--- title: Mixed-State Deep Thermal Ensemble url: https://www.emergentmind.com/topics/mixed-state-deep-thermal-ensemble type: topic --- # Mixed-State Deep Thermal Ensemble Searching arXiv for the cited works to ground the article and verify bibliographic details. arXiv search query: 2505.07795 Mixed state deep thermalization arXiv search query: 2507.14135 Do mixed states exhibit deep thermalisation? arXiv search query: 2208.10542 Solvable model of deep thermalization with distinct design times arXiv search query: 1407.6136 Thermal states of random quantum many-body systems arXiv search query: 2508.00809 Statistical Mechanics of Random Mixed State Ensembles with Fixed Energy Mixed-state deep thermal ensemble denotes a family of constructions in which the object of interest is not a single reduced density matrix but a probability distribution over mixed states, typically attached to a local subsystem and generated by conditioning, tracing, or imposing thermodynamic constraints. In the narrow sense used in recent work on deep thermalisation, it is the universal late-time ensemble of mixed subsystem states produced either by incomplete measurements on a complementary region or by unitary evolution from an initially mixed state, after reformulating the maximum-entropy principle on an augmented purification [2505.07795], [2507.14135]. In a broader statistical-mechanical sense, closely related ensembles arise from random Hamiltonians at finite temperature, from fixed-energy measures over density matrices, and from variational multiscale ansätze for Gibbs-state preparation [1407.6136], [2508.00809], [2210.16419]. This suggests that the topic sits at the intersection of quantum thermalisation, random-state theory, and nonequilibrium quantum information. ## 1. Conceptual framework and basic definitions Deep thermalisation is stronger than ordinary thermalisation. In the projected-ensemble formulation, a bipartite pure state $\ket{\Psi}_{AB}$ defines an ensemble of conditional pure states on $A$ by \[ \mathcal{E} = \{ (p(z), \ket{\psi_z}_A):\ z = 1, \dots d_B \}, \] with \[ p(z) = \| \, {}_B\!\braket{z}{\Psi}_{AB} \|^2, \qquad \ket{\psi_z}_A = {}_B\!\braket{z}{\Psi}_{AB} / \sqrt{p(z)}. \] The $k$-th moment operator is \[ \rho^{(k)} = \sum_z p(z)\, (\ket{\psi_z}\bra{\psi_z})^{\otimes k}, \] and exact Haar-like behaviour means $\rho^{(k)}$ matches the Haar moment operator $\rho_H^{(k)}$ [2208.10542]. For mixed states, the conditioning procedure no longer produces pure states in general. One formulation defines the projected ensemble from a mixed state $\varrho$ as \[ {\cal E}_{\rm PE}[\varrho] := \{p(z_B), \varrho_A(z_B)\}_{z_B}, \] with \[ \varrho_A(z_B)=\frac{1}{p(z_B)}\operatorname{Tr}_B[\Pi_{z_B}\varrho],\qquad p(z_B)=\operatorname{Tr}[\Pi_{z_B}\varrho]. \] A related construction, the mixed state projected ensemble (MSPE), appears when measurements on the complement are incomplete, so that some outcomes are lost and the conditional state on $A$ is mixed rather than pure [2507.14135], [2505.07795]. The moment language remains central in the mixed-state setting. In the MSPE approach, the $k$-th moment is \[ \rho_{N_A}^{(k)} = \sum_{\boldsymbol{\alpha}} P_{\boldsymbol{\alpha}}\, \rho_A(\boldsymbol{\alpha})^{\otimes k}, \] while in the mixed-state deep-thermalisation approach the natural reference ensemble is \[ [\rho_0] := {\cal E}_{\rm PE}[U_{\rm Haar}\rho_0 U^\dagger_{\rm Haar}], \] whose moments encode the universal mixed-state distribution associated with the initial state $\rho_0$ [2505.07795], [2507.14135]. ## 2. Failure of the pure-state paradigm and the augmented-system construction A central 2025 result is that the standard pure-state deep-thermalisation framework fails sharply for mixed initial states, even with infinitesimal initial mixedness. For the Haar-scrambled reference ensemble $[\rho_0]$, the second moment is \[ \rho^{(2)}_{[\rho_0]}= \frac{\mathbb{I}_{D_A^2}+(\operatorname{Tr}\rho_0^2)\,\mathbb{S}_{D_A^2}}{D_A^2+(\operatorname{Tr}\rho_0^2)\,D_A}, \] which implies the average purity \[ {\cal P}_{[\rho_0]} =\operatorname{Tr}\!\left[\rho^{(2)}_{[\rho_0]}\mathbb{S}_{D_A^2}\right] =\frac{1+D_A\,\operatorname{Tr}\rho_0^2}{D_A+\operatorname{Tr}\rho_0^2}\le 1. \] Equality holds only if $\rho_0$ is pure. For any finite mixedness, however small, $\operatorname{Tr}\rho_0^2<1$, so the ensemble is not an ensemble of pure states. It therefore cannot be the Haar ensemble of pure states, and it cannot be the usual Scrooge ensemble of pure states either [2507.14135]. The replacement paradigm introduces a mixed-state deep thermal ensemble by purification. One purifies the initial mixed state as \[ \rho_0=\operatorname{Tr}_X\bigl[|\phi_0\rangle\langle\phi_0|\bigr], \] constructs a maximum-entropy projected ensemble of pure states on the augmented system $X\cup A$, and then traces out $X$. The resulting moment formula is \[ \rho_{[\rho_0]}^{(k)} = \int d\psi_{\rm Haar}\, D_{XA}\, \frac{ \left(\operatorname{Tr}_X\left[\sqrt{\rho_{XA}\,|\psi\rangle\langle\psi|\,\rho_{XA}}\right]\right)^{\otimes k} }{ \langle\psi|\rho_{XA}|\psi\rangle^{k-1} }, \qquad \rho_{XA}=\frac{\rho_0^T\otimes \mathbb{I}_{D_A}}{D_A}. \] This is exactly the form of a Scrooge ensemble on $X\cup A$, followed by partial trace over $X$ [2507.14135]. The dependence on the initial state is explicit and spectral. The general $k$-moment is \[ \rho^{(k)}_{[\rho_0]} = \frac{\sum_{\sigma \in \mathcal{S}_k} h_\sigma(\rho_0)\, \mathrm{Perm}_{\mathcal{H}_A^{\otimes k}(\sigma)} }{ \sum_{\sigma \in \mathcal{S}_k} h_\sigma(\rho_0)\,D_A^{|\sigma|} }. \] For $k=2$, the dependence reduces to $\operatorname{Tr}\rho_0^2$, so the distance from the Haar ensemble increases monotonically with the second Rényi entropy \[ S_2(\rho_0)=-\ln \operatorname{Tr}\rho_0^2 \] for $k\ge 2$. In the special case where all nonzero eigenvalues of $\rho_0$ are equal, the ensemble reduces to a generalised Hilbert-Schmidt ensemble, \[ [\rho_0] = \{\operatorname{Tr}_R[|\psi\rangle\langle\psi|]: |\psi\rangle\sim {\rm Haar}(D_{RA})\}, \qquad D_R=\operatorname{rank}(\rho_0). \] This establishes a maximum-entropy principle for mixed states that is fundamentally distinct from the pure-state case [2507.14135]. ## 3. Mixed state projected ensembles and incomplete measurement A second major route to mixed-state deep thermal ensembles begins from imperfect measurements. In the MSPE construction, a many-body system is partitioned into subsystem $A$ and complementary subsystem $B$, with $B$ further split into contiguous regions $B_1,B_2,B_3$. Outcomes on $B_1$ and $B_3$ are retained, while the outcomes on $m$ sites in $B_2$ are lost. If $\boldsymbol{\alpha}=(\boldsymbol{\alpha}_{B_1},\boldsymbol{\alpha}_{B_3})$ denotes the retained outcome string, then \[ \rho_A(\boldsymbol{\alpha}) = \frac{ \mathrm{Tr}_{B_2}\!\left[ \left(I_{A+B_2}\otimes \ket{\boldsymbol{\alpha}}\bra{\boldsymbol{\alpha}}\right) \ket{\Psi}\bra{\Psi} \left(I_{A+B_2}\otimes \ket{\boldsymbol{\alpha}}\bra{\boldsymbol{\alpha}}\right) \right] }{ P_{\boldsymbol{\alpha}} }, \] with \[ P_{\boldsymbol{\alpha}} = \bra{\Psi}\left(I_{A+B_2}\otimes \ket{\boldsymbol{\alpha}}\bra{\boldsymbol{\alpha}}\right)\ket{\Psi}. \] The MSPE is \[ \mathcal{E}=\{P_{\boldsymbol{\alpha}},\rho_A(\boldsymbol{\alpha})\}. \] If no measurement outcomes are lost, $m=0$, then $\rho_A(\boldsymbol{\alpha})$ is pure and the construction reduces to the usual pure-state projected ensemble [2505.07795]. The solvable setting studied for this problem is a $(1+1)$D dual-unitary brick-wall circuit with an initial state given by a tensor product of Bell pairs, \[ \ket{\varphi_0}=\sum_i^d \ket{ii}/\sqrt d, \] and local measurements in the Bell-like basis \[ \ket{\varphi_\alpha} = (I\otimes \sigma_\alpha)\ket{\varphi_0}, \qquad \alpha=1,\dots,d^2. \] Lost outcomes act as a totally depolarizing channel or partial trace, which is the mechanism that converts conditional pure states into mixed states [2505.07795]. The late-time universality statement is explicit: for any fixed ratio $m/N_A$, the MSPE deep thermalizes to a generalized Hilbert-Schmidt ensemble, \[ \mathcal{E} = \left\{ \mathrm{Tr}_m \ket{\psi}_{N_A+m}\bra{\psi} \ \middle|\ \ket{\psi}\sim \mathrm{Haar}(d^{N_A+m}) \right\}. \] When $m=N_A$, this becomes the usual Hilbert-Schmidt ensemble of random density matrices. The derivation uses the permutation-operator basis \[ \rho_{N_A}(g) = \sum_{i_1,\dots,i_k=1}^{d^{N_A}} \ket{i_{g(1)}\cdots i_{g(k)}}\bra{i_1\cdots i_k}, \qquad g\in \mathbb{S}_k, \] with expansion \[ \rho_{N_A}^{(k)}=\sum_g \alpha(g)\rho_{N_A}(g), \] and asymptotic coefficients \[ \alpha(g)=\frac{d^{m\,l(g)}}{d^{mk}}, \] where $l(g)$ is the number of cycles of the permutation $g$ [2505.07795]. The convergence is exponentially fast in the circuit depth. For large $t$ and $k\ll d^m$, \[ \Delta_1^{(k)} \lesssim \frac{1}{d^{t+1}} \frac{1}{d^m\!\left(1-\frac{1}{d^m}\right)} \frac{k(k-1)}{2}, \] and \[ \Delta_2^{(k)} \approx \frac{1}{d^{t+1}} \frac{1}{d^m\!\left(1-\frac{1}{d^m}\right)} \sqrt{\frac{k(k-1)}{2}}. \] Accordingly, the thermalization time for fixed accuracy $\epsilon$ scales as \[ t_k \sim \log(k/\epsilon). \] This establishes a mixed-state version of design-time separation in a fully solvable setting [2505.07795]. ## 4. Design times, GAP/Scrooge ensembles, and solvable chaotic dynamics An exactly solvable precursor to the mixed-state theory studies deep thermalization in a model consisting of a finite subsystem $A$, a small bottleneck subsystem $B_1$, and a large bath $B_2$, with dynamics generated by alternating Haar-random gates on $AB_1$ and $B_1B_2$. After projective measurement of the bath, the projected ensemble on $A$ approaches a universal limit. In the thermodynamic limit $q\to\infty$, the limiting ensemble on $A$ is the Scrooge or Gaussian Adjusted Projected (GAP) ensemble, with induced density matrix \[ \rho_A = d_C^{-1/2} EE^\dagger \] and projected-ensemble moment \[ \rho_A^{(k)} = d_C^{1/2} \int {\rm d}\phi\ \bra{\phi}E^\dagger E \ket{\phi}^{1-k} (E\ketbra{\phi}{\phi} E^\dagger)^{\otimes k}. \] This is the maximally entropic ensemble compatible with $\rho_A$ [2208.10542]. The same work makes the distinction between regular and deep thermalization sharp. Regular thermalization concerns only \[ \rho_A = \mathrm{Tr}_B\, \ketbra{\Psi}{\Psi}, \] while deep thermalization requires convergence of all projected-ensemble moments toward Haar moments. The deviation is measured by \[ \Delta^{(k)} := \frac{\| \rho^{(k)} - \rho_H^{(k)}\|_2}{ \|\rho_H^{(k)} \|_2}, \] and the design time $t_k$ is the earliest time when $\Delta^{(k)}\le \epsilon$. Since $\Delta^{(k+1)}\ge \Delta^{(k)}$, one has \[ t_{k+1}\ge t_k. \] Near infinite temperature, \[ \Delta^{(k)} = f(k,d_A)\,\Delta^{(1)} + O((\Delta^{(1)})^2), \qquad f(k,d_A) = \left( \frac{1+d_A}{1+d_A/k} \right)^{1/2}, \] and the exact result for the design times is \[ t_k = t_1 + \frac{1}{v_E} \log_2 \frac{1+d_A}{1+d_A/k}. \] For $k\ll d_A$, \[ t_k \approx t_1 + \frac{1}{v_E}\log_2 k, \] while \[ t_\infty = t_1 + \frac{\log_2(d_A+1)}{v_E} \simeq \frac{2N_A}{v_E} + \frac{2\log_2(1/\epsilon)}{v_E}, \qquad v_{\text{Haar}} = v_E/2. \] These formulas show that deep thermalization occurs later than ordinary thermalization and that higher moments equilibrate more slowly [2208.10542]. This framework is not itself a mixed-state theory, because its conditional states are pure. However, it supplies the structural template later generalized to mixed-state deep thermalization: projected ensembles, maximum-entropy limiting distributions, and a hierarchy of moment-dependent equilibration times. The mixed-state theories of 2025 can therefore be read as extensions of the GAP/Scrooge paradigm to settings with incomplete measurement or finite initial entropy [2208.10542], [2507.14135]. ## 5. Thermal-state ensembles, random density matrices, and fixed-energy constraints A broader line of work studies thermal or constrained ensembles of density matrices without deriving them from measurement-conditioned subsystem states. For a random Hamiltonian $H$, the thermal state is \[ \rho_H(\beta) := \frac{e^{-\beta H}}{Z_H(\beta)}, \qquad Z_H(\beta)=\operatorname{tr}(e^{-\beta H}), \] and the ensemble of thermal states is $\{\rho_H(\beta)\}_{H\in\mathfrak H_k}$. Randomness is quantified by the mixed-state $t$-design distance \[ T_t^{(k)}(\beta) := \frac12 \left\| \mathbb{E}_{H\in\mathfrak H_k}\big[\rho_H(\beta)^{\otimes t}\big] - \mathbb{E}_{\Psi\in\Upsilon}\big[\Psi^{\otimes t}\big] \right\|_1, \] with \[ \mathbb{E}_{\Psi\in\Upsilon}\big[\Psi^{\otimes t}\big] = \frac{\Pi_{\rm sym}^{(t)}}{d_{\rm sym}^{(t)}}. \] For random global Hamiltonians $H\sim \mathrm{GUE}(D)$, the distance decreases monotonically as $\beta$ increases, \[ \frac{\partial}{\partial \beta}T_t^{(n)}(\beta)\le 0, \] interpolating between the maximally mixed state at $\beta=0$ and the Haar-random ground-state ensemble as $\beta\to\infty$. The ensemble becomes an $\epsilon$-approximate $t$-design at \[ \beta = \Omega(\log t+\log(1/\epsilon)), \] equivalently at temperature \[ T_\epsilon = O\!\left(\frac{1}{\log t+\log(1/\epsilon)}\right). \] For random $k$-local Hamiltonians, by contrast, \[ \mathbb{E}_{H\in\mathfrak H_k}[\rho_H(\beta)] = \frac{I_D}{D} \qquad \text{for all } k \text{ and all } \beta, \] so the ensemble is always an exact state $1$-design, but it does not become a $t$-design for $t>1$. Numerical evidence indicates two temperature regimes separated by a singular point, with values reported roughly around \[ \beta_c^{(2)} \sim 0.8,\qquad \beta_c^{(k)} \sim 0.85,\qquad \beta_c^{(k)} \sim 1.05 \] for different interaction ranges [1407.6136]. A different generalization treats density matrices themselves as microstates. In that approach, the space of states is $\mathcal S(\mathcal H)$, equipped with a measure $dV(\rho)=d\mu_\rho\,d\mu_U$, and one imposes the fixed-average-energy constraint \[ d\mu_E(\rho) = \omega(E)^{-1} \delta\!\left[E-\operatorname{Tr}(H\rho)\right]\,dV(\rho), \] with density of states \[ \omega(E) = \int_{\mathcal S(\mathcal H)} dV(\rho)\, \delta\!\left[E-\operatorname{Tr}(H\rho)\right]. \] The ensemble average of an observable is \[ \langle \mathcal O\rangle_E = \int d\mu_E(\rho)\,\operatorname{Tr}(\mathcal O\rho), \] and the average density matrix \[ \bar\rho(E) = \int_{\mathcal S(\mathcal H)} d\mu_E(\rho)\,\rho \] obeys \[ [\bar\rho(E),H]=0. \] The associated thermodynamic quantities are \[ S(E)=\ln \Omega(E), \qquad \Omega(E)=\int_{E_0}^{E} dE'\,\omega(E'), \] and \[ \frac{1}{T(E)}=\frac{\partial S}{\partial E} = \frac{\omega(E)}{\Omega(E)} \ge 0. \] A useful response identity is \[ \langle \mathcal O\rangle_E = -T(E)\left.\frac{\partial S_\lambda(E)}{\partial \lambda}\right|_{\lambda=0}, \] for the perturbation $H\mapsto H+\lambda\mathcal O$ [2508.00809]. For a qubit with \[ H=-\epsilon \sigma_z, \] the average density matrix is \[ \bar\rho(E) = \frac{1}{\epsilon} \begin{pmatrix} E & 0 \\ 0 & \epsilon-E \end{pmatrix}, \] which can also be written in canonical form only in this special two-level case. For $N$ non-interacting spins in a magnetic field, \[ H=-\frac{B}{2}\sum_{i=1}^N \sigma_z^{(i)}, \] the Bures-Hall ensemble exhibits a phase-transition-like phenomenon without interactions, and the relative energy fluctuations do not vanish generically in the thermodynamic limit. This suggests that mixed-state ensemble geometry alone can generate nonstandard thermodynamic behaviour [2508.00809]. ## 6. Operational significance: teleportation, imperfect measurements, and multiscale Gibbs ansätze Mixed-state deep thermal ensembles are operationally relevant because they arise naturally when measurements are lossy or when the initial state has finite entropy. In the MSPE setting, the states composing the ensemble can mediate teleportation of quantum information from the right end of the system to the left subsystem $A$. The relevant diagnostic is the annealed Rényi-$k$ conditional entropy \[ I_{R:A}^{(k)} = S_{AR}^{(k)}-S_A^{(k)} = -\log \frac{\overline{\mathrm{Tr}\,\rho_{AR}^k}}{\overline{\mathrm{Tr}\,\rho_A^k}}. \] Its values distinguish three regimes: \[ I_{R:A}^{(k)}=-\log d \quad \text{(perfect teleportation)}, \] \[ I_{R:A}^{(k)}=+\log d \quad \text{($A$ and $R$ decoupled)}, \] and \[ I_{R:A}^{(k)}=0 \quad \text{(transition point)}. \] The sharp transition occurs at \[ N_A=m. \] For $N_A>m$, one has the teleportation phase; for $N_A 0.4$ for $512$-site systems across all temperatures; even $D=3$ gives fidelity above $0.2$ across all temperatures. The per-site infidelity behaves approximately as \[ 1-\mathcal F^{1/L} = s(T),\qquad s(T)<10^{-2}. \] At low temperatures, the energy above the ground state scales roughly as $T^2$, and the entropy per qubit scales roughly as $T$. The largest reported entropy mismatch for $D=6$ is about $0.014$ bits per qubit near $T=0.28$, and the largest energy mismatch is about $0.011$ per qubit near $T=1.3$ [2210.16419]. Taken together, these results show that mixed-state deep thermal ensembles are not merely abstract limiting distributions. They encode experimentally relevant imperfect-measurement statistics, furnish solvable universality classes for chaotic dynamics, connect to random density-matrix measures such as the generalized Hilbert-Schmidt, Hilbert-Schmidt, Bures-Hall, and GAP/Scrooge ensembles, and provide constructive ansätze for thermal state preparation on quantum devices [2505.07795], [2507.14135], [2210.16419]. Source: https://www.emergentmind.com/topics/mixed-state-deep-thermal-ensemble