . Then E_{k}:=\widehat{E}\bigm|_{\mathcal{M}_{k}}:\mathcal{M}_{k}\longrightarrow\mathcal{N}_{k} is the \tau_{k} - preserving conditional expectation onto \mathcal{N}_{k} . Combining the preceding results with ( haagerup2009reductionmethodnoncommutativelpspaces , Eq. (6.4)) , which gives \widehat{E}\circ{\mathcal{E}}={\mathcal{E}}\circ\widehat{E}\quad\text{and}\quad\widehat{E}\circ{\mathcal{E}}_{k}={\mathcal{E}}_{k}\circ\widehat{E}, we obtain the following commuting diagrams Figure 3 . {\lx@inpgf@ignorespaces\mathcal{M}} {\lx@inpgf@ignorespaces\widehat{\mathcal{M}}} {\lx@inpgf@ignorespaces\mathcal{M}_{k}} {\lx@inpgf@ignorespaces\mathcal{M}} {\lx@inpgf@ignorespaces\widehat{\mathcal{M}}} {\lx@inpgf@ignorespaces\mathcal{M}_{k}} {\lx@inpgf@ignorespaces\mathcal{N}} {\lx@inpgf@ignorespaces\widehat{\mathcal{N}}} {\lx@inpgf@ignorespaces\mathcal{N}_{k}} \scriptstyle{\lx@inpgf@ignorespaces T_{t}} \scriptstyle{\lx@inpgf@ignorespaces\widehat{T}_{t}} \scriptstyle{\lx@inpgf@ignorespaces{\mathcal{E}}} \scriptstyle{\lx@inpgf@ignorespaces{\mathcal{E}}_{k}} \scriptstyle{\lx@inpgf@ignorespaces T_{t,k}} \scriptstyle{\lx@inpgf@ignorespaces E} \scriptstyle{\lx@inpgf@ignorespaces\widehat{E}} \scriptstyle{\lx@inpgf@ignorespaces{\mathcal{E}}} \scriptstyle{\lx@inpgf@ignorespaces{\mathcal{E}}_{k}} \scriptstyle{\lx@inpgf@ignorespaces E_{k}} \scriptstyle{\lx@inpgf@ignorespaces{\mathcal{E}}} \scriptstyle{\lx@inpgf@ignorespaces{\mathcal{E}}_{k}} Figure 3: Haagerup reduction commuting diagram: the quantum Markov semigroup T_{t} on \mathcal{M} , its finite approximants \mathcal{M}_{k} , and the conditional expectations E , E_{k} down to \mathcal{N} . 4.6 Proof of L^{p} version’s of \mathrm{PI}(p,p) We are now ready to prove our main theorem Theorem 4.17 . Unless otherwise stated, we use the notation established in previous sections. The outline is to prove that \mathrm{PI}(p,p) holds on the tracial slices (\mathcal{M}_{k},\tau_{k}) for all k\geq 1 and extend it to (\mathcal{M},\phi) by Haagerup’s reduction. Proof. We first verify that the spectral gap \alpha is inherited by T_{t,k} . Combining this with (iii) of Theorem 4.20 , it follows from Theorem 1.3 that \mathrm{PI}(p,p) holds on each slice (\mathcal{M}_{k},\tau_{k}) . Indeed, given \alpha>0 spectral gap for T_{t} , we have for x\in\mathcal{M} , \|T_{t}(x)-E(x)\|_{L^{2}(\mathcal{M})}\leq e^{-\alpha t}\|x-E(x)\|_{L^{2}(\mathcal{M})}. Recall that \widehat{T}_{t}:=T_{t}\otimes\operatorname{Id}_{B(\ell_{2}(G))}|_{\widehat{\mathcal{M}}} . By direct calculation, or equivalently by Proposition B.1 , using that the spectral gap of \operatorname{Id} is \alpha_{\operatorname{Id}}=\infty , we see that \widehat{T}_{t} has the same spectral gap \alpha as T_{t} . By isometry of norms for {\mathcal{E}}_{k} on L^{2} spaces using Lemma 4.12 , it follows that \widehat{T}_{t} has the same spectral gap \alpha when restricted to each slice (\mathcal{M}_{k},\tau_{k}) for k\geq 1 . In the sequel, we adopt the convention that all objects associated with the restricted semigroup T_{t,k} carry the subscript k . In particular, we set L_{k}:=\widehat{L}\bigl|_{\mathcal{M}_{k}},\qquad\mathcal{L}_{p,k}^{K}:=\widehat{\mathcal{L}}^{K}_{p}\bigl|_{\mathcal{M}_{k}},\qquad\Gamma^{(p,k)}_{\eta}:=\widehat{\Gamma}^{(p)}_{\eta}\bigl|_{\mathcal{M}_{k}\times\mathcal{M}_{k}}, so that L_{k} , \mathcal{L}^{K}_{p,k} and \Gamma^{(p,k)}_{\eta} are the Lindbladian, the Kosaki’s L^{p} Lindbladian and L^{p} gradient form of T_{t,k} on \mathcal{M}_{k} . Given k\geq 1 , let x_{k}\in(\mathcal{M}_{k})_{sa}\cap\widehat{\mathcal{M}}_{a}^{\sigma}\cap\operatorname{Dom}(L_{k}) , where \widehat{\mathcal{M}}_{a}^{\sigma} the set of all entire elements of \widehat{\mathcal{M}} , and d_{k} the density of \phi_{k}:=\widehat{\phi}|_{\mathcal{M}_{k}} under trace \tau_{k} . By (iv) of Theorem 4.19 , d_{k}\in L(G)\subset\widehat{\mathcal{N}}:=\mathcal{N}\rtimes_{\sigma^{\phi}}G and d_{k}\in\mathcal{M}_{k} , it follows that d_{k}\in\mathcal{N}_{k}=\mathcal{M}_{k}\cap\widehat{\mathcal{N}} . We can assert the following \mathcal{L}_{p,k}^{K}(d_{k}^{\frac{1}{2}}x_{k}d_{k}^{\frac{1}{2}})=d_{k}^{\frac{1}{2}}L_{k}(x_{k})d_{k}^{\frac{1}{2}}, (52) which agrees with Definition 4.9 . Thus, we can write the \mathrm{PI}(p,p) in Kosaki’s L^{p} spaces as in Proposition 4.16 . For all p=2 and p\geq 3 , and let \iota_{\eta}(x)=h_{\phi}^{\eta}xh_{\phi}^{(1-\eta)} 10 10 10 We omit the \phi -subscript on \iota whenever the norm’s subscript makes the state clear. It is also known that L^{p}(\mathcal{N}) is naturally isometrically embedded in L^{p}(\mathcal{M}) Junge_Xu_2003 whenever there exists a conditional expectation E:\mathcal{M}\to\mathcal{N} , so no further distinction is needed. for \eta\in[0,1] , \bigl\|\iota_{\frac{1}{2}}(x_{k}-E_{k}(x_{k}))\bigr\|_{p,\phi_{k},\frac{1}{2}}\leq\frac{p}{\sqrt{2\alpha}}\bigl\|\iota_{\frac{1}{2}}\bigl(\Gamma_{\frac{1}{2}}^{(p,k)}(x_{k},x_{k})\bigr)\bigr\|_{\frac{p}{2},\phi_{k},\frac{1}{2}}^{1/2}. As we are working with self-adjoint elements we denote \|\cdot\|_{p,\phi,\frac{1}{2}} just as \|\cdot\|_{p,\phi} in the following. Taking the limit k\to\infty , the above relations remain valid, and hence \mathrm{PI}(p,p) holds in \widehat{\mathcal{M}} and thus in \mathcal{M} . To see this, for {x}\in{\mathcal{B}}:={\mathcal{M}}_{sa}\cap{\mathcal{M}}^{\sigma}_{a}\cap\operatorname{Dom}(L)\subset\mathcal{M}\subset\widehat{\mathcal{M}} , we have x_{k}:={\mathcal{E}}_{k}({x})\in(\mathcal{M}_{k})_{sa} , {\mathcal{E}}_{k,1}\circ\iota_{\eta}=\iota_{\eta}\circ{\mathcal{E}}_{k} and E_{k}\circ{\mathcal{E}}_{k}={\mathcal{E}}_{k}\circ\widehat{E}={\mathcal{E}}_{k}\circ E 11 11 11 Since \widehat{E} is the canonical extension of E , we write E for \widehat{E} when acting on x\in\mathcal{M} . from Figures 2 and 3 , the left-hand side rewrites as \bigl\|\iota_{\frac{1}{2}}(x_{k}-E_{k}(x_{k}))\bigr\|_{p,\phi_{k}}=\bigl\|\iota_{\frac{1}{2}}({\mathcal{E}}_{k}(x)-E_{k}({\mathcal{E}}_{k}(x)))\bigr\|_{p,\phi_{k}}=\bigl\|{\mathcal{E}}_{k,1}\bigl(\iota_{\frac{1}{2}}(x-E(x))\bigr)\bigr\|_{p,\phi_{k}}. (53) Taking the limit k\to\infty and applying (v) of Theorem 4.19 , \lim_{k\to\infty}{\mathcal{E}}_{k}(x)=x in \sigma -strong topology for any x\in{\mathcal{B}}\subset\widehat{\mathcal{M}} give us \lim_{k\to\infty}\bigl\|{\mathcal{E}}_{k,1}\bigl(\iota_{\frac{1}{2}}(x-E(x))\bigr)\bigr\|_{p,\phi_{k}}=\bigl\|\iota_{\frac{1}{2}}(x-E(x))\bigr\|_{p,\widehat{\phi}}=\bigl\|\iota_{\frac{1}{2}}(x-E(x))\bigr\|_{p,\phi}, (54) where the last equality uses Lemma 4.12 where the L^{p} extensions of {\mathcal{E}}:\widehat{\mathcal{M}}\to\mathcal{M} is a contractive projection. Similarly, for the RHS, we can remove one conditional expectation as in eqs. 53 and 54 , \|\iota_{\frac{1}{2}}\bigl(\Gamma_{\frac{1}{2}}^{(p,k)}(x_{k},x_{k})\bigr)\bigr\|_{\frac{p}{2},\phi_{k}}^{1/2}\leq\|\iota_{\frac{1}{2}}\bigl(\widehat{\Gamma}_{\frac{1}{2}}^{(p)}(x,{\mathcal{E}}_{k}(x))\bigr)\bigr\|_{\frac{p}{2},\widehat{\phi}}^{1/2} We know \widehat{\Gamma}_{\frac{1}{2}}^{(p)}(x,{\mathcal{E}}_{k}(x)) is linear on {\mathcal{E}}_{k}(x) from Definition 4.14 , \widehat{\Gamma}_{\frac{1}{2}}^{(p)}(x,{\mathcal{E}}_{k}(x))=\Gamma(\sigma^{\widehat{\phi}}_{-\frac{i}{2p}}(x),\sigma^{\widehat{\phi}}_{-\frac{i}{2p}}({\mathcal{E}}_{k}(x)))=\Gamma(\sigma^{\widehat{\phi}}_{-\frac{i}{2p}}(x),{\mathcal{E}}_{k}(\sigma^{\widehat{\phi}}_{-\frac{i}{2p}}(x))), where the second equality is due to \sigma_{t}^{\widehat{\phi}}\circ{\mathcal{E}}_{k}={\mathcal{E}}_{k}\circ\sigma_{t}^{\widehat{\phi}} by (iii) of Theorem 4.19 . Let y:=\sigma^{\widehat{\phi}}_{-\frac{i}{2p}}(x)=\sigma^{\widehat{\phi}}_{-\frac{i}{2p}}\bigl|_{\mathcal{M}}(x)=\sigma^{\phi}_{-\frac{i}{2p}}(x)\in\mathcal{M} for x\in\mathcal{M}_{a}^{\sigma} , and y_{k}:={\mathcal{E}}_{k}(y) . We aim to prove that \lim_{k\to\infty}\|\iota_{\frac{1}{2}}\bigl(\widehat{\Gamma}(y,y_{k})\bigr)\bigr\|_{\frac{p}{2},\widehat{\phi},\frac{1}{2}}^{1/2}\leq\|\iota_{\frac{1}{2}}\bigl(\widehat{\Gamma}(y,y)\bigr)\bigr\|_{\frac{p}{2},\widehat{\phi},\frac{1}{2}}^{1/2}=\|\iota_{\frac{1}{2}}\bigl(\Gamma(y,y)\bigr)\bigr\|_{\frac{p}{2},\phi,\frac{1}{2}}^{1/2}, (55) where the last equality is due to the isometry of norm from Lemma 4.12 . There are technical complications due to the domain issue for the gradient form \Gamma : \operatorname{Dom}(L) is not closed under multiplication. In particular, for y,y_{k}\in\operatorname{Dom}(L) one may have y^{*}y_{k}\notin\operatorname{Dom}(L)\subset\operatorname{Dom}({\widehat{L}}) . To work around this issue, we employ the \epsilon -regularization, so that all expressions are well defined, and then pass to the limit as \epsilon\to 0 . Recall in this framework, for all x,y\in{\mathcal{A}}:=\operatorname{Dom}(L)\cap\mathcal{M}\subset\operatorname{Dom}(\widehat{L})\cap\widehat{\mathcal{M}} , \widehat{\Gamma}_{\epsilon}(x,y):=\frac{1}{2}\Bigl(x^{*}\widehat{L}_{\epsilon}(y)+\widehat{L}_{\epsilon}(x)^{*}y-\widehat{L}_{\epsilon}(x^{*}y)\Bigr),\text{ where }\widehat{L}_{\epsilon}:=\frac{\widehat{L}}{1+\epsilon\widehat{L}}. (56) From ( cipriani_derivations_2003 , Prop. 2.5) that for \epsilon>0 the resolvent maps (I+\epsilon\widehat{L})^{-1} are bounded, completely positive contractions on L^{2}({\widehat{\mathcal{M}}}) and the regulated operators \widehat{L}_{\epsilon} are completely bounded endomorphisms of \widehat{\mathcal{M}} . These properties guarantee that the regularized gradient form above is well defined and uniformly bounded ( Remark C.3 ). As shown in Appendix C , detailed balance for L passes to the regulator L_{\epsilon} , so that the tracial \mathrm{PI}(p,p) inequality holds for the regularization L_{k,\epsilon}={L_{k}}({I+\epsilon L_{k}})^{-1} on \mathcal{M}_{k} , and the spectral gap satisfies \alpha_{\epsilon}={\alpha}/({1+\epsilon\alpha}) . Meanwhile, we restrict to the simplified case in which we assume that {\mathcal{A}} is a weak ∗ -dense * -algebra, given by {\mathcal{A}}=\operatorname{Dom}(L)\cap\mathcal{M} . By Proposition C.2 , \widehat{\Gamma}_{\epsilon}(x,y)\to\widehat{\Gamma}(x,y) strongly for all x,y\in{\mathcal{A}} . To prove eq. 55 , we first work with the regularized operators \widehat{\Gamma}_{\epsilon} . By (v) of Theorem 4.19 , we have \lim_{k\to\infty}y_{k}=y in the \sigma -strong topology. Since \widehat{L}_{\epsilon}=\widehat{L}(1+\epsilon\widehat{L})^{-1} is bounded for each \epsilon>0 , it is straightforward to check that \lim_{k\to\infty}\widehat{\Gamma}_{\epsilon}(y,y_{k})=\lim_{k\to\infty}\frac{1}{2}\bigl(\widehat{L}_{\epsilon}(y^{*})y_{k}+y^{*}\widehat{L}_{\epsilon}(y_{k})-\widehat{L}_{\epsilon}(y^{*}y_{k})\bigr)=\widehat{\Gamma}_{\epsilon}(y,y) (57) strongly. After taking \epsilon\downarrow 0 and applying Lemma C.1 , we obtain the equality form of eq. 55 in \frac{p}{2} -norm. For the general case when there is no domain *-algebra {\mathcal{A}} , we use the weak ∗ convergence definition extended by ( cipriani_derivations_2003 , Lemma 9.1) ( eq. 15 ). \Gamma(x,y):=\mathrm{w}^{*}\text{-}\lim_{\epsilon\downarrow 0}\Gamma_{\epsilon}(x,y). We defer the details of proving this general case of eq. 55 to Appendix D . The above arguments establish the L^{p} version of \mathrm{PI}(p,p) on (\mathcal{M},\phi) for any x\in{\mathcal{B}}=\mathcal{M}_{sa}\cap\mathcal{M}_{a}^{\sigma}\cap\operatorname{Dom}(L) by (ii) (density of \cup_{k}\mathcal{M}_{k} ) in Theorem 4.19 . Equivalently, in Haagerup’s L^{p} -space functorial formulation, for dense a=h_{\phi}^{1/2p}xh_{\phi}^{1/2p}\in L^{p}(\mathcal{M})_{sa} where x\in{\mathcal{B}} , \|a-E_{p}^{\phi}(a)\|_{L^{p}(\mathcal{M})}\leq\frac{p}{\sqrt{2\alpha}}\|\Gamma_{p}(a,a)\|_{L^{\frac{p}{2}}(\mathcal{M})}^{1/2}. Haagerup’s functorial construction shows that the spaces L^{p}(\mathcal{M}) do not depend on any particular choice of trace, so the L^{p} inequality above is canonically well defined for all a\in L^{p}(\mathcal{M})_{sa} in Haagerup’s L^{p} space. ∎ 5 Examples and Applications 5.1 Khintchine inequalities with Rademacher random variables We present a degree-one example in which the \mathrm{PI}(p,p) inequality yields Khintchine-type estimates. This example also shows that, in the non-self-adjoint extension \mathrm{PI}(p,p) from 4.18 , the second term (the \Gamma_{p}(a^{*},a^{*}) contribution) is genuinely necessary and cannot be removed. Let \Omega=\{-1,1\}^{n} be equipped with the uniform probability measure, and let \varepsilon_{1},\dots,\varepsilon_{n} denote the coordinate (Rademacher) functions. We consider the von Neumann algebra {\mathcal{A}}:=L_{\infty}(\Omega)\bar{\otimes}\mathcal{M}\cong L_{\infty}(\Omega,\mathcal{M}), equipped with the product state \psi:=\mathbb{E}\otimes\phi , where \mathbb{E} is the uniform expectation on L_{\infty}(\Omega) . Concretely, elements of {\mathcal{A}} are bounded \mathcal{M} -valued functions on \Omega ; in particular, we will often restrict to the degree-one subspace consisting of elements of the form x=\sum_{i=1}^{n}\varepsilon_{i}\times x_{i},\qquad x_{i}\in\mathcal{M}. For each i , let {\mathbb{E}}_{i}:{\mathcal{A}}\to{\mathcal{A}} denote the conditional expectation onto the von Neumann subalgebra generated by \{\varepsilon_{j}:j\neq i\} , acting trivially on \mathcal{M} . We define the L=\sum_{i=1}^{n}L_{i},\qquad L_{i}=\operatorname{Id}-{\mathbb{E}}_{i}. (58) One checks that the spectral gap of L is 1 and L is GNS- \psi -detailed balance, since each L_{i} is a projection; moreover E_{fix}={\mathbb{E}}\otimes\operatorname{Id}_{\mathcal{M}} and E_{fix}(x)=0 . The associated gradient form for x=\sum_{i=1}^{n}\varepsilon_{i}\times x_{i} decomposes as \Gamma=\sum_{i=1}^{n}\Gamma_{i},\quad\Gamma_{i}(x,x)=x_{i}^{*}x_{i},\quad\Gamma_{i}(x^{*},x^{*})=x_{i}x_{i}^{*}. Denote by h_{\phi} the density of \phi , and set a=xh_{\phi}^{1/p}=\sum_{i=1}^{n}\varepsilon_{i}\times a_{i},\qquad a_{i}=x_{i}h_{\phi}^{1/p}. Following Equation 43 , E_{p}^{\phi}(\cdot)=E(\cdot)h_{\phi}^{\frac{1}{p}} and thus E_{p}^{\phi}(a)=0 . By 4.18 and density, for all a_{i}\in L^{p}(\mathcal{M}) and 2\leq p<\infty , \|a-E_{p}^{\phi}(a)\|_{L^{p}({\mathcal{A}})}=\|\sum_{i=1}^{n}\varepsilon_{i}\times a_{i}\|_{L^{p}({\mathcal{A}})}\leq\frac{p}{\sqrt{2}}\Bigl(\bigl\|(\sum_{i=1}^{n}a_{i}^{*}a_{i})^{1/2}\bigr\|_{L^{p}(\mathcal{M})}+\bigl\|(\sum_{i=1}^{n}a_{i}a_{i}^{*})^{1/2}\bigr\|_{L^{p}(\mathcal{M})}\Bigr). Our application gives a suboptimal constant for noncommutative Khintchine inequalities in Lust-Piquard_Pisier_1991 with constant O(\sqrt{p}) . 5.2 Sub-exponential concentration inequalities It is well-known that Poincaré inequalities imply concentration inequalities. We establish that our \mathrm{PI}(p,p) implies sub-exponential concentration inequality, which in the semi-commutative case is established in ( huang2021poincare , Thm 2.7) . Corollary 5.1 ( \mathrm{PI}(p,\infty) with O(p) constant) . Given the same conditions as the 4.18 , for a:=h_{\phi}^{\frac{\eta}{p}}xh_{\phi}^{\frac{1-\eta}{p}}\in L^{p}(\mathcal{M}) , x\in\mathcal{M}^{\sigma}_{a} , \|a-E_{p}^{\phi}(a)\|_{L^{p}(\mathcal{M})}\leq\frac{p}{\sqrt{2\alpha}}\|x\|_{\mathrm{Lip}_{\Gamma}}, (59) where the Lipschitz semi-norm gao_fisher_2020 is \|x\|_{\mathrm{Lip}_{\Gamma}}:=\max\{\|\Gamma(x,x)\|^{1/2},\|\Gamma(x^{*},x^{*})\|^{1/2}\} . Proof. The result follows from the norm comparison in the tracial case and the modular covariance \Gamma(\sigma_{t}(x),\sigma_{t}(y))=\sigma_{t}(\Gamma(x,y)) , which gives \|\sigma_{t}(x)\|_{\mathrm{Lip}_{\Gamma}}=\|x\|_{\mathrm{Lip}_{\Gamma}} . For entire elements x\in\mathcal{M}^{\sigma}_{a} , the above also applies to eq. 45 . ∎ Remark 5.2 . 5.1 may be compared with ( Gao_Junge_LaRacuente_Li_2025 , Theorem 5.12) , where a \mathrm{PI}(p,\infty) inequality with constant O(\sqrt{p}) is obtained under the assumption \alpha_{\mathrm{MLSI}}>0 , rather than an L^{2} spectral gap. In view of the discussion in the introduction, the two results should be regarded as independent. We follow the procedure as in Gao_Junge_LaRacuente_Li_2025 for sub-exponential concentration property. Definition 5.3 . For x\in\mathcal{M} , we write {\mathbb{P}}_{\phi}(|x|>t)\hskip 2.84544pt\leq\hskip 2.84544pt\epsilon if there exists a projection e\in\mathcal{M} such that \|exe\|_{\infty}\leq t\quad\text{and}\quad\phi(1-e)\leq\epsilon. Lemma 5.4 ( ( Gao_Junge_LaRacuente_Li_2025 , Lemma 5.14) ) . Let x\in L_{p}(\mathcal{M},\phi) with 1
t)\hskip 2.84544pt\leq\hskip 2.84544pt2\Bigl(\frac{t}{4}\Bigr)^{-p}\|x\|_{L_{p}(\mathcal{M})}^{p}\hskip 2.84544pt.
Corollary 5.5
(Sub-exponential Concentration)
.
Let
T_{t}=e^{-tL}
be a GNS–
\phi
-detailed balance QMS with spectral gap
\alpha>0
and conditional expectation
E
onto its fixed-point algebra. Then for any
x\in\mathcal{M}
and large
t>0
,
{\mathbb{P}}_{\phi}(|x-E(x)|>t)\leq 2\exp\Bigl(-\frac{\sqrt{\alpha}t}{2\sqrt{2}e\|x\|_{\mathrm{Lip}_{\Gamma}}}\Bigr).
Proof.
By
Lemma
5.4
and
5.1
,
{\mathbb{P}}_{\phi}(|x-E(x)|>t)\leq 2\Bigl(\frac{t}{4}\Bigr)^{-p}\|x-E(x)\|_{L_{p}(\mathcal{M})}^{p}\leq 2\Bigl(\frac{4p\|x\|_{\mathrm{Lip}_{\Gamma}}}{\sqrt{2\alpha}t}\Bigr)^{p}\hskip 2.84544pt.
Optimizing in
p
gives
p=\frac{\sqrt{2\alpha}t}{4e\|x\|_{\mathrm{Lip}_{\Gamma}}}\geq 3
for
t
large, which implies the claim.
∎
This is a generalization of the sub-exponential concentration obtained in the semi-commutative case
(
huang2021poincare
, Thm 2.7)
. Under only spectral gaps, one cannot expect tail decay faster than exponential, since this hypothesis does not imply stronger concentration phenomena. In contrast,
(
junge2015noncommutative
, Cor 5.15)
proved Gaussian concentration under the additional assumption of MLSI, which is strictly stronger.
Birth-death process.
We illustrate this point with a birth-death process example showing the failure of geometric Talagrand’s inequality under a uniform spectral gap assumption. As shown in
(
gao_fisher_2020
, Thm 6.10)
, the
\mathrm{PI}(p,p)
(also argument of
5.1
) with constant
O(\sqrt{p})
implies a bound in Wasserstein-1 distance
WA_{1}
, which in turn gives geometric Talagrand’s inequality. This shows that
\mathrm{PI}(p,p)
with
O(p)
constants is genuinely weaker than Talagrand or Gaussian concentration, and cannot be improved under a just spectral gap assumption.
Definition 5.6
(Geometric Talagrand inequality)
.
Let
T_{t}=e^{-tL}
be a QMS on
(\mathcal{M},\tau)
with fixed-point algebra
\mathcal{N}
and conditional expectation
E\colon\mathcal{M}\to\mathcal{N}
.
We say that
T_{t}
satisfies the geometric Talagrand inequality if there exists
c>0
such that for every normal state
\rho
,
\|\rho-E_{*}(\rho)\|_{\Gamma^{*}}\leq c\sqrt{D(\rho\|E_{*}(\rho))},
(60)
where
\displaystyle\|\rho\|_{\Gamma^{*}}:=\sup\{\bigl|\tau(\rho x)\bigr|:E(x)=0,{x=x^{*},\ \|x\|_{\mathrm{Lip}_{\Gamma}}\leq 1}\}
and
E_{*}
is the predual of
E
.
Example 5.7
(Birth-death Process)
.
We consider a 1D chain of length n, and define the birth-death process from the edge, which gives
\displaystyle L_{j,j+1}(x)
\displaystyle=e^{\beta/2}L_{e_{j,j+1}}(x)+e^{-\beta/2}L_{e_{j+1,j}}(x)
\displaystyle=e^{\beta/2}\bigl(e_{j+1,j+1}x+xe_{j+1,j+1}-2e_{j+1,j}xe_{j,j+1}\bigr)+e^{-\beta/2}\bigl(e_{j,j}x+xe_{j,j}-2e_{j,j+1}xe_{j+1,j}\bigr),
where in the quantum notation
e_{j,j+1}=\ket{j}\bra{j+1}
. The total birth-death Lindbladian is
L=\sum_{j=1}^{n-1}L_{j,j+1}.
It is GNS-
\mu
-detailed balanced for the thermal state
\mu:=Z_{\beta}^{-1}(e^{-\beta k})_{k=1}^{n}
, where
Z_{\beta}:=\sum_{j=1}^{n}e^{-\beta j}
the partition function (constant). We work on the diagonal (commutative) part
L^{\mathrm{diag}}
and
Miclo1999DiscreteHardy
gives the spectral gap
\alpha(L^{\mathrm{diag}})=O(1)
.
We show a violation of Talagrand’s inequality
Equation
60
with any constant
c=c(\beta)
independent of
n
.
Following
(
gao_fisher_2020
, Remark 6.9)
, consider
e_{A}
a projection in region
A
and
\mu_{A}:=\frac{e_{A}}{\mu(e_{A})}
. Choose
f=\sum_{j=1}^{n}\frac{j}{\sqrt{2n\cosh{\beta/2}}}\ket{j}\bra{j}
so
\|f\|_{\mathrm{Lip}_{\Gamma}}=\|\sum_{j}|[e^{\beta/4}e_{j,j+1},f]|^{2}+\sum_{j}|[e^{-\beta/4}e_{j+1,j},f]|^{2}\|^{1/2}\leq 1.
Let
A=\{1\}
and
B=\{n\}
. Since
\mu_{A}
and
\mu_{B}
are supported on
A
and
B
,
\|\mu_{A}-\mu_{B}\|_{\Gamma^{*}}=\sup_{x=x^{*},\ \|x\|_{\mathrm{Lip}_{\Gamma}}\leq 1}\bigl|\mu_{A}(x)-\mu_{B}(x)\bigr|\geq|\mu_{A}(f)-\mu_{B}(f)|=O(\sqrt{n}e^{\beta n}).
Moreover,
D\Big(\mu_{A}||E(\mu_{A})\Big)\leq-\ln\mu(A).
Comparing the right-hand side
c(\beta)(\sqrt{D(\mu_{A}||E(\mu_{A}))}+\sqrt{D(\mu_{B}||E(\mu_{B}))})=O(\sqrt{n})
with left-hand side
\|\mu_{A}-\mu_{B}\|_{\Gamma^{*}}=O(\sqrt{n}e^{\beta n})
gives a contradiction of
Equation
60
for
c=c(\beta)
as
n\to\infty
.
The birth-death process above provides an example showing that a subgaussian-type estimate where
K
is a universal constant
\|f-E(f)\|_{p}\leq Kc(\beta)\sqrt{p}\|f\|_{\mathrm{Lip}_{\Gamma}}
cannot hold uniformly in
n
with a constant depending only on
\beta
.
Indeed, on the diagonal subalgebra,
\mu
violates Talagrand’s inequality
eq.
60
with any
c(\beta)
independent of
n
, hence one cannot have an
O(\sqrt{p})
version of
\mathrm{PI}(p,\infty)
.
In contrast,
L^{\mathrm{diag}}
has a spectral gap of
O(1)
(for fixed
\beta
), so the
\mathrm{PI}(p,p)
inequality with constant
O(p)
still holds.
5.3
Semigroup diameter
As an application of
Theorem
1.4
, we obtain a universal estimate on the GNS-
\phi
-detailed balanced QMS diameter in finite dimensions depending on the spectral gap
\alpha
and the spectrum of density of
\phi
.
Corollary 5.8
(Semigroup diameter)
.
Let
{\mathbb{M}}_{d}
be a finite-dimensional matrix algebra equipped with the unnormalised trace
\Tr
and large dimension say
d>20
,
and denote by
\|\cdot\|_{S_{p}}:=\bigl(\Tr|\cdot|^{p}\bigr)^{1/p}
the Schatten–
p
norm.
Let
(T_{t})_{t\geq 0}
be a GNS-detailed balanced quantum Markov semigroup with respect to
\phi(\cdot)=\Tr(D_\phi\cdot)
and spectral gap
\alpha>0
.
Then for any
x\in{\mathbb{M}}_{d}
,
\|x-E(x)\|_{\infty}\leq\mathrm{diam}_{\Gamma}\|x\|_{\mathrm{Lip}_{\Gamma}},
where
\mathrm{diam}_{\Gamma}=\frac{e\log(\lambda_{\min}^{-1})}{\sqrt{2\alpha}}
and
\lambda_{\min}
denotes the minimal eigenvalue of
D_{\phi}
.
Proof.
We write
D
as shorthand for
D_{\phi}
and WLOG
x\in\mathring{{\mathbb{M}}_{d}}
. Since
x=D^{-\frac{1}{2p}}\bigl(D^{\frac{1}{2p}}xD^{\frac{1}{2p}}\bigr)D^{-\frac{1}{2p}},
the three-factor Hölder inequality with exponents
(2p,p,2p)
and
5.1
gives
\displaystyle\|x\|_{\infty}
\displaystyle\leq\|x\|_{S_{p}}\leq\bigl\|D^{-\tfrac{1}{2p}}\bigr\|_{S_{2p}}^{2}\bigl\|D^{\tfrac{1}{2p}}xD^{\tfrac{1}{2p}}\bigr\|_{S_{p}}\leq(\lambda_{\min})^{-\tfrac{1}{p}}\|\iota_{1/2}(x)\|_{p,\phi,1/2}\leq(\lambda_{\min})^{-\tfrac{1}{p}}\frac{p}{\sqrt{2\alpha}}\|x\|_{\mathrm{Lip}_{\Gamma}},
Now choose
p=\log\bigl(\lambda_{\min}^{-1}\bigr)\geq\log\bigl(d\bigr)>3
,
so that
(\lambda_{\min})^{-1/p}=e
. Substituting back leads to the desired inequality.
∎
Appendix A
Markov dilation and proof of
Proposition
3.7
Before proving the noncommutative convex-chain estimate for gradient forms, we first
recall the framework of Markov dilation, which lifts a quantum Markov semigroup to a
multiplicative level.
This allows the Dirichlet form to be written as a square and enables the
noncommutative chain rule.
Markov dilation embeds quantum Markov semigroups into larger operator-algebraic structures that carry probabilistic interpretations. The foundational work of Kümmerer and Maassen
Kümmerer_Maassen_1987
introduced noncommutative dilation in which the enlarged algebra remains essentially commutative, and the first author and Mei
junge-mei
extended this framework to the setting of tracial von Neumann algebras. In
wu2023
, the author provided a probabilistic construction of Markov dilation. In the context of quantum physics, such dilations are interpreted as providing a reversible microscopic model of systems interacting with thermal environments. We recall the definition from
(
junge-mei
, Section 2.2)
.
Definition A.1
(Markov dilation)
.
Let
(T_{t})_{t\geq 0}
be a quantum Markov semigroup on a tracial von Neumann algebra
(\mathcal{M},\tau)
.
We say that
(T_{t})
admits a Markov dilation if there exists an increasing family of tracial von Neumann algebras
\mathcal{M}\subset\widetilde{\mathcal{M}}_{t}\subset\widetilde{\mathcal{M}},\quad t\geq 0,
that is, an increasing filtration of
\widetilde{\mathcal{M}}_{t}
, together with
•
normal, trace-preserving conditional expectations
E_{t}:\widetilde{\mathcal{M}}\to\widetilde{\mathcal{M}}_{t}
,
•
*
–homomorphisms
\pi_{t}:\mathcal{M}\to\widetilde{\mathcal{M}}
which adapted to the filtration, i.e.
\pi_{t}(\mathcal{M})\subset\widetilde{\mathcal{M}}_{t}
, for each
t\geq 0
,
such that for all
0\leq s 0,\lambda_{j}\geq 0
and
\sum_{j}\lambda_{j}=1
. In this way, norm convergence is recovered at the expense
of convexification.
We begin by introducing some definitions and results from
JungeRicardShlyakhtenko2025
.
Definition D.1
(noncommutative Diffusion QMS)
.
A semigroup
(T_{t})
is called a noncommutative diffusion process if for every
x\in\dom(L)\cap\mathcal{M}
the gradient form satisfies
\Gamma(x,x)\in L^{1}(\mathcal{M})
.
Proposition D.2
.
(T_{t})
is noncommutative diffusion if
\Gamma_{\epsilon}(x,x)\to\Gamma(x,x)\in L^{1}(\mathcal{M})
(65)
in
\sigma(L_{1}(\mathcal{M}),\mathcal{M})
topology. This implies
\Gamma(x,x)\in\overline{\mathrm{conv}(\Gamma_{\epsilon})}^{\|\|_{L^{1}(\mathcal{M})}}
by Hahn-Banach. Thus, there exists a family of Lindbladians
L_{\epsilon_{j}}
, and the gradient form of convex combinations of Lindbladians also converges
\Gamma_{\sum_{j}\lambda_{j}L_{\epsilon_{j}}}(x,y)=\sum_{j}\lambda_{j}\Gamma_{\epsilon_{j}}(x,y)\to\Gamma(x,y)
in
\|\cdot\|_{L^{1}(\mathcal{M})}
as
\{\epsilon_{j}\}\downarrow 0
, for some
\epsilon_{j}>0,\lambda_{j}\geq 0
and
\sum_{j}\lambda_{j}=1
.
We work with a convex
combination of
\epsilon_{j}
-regularized operators
\widehat{L}_{\epsilon_{j}}
and then
let
\epsilon_{j}\downarrow 0
. The corresponding spectral gap is
\overline{\alpha}_{\epsilon}=\sum_{j}\lambda_{j}\tfrac{\alpha}{1+\epsilon_{j}\alpha}
as in
Appendix
C
.
By similar arguments as in
Lemma
C.1
, given
L^{1}
convergence and uniformly boundedness of
\Gamma_{\sum_{j}\lambda_{j}L_{\epsilon_{j}}}(x,x)
for
x\in\operatorname{Dom}(L)
, we obtain
L^{p/2}
-norm convergence.
We are now ready to prove
eq.
55
. For
y\in\mathcal{M},y_{k}\in\mathcal{M}_{k}
we obtain
\displaystyle\frac{p}{\sqrt{2\alpha}}\lim_{k\to\infty}\|\iota_{\frac{1}{2}}(\widehat{\Gamma}(y,y_{k}))\|_{\frac{p}{2},\widehat{\phi},\frac{1}{2}}^{1/2}
\displaystyle=\lim_{\begin{subarray}{c}\epsilon_{j}\downarrow 0\\
\forall j\end{subarray}}\frac{p}{\sqrt{2\overline{\alpha}_{\epsilon}}}\lim_{k\to\infty}\Bigl\|\iota_{\frac{1}{2}}\Bigl(\sum_{j}\lambda_{j}\widehat{\Gamma}_{\epsilon_{j}}(y,y_{k})\Bigr)\Bigr\|_{\frac{p}{2},\widehat{\phi},\frac{1}{2}}^{1/2}
\displaystyle\leq\lim_{\begin{subarray}{c}\epsilon_{j}\downarrow 0\\
\forall j\end{subarray}}\frac{p}{\sqrt{2\overline{\alpha}_{\epsilon}}}\sum_{j}\lambda_{j}\lim_{k\to\infty}\|\iota_{\frac{1}{2}}(\widehat{\Gamma}_{\epsilon_{j}}(y,y_{k}))\|_{\frac{p}{2},\widehat{\phi},\frac{1}{2}}^{1/2}
\displaystyle=\lim_{\begin{subarray}{c}\epsilon_{j}\downarrow 0\\
\forall j\end{subarray}}\frac{p}{\sqrt{2\overline{\alpha}_{\epsilon}}}\sum_{j}\lambda_{j}\|\iota_{\frac{1}{2}}(\widehat{\Gamma}_{\epsilon_{j}}(y,y))\|_{\frac{p}{2},\widehat{\phi},\frac{1}{2}}^{1/2}
\displaystyle=\frac{p}{\sqrt{2\alpha}}(\sum_{j}\lambda_{j})\|\iota_{\frac{1}{2}}(\widehat{\Gamma}(y,y))\|_{\frac{p}{2},\widehat{\phi},\frac{1}{2}}^{1/2}=\frac{p}{\sqrt{2\alpha}}\|\iota_{\frac{1}{2}}(\widehat{\Gamma}(y,y))\|_{\frac{p}{2},\widehat{\phi},\frac{1}{2}}^{1/2},
where the inequality follows from Minkowski’s inequality, and the third line follows from
eq.
57
and
Lemma
C.1
. Finally, we pick
\sum_{j}\lambda_{j}=1
to prove
eq.
55
.
Acknowledgments.
MJ was partially supported by NSF Grant DMS-2247114. JW acknowledges support from the Spring 2025 Graduate Travel Award. The authors thank the Institute for Pure and Applied Mathematics (IPAM), which is supported by the
National Science Foundation (Grant No. DMS-1925919), for its hospitality, during which the initial
idea of this project was formulated. The authors thank Joel Tropp for helpful discussions.
JW also thanks Roy Araiza, Jihong Cai, David Jekel, Mark Klinger, Rolando de Santiago,
Vincent Villalobos, and Peixue Wu for helpful discussions.
Data Availability.
Data sharing not applicable to this article as no datasets were generated or
analyzed during the current study.
Conflict of interest.
The authors have no conflicts of interest to declare that are relevant to the
content of this article.