Unitary Designs from Two Chaotic Hamiltonians and a Random Pauli Operation Ning Sun, Pengfei Zhang Abstract The realization of unitary designs is of fundamental interest in quantum science and typically requires the ability to implement structured quantum circuits. Recent developments have explored the possibility of generating unitary designs using only a small number of quantum quenches, in which the evolution during each interval is governed by a static Hamiltonian. In particular, it has been established that at least three chaotic Hamiltonians are required when only Hamiltonian evolutions are employed. In this work, we propose the emergence of unitary designs in the temporal ensemble of qubit systems evolved under two distinct chaotic Hamiltonians for sufficiently long times, supplemented by an intermediate random Pauli operation inserted between them. This result follows from the universal Pauli spectrum of chaotic Hamiltonians, a central concept in the study of non-stabilizerness. Our theoretical predictions are verified numerically using explicit examples, including Gaussian unitary ensemble Hamiltonians and random spin models. We further investigate finite-time and finite-size corrections to the protocol. Our results provide new insights into the dynamical generation of quantum randomness and offer a new route toward realizing unitary designs in chaotic systems. Introduction.— Understanding how the complexity of evolution operators grows in chaotic many-body systems is a central topic in modern quantum many-body physics, as it is intimately linked to information scrambling Sekino and Susskind 2008 ; Hayden and Preskill 2007 ; Kitaev 2014 ; Shenker and Stanford 2014 ; Shenker and Stanford 2015 ; Maldacena et al. 2016 ; Roberts et al. 2015 and quantum thermalization Srednicki 1999 ; Deutsch 1991 ; Kaufman et al. 2016 . A key concept is unitary k -designs DiVincenzo et al. 2002 ; Dankert et al. 2009 ; Ambainis and Emerson 2007 ; Emerson et al. 2005 ; Scott 2008 ; Roberts and Yoshida 2017 ; Cotler et al. 2017 , which refer to ensembles of unitary operators whose first k moments coincide with those of the Haar-random unitary ensemble. For example, random Pauli operators form a unitary 1-design Roberts and Yoshida 2017 , leading to thermalization to an infinite-temperature ensemble. On the other hand, random Clifford operations in qubit systems form a unitary 3-design Webb 2016 ; Zhu 2017 and result in the saturation of out-of-time-order correlators (OTOCs) Larkin and Ovchinnikov 1969 . The recent study on higher-order OTOCs Abanin et al. 2025 ; Vardhan and Wang 2025 ; Guo 2025 ; Liu and Zhang 2026 is also connected to higher-order designs. In addition, unitary designs have broad applications in quantum computation: they can be used for randomized measurements Qi et al. 2019 ; Huang et al. 2020 ; Elben et al. 2023 and serve as a resource for demonstrating quantum advantage Arute et al. 2019 ; Zhong et al. 2020 ; Madsen et al. 2022 . Figure 1: The schematic illustration of our main results. We consider the temporal ensemble generated by evolution under two fixed chaotic Hamiltonians H_{1} and H_{2} , with a random Pauli operation inserted between them. We show that this protocol enables the emergence of unitary k -designs at sufficiently long times in the thermodynamic limit. As a consequence, generating unitary designs in realistic setups has become an important question. In particular, it has been established that the ability to implement structured quantum circuits enables the generation of unitary designs Emerson et al. 2003 ; Dankert 2005 ; Gross et al. 2007 ; Ambainis and Emerson 2007 ; Roy and Scott 2009 ; Brandão et al. 2016 ; Brandão et al. 2016 ; Brandão et al. 2021 ; Schuster et al. 2025 ; Hou et al. 2025 ; Zhang et al. 2025 , with an optimized depth that scales logarithmically with system size Schuster et al. 2025 . This makes the circuit-based approach well suited to quantum computing platforms. In contrast, quantum simulation platforms typically lack the ability to implement complex quantum circuits. Therefore, understanding how to generate unitary designs through Hamiltonian evolution has likewise attracted significant attention Vermersch et al. 2018 ; Brandão et al. 2021 ; Onorati et al. 2017 ; Jian et al. 2023 ; Tiutiakina et al. 2024 ; Guo et al. 2024 ; Zhou et al. 2025 . Recently, Zhou et al. have shown that unitary designs can be realized either through a single quantum quench separating evolutions governed by two random Hamiltonians Zhou et al. 2025 , or within the temporal ensemble generated by evolution under three fixed chaotic Hamiltonians Zhou et al. 2026 . To our knowledge, these constitute the simplest schemes for realizing unitary designs when only Hamiltonian evolutions are employed. In this work, we take one step back: in addition to Hamiltonian evolution, we allow random Pauli operations, which correspond to applying a random Pauli operator to each qubit. Such operations can be implemented by applying control fields to specific subsystems and are therefore feasible on quantum simulation platforms. With this new ingredient, we show that two chaotic Hamiltonians are sufficient to generate unitary designs in the temporal ensemble, as illustrated in Fig. 1 . Interestingly, this mechanism is closely related to the universal property of the Pauli spectrum in generic chaotic systems Beverland et al. 2020 ; Turkeshi et al. 2025 , which has recently been proposed in the context of quantum magic Bravyi and Kitaev 2005 ; Bravyi and Haah 2012 ; Veitch et al. 2012 ; Emerson et al. 2014 . We explicitly verify our protocol using concrete Hamiltonians sampled from the Gaussian unitary ensemble (GUE) and from random spin models. Our results provide a new route toward generating unitary designs in many-body systems and shed new light on the interplay between quantum many-body chaos and quantum magic. Setup.— We now describe our scheme, illustrated in Fig. 1 . We focus on systems consisting of N qubits. The system is first evolved under a chaotic Hamiltonian H_{1} for a time t_{1} , where t_{1} is sampled uniformly from the interval [0,T] . We then apply a random Pauli operation with unitary operator P=P_{1}P_{2}\cdots P_{N} , where P_{i}\in\{I,X,Y,Z\} denotes a single-qubit Pauli operator acting on site i . For now, we assume that all Pauli operators are chosen with equal probability. The consequences of modifying the Pauli ensemble will be discussed later. Finally, we apply a second Hamiltonian evolution governed by a distinct Hamiltonian H_{2} for a time t_{2} , where t_{2} is sampled from the same distribution as t_{1} . The resulting ensemble of unitary operators is given by \mathcal{E}=\big\{e^{-iH_{2}t_{2}}Pe^{-iH_{1}t_{1}},\big|t_{1},t_{2}\in[0,T],P\big\}. (1) We aim to study whether the ensemble \mathcal{E} forms a unitary k -design in the thermodynamic limit N\rightarrow\infty . This is typically quantified by the k -th frame potential F^{(k)}_{\mathcal{E}} Scott 2008 ; Roberts and Yoshida 2017 ; Cotler et al. 2017 , defined as F^{(k)}_{\mathcal{E}}\equiv\mathbb{E}_{U,V\in\mathcal{E}}~\left|\text{tr}[V^{\dagger}U]\right|^{2k},\ \ \ k\in\mathbb{N^{+}}. (2) It has been shown that the Haar-random unitary ensemble achieves the minimal k -th frame potential among all ensembles of unitary operators, with F^{(k)}_{\text{Haar}}=k! Scott 2008 . Therefore, an ensemble \mathcal{E} forms a unitary k -design if and only if F^{(k)}_{\mathcal{E}}=k! . The frame potential can also be related to the Rényi entropy of EPR states when one member of each pair is subjected to noisy unitary evolution drawn from the ensemble \mathcal{E} Wang et al. 2024 . General arguement.— We proceed to calculate the frame potential for our protocol \mathcal{E} , defined in Eq. ( 1 ). Random Pauli operations with a uniform distribution already form an exact 1-design, and this property remains valid after composition with Hamiltonian evolutions Roberts and Yoshida 2017 . Therefore, it suffices to analyze the case k\geq 2 . Now, we focus on k=2 as a concrete example. We represent the eigenstates and eigenvalues of the two Hamiltonians as H_{1}|E_{i}\rangle=E_{i}|E_{i}\rangle and H_{2}|\epsilon_{a}\rangle=\epsilon_{a}|\epsilon_{a}\rangle , where i,a\in\{1,2,\dots,D\} with the Hilbert space dimension D=2^{N} . This allows us to expand \text{tr}[V^{\dagger}U]=\sum_{i,a}e^{-i\delta t_{1}E_{i}-i\delta{t_{2}}\epsilon_{a}}\langle\epsilon_{a}|P|E_{i}\rangle\langle E_{i}|\tilde{P}|\epsilon_{a}\rangle. (3) Here, P and \tilde{P} denote the Pauli operators associated with U and V , respectively. We have also introduced \delta t_{r}=t_{r}-t_{r}^{\prime} with r\in\{1,2\} , representing the differences between the random evolution times t_{r} and t_{r}^{\prime} for U and V . To calculate F_{\mathcal{E}}^{(2)} , we introduce four replicas of Eq. ( 3 ) labeled by n\in\{1,2,3,4\} . The result reads \displaystyle F_{\mathcal{E}}^{(2)}=\sum_{i_{n},a_{n}}I(E_{i_{n}}, \displaystyle\epsilon_{a_{n}})~P_{a_{1}i_{1}}P_{a_{2}i_{2}}^{*}P_{a_{3}i_{3}}P_{a_{4}i_{4}}^{*} (4) \displaystyle\times \displaystyle\tilde{P}_{i_{1}a_{1}}\tilde{P}_{i_{2}a_{2}}^{*}\tilde{P}_{i_{3}a_{3}}\tilde{P}_{i_{4}a_{4}}^{*}. Here, we keep the average over P and P^{\prime} implicit and introduce P_{ai}=\langle\epsilon_{a}|P|E_{i}\rangle and \tilde{P}_{ia}=\langle E_{i}|\tilde{P}|\epsilon_{a}\rangle for brevity. The time-filter function I(E_{i_{n}},\epsilon_{a_{n}}) reads I(E_{i_{n}},\epsilon_{a_{n}})\equiv\mathbb{E}_{\delta t_{r}}e^{i\sum_{n=1}^{4}(-1)^{n}(E_{i_{n}}\delta t_{1}+\epsilon_{a_{n}}\delta t_{2})}, (5) where (-1)^{n} arises from the complex conjugation in the replicas involving U^{\dagger}V . When the total evolution time T\rightarrow\infty , time averaging over \delta t_{1} imposes the constraint E_{i_{1}}+E_{i_{3}}=E_{i_{2}}+E_{i_{4}} . In the absence of two-level resonances in chaotic systems, the only solutions are (i_{1}=i_{2},i_{3}=i_{4}) or (i_{1}=i_{4},i_{3}=i_{2}) . Similar relations holds for a_{n} . This is referred to as the perfect time-filter limit in Zhou et al. 2026 . We find \displaystyle\frac{F_{\mathcal{E}}^{(2)}}{2}= \displaystyle\sum_{i_{n},a_{n}}P_{a_{1}i_{1}}P_{a_{1}i_{1}}^{*}P_{a_{3}i_{3}}P_{a_{3}i_{3}}^{*}\tilde{P}_{i_{1}a_{1}}\tilde{P}_{i_{1}a_{1}}^{*}\tilde{P}_{i_{3}a_{3}}\tilde{P}_{i_{3}a_{3}}^{*} (6) \displaystyle+P_{a_{1}i_{1}}P_{a_{1}i_{3}}^{*}P_{a_{3}i_{3}}P_{a_{3}i_{1}}^{*}\tilde{P}_{i_{1}a_{1}}\tilde{P}_{i_{1}a_{3}}^{*}\tilde{P}_{i_{3}a_{3}}\tilde{P}_{i_{3}a_{1}}^{*}. The first and second lines represent the contributions in which the solutions for the i and a either share the same pairing pattern or follow different pairing patterns. The evaluation of Eq. ( 6 ) therefore requires analyzing the matrix elements P_{ai}=\langle\epsilon_{a}|P|E_{i}\rangle for typical Pauli operators P . To gain some intuition, we consider the scenario in which both H_{1} and H_{2} are sampled from the GUE Fyodorov 2004 . In this case, the corresponding eigenstates are Haar-random states, which can be written as |E_{i}\rangle=W_{1}|i\rangle and |\epsilon_{a}\rangle=W_{2}|a\rangle , where W_{1} and W_{2} are typical Haar-random unitaries. As a consequence, P_{ai}=\langle a|W_{2}^{\dagger}PW_{1}|i\rangle is also Haar random. We further express \tilde{P}_{ia}=\langle i|W_{1}^{\dagger}\tilde{P}W_{2}|a\rangle=\sum_{c}P_{ci}^{*}\langle c|W_{2}^{\dagger}P_{r}W_{2}|a\rangle, (7) with P_{r}\equiv P\tilde{P} . Next, we neglect sample fluctuations and perform the Haar average over P_{ai} . Standard Weingarten calculus Roberts and Yoshida 2017 in the limit D\rightarrow\infty implies that matrix elements P_{ai} with different indices behave as independent Gaussian variables with variance 1/D . We assume this remains valid for general chaotic Hamiltonians. To extract the dominant contribution to Eq. ( 6 ), we should avoid introducing additional constraints on the indices i_{n} and a_{n} , as such constraints would render the contribution subleading in 1/D . We begin by considering the special case P=\tilde{P} , which occurs with probability p_{0}=4^{-N} for Pauli operations drawn from a uniform distribution. In this case, the Haar average takes the same form as in the two-step protocol without Pauli operations, which has already been analyzed in Zhou et al. 2026 . In both lines of Eq. ( 6 ), Wick’s theorem allows contractions between P_{a_{n}i_{n}} and \tilde{P}_{i_{n}a_{n}}=P_{a_{n}i_{n}}^{*} . In addition, the first line permits an extra contraction pattern between P_{a_{n}i_{n}} and P_{a_{n}i_{n}}^{*} . Putting all these contributions together, we find that the resulting contribution to F^{(2)}_{\mathcal{E}} is 6p_{0} , consistent with Ref. Zhou et al. 2026 . Next, we consider the contribution from P\neq P^{\prime} . The key difference is that, if we attempt to contract P_{a_{n}i_{n}} with \tilde{P}_{i_{n}a_{n}} , Eq. ( 7 ) predicts \overline{P_{a_{n}i_{n}}\tilde{P}_{i_{n}a_{n}}}=\langle\epsilon_{a}|P_{r}|\epsilon_{a}\rangle/D. (8) The collection of c(P_{r})=\langle\epsilon_{a}|P_{r}|\epsilon_{a}\rangle is known as the Pauli spectrum Beverland et al. 2020 ; Turkeshi et al. 2025 , which provide detailed information of the many-body state |\epsilon_{a}\rangle . It has been established that for chaotic systems c(P_{r}) satisfies a real Gaussian distribution with a zero mean and a variance of 1/(D+1) Turkeshi et al. 2025 . This leads to an additional suppression, causing the corresponding contraction to vanish in the thermodynamic limit. Therefore, the only remaining contribution comes from the contraction between P_{a_{n}i_{n}} and P_{a_{n}i_{n}}^{*} in the first line. Putting all ingredients together, we find F^{(2)}_{\mathcal{E}}=6p_{0}+2(1-p_{0})~\xrightarrow{N\rightarrow\infty}~2. (9) Finally, we briefly describe the generalization of the above analysis to arbitrary k . Similar to Eq. ( 6 ), in the perfect-filter limit, the calculation of F^{(k)} requires analyzing the 2k -th moments of P_{a_{n}i_{n}} and \tilde{P}_{i_{n}a_{n}} , where the indices i_{n} and a_{n} can follow different pairing patterns. However, for P\neq\tilde{P} , only the terms with the same pairing pattern contribute in the thermodynamic limit. This reduces the double sum over the permutation group to a single sum, resulting in a contribution of (1-p_{0})k! . Adding the contribution from the special case with P=\tilde{P} , denoted as F^{(k)}_{\text{2SP}} Zhou et al. 2026 , we obtain F^{(k)}_{\mathcal{E}}=p_{0}F^{(k)}_{\text{2SP}}+(1-p_{0})k!~\xrightarrow{N\rightarrow\infty}~k!. (10) Here, we have used the fact that F^{(k)}_{\text{2SP}} is independent of the system size N . This demonstrates the emergence of the unitary k -design in our protocol for sufficiently long evolution times in the thermodynamic limit. We further note that, for a Hamiltonian ensemble consisting of a single evolution followed by random Pauli operations, the contribution from P=\tilde{P} already leads to F_{\mathcal{E}}^{(k)}\geq p_{0}k!D^{k} , which diverges for k>2 Roberts and Yoshida 2017 . Consequently, our protocol represents the simplest scheme that combines Hamiltonian evolutions with random Pauli operations to generate unitary designs. Figure 2: The numerical results of our protocol. The Hamiltonians are drawn from either the GUE (a,c) or random spin models (b,d), for a system size of N=7 . The results are obtained by generating 400 random evolution operators and computing their overlaps according to Eq. ( 2 ). In panels (a-b), we set T=10^{6} to ensure sufficiently long evolution times, and we compare protocols with and without the Pauli operations. The dashed lines correspond to either F_{\text{2SP}}^{(k)} or k! . The error bars, which reflect statistical fluctuations in the evaluation of the expectation, are smaller than the marker size for most data points. In panels (c-d), the dashed lines indicate the error threshold \epsilon=10^{-1} . Numerical demonstration.— We now provide numerical demonstrations of our theoretical predictions. We consider two different models for the chaotic Hamiltonians: the GUE and random spin models. For the GUE Hamiltonians, the matrix elements H_{ij} are sampled as independent Gaussian variables with variance 1/2D , subject to the constraint of Hermiticity. The resulting many-body spectrum follows a semicircular distribution with a radius of \sqrt{2} Fyodorov 2004 . For random spin models, the Hamiltonian reads H=\sum_{i