betaFeedback

WorksPhysics and Astronomy

Qubit-qumode simulation of molecular torsion in photoisomerization

Abstract

Photochemical reactions driven by large nuclear motion, such as the photoisomerization of retinal, depend strongly on both the global anharmonic forces of the potential and the periodic topology of molecular torsion. In this work, we develop a hybrid discrete (qubit) and continuous variable (qumode) quantum simulation framework for photochemical reactions that maps electronic states to qubits, vibrations to qumodes, and the periodic torsion to a compact qumode encoding. Applying this approach to the two-mode Hahn-Stock and three-mode PSB3 models of retinal photoisomerization, we compare i) local vibronic approximations, ii) the full trigonometric potential on the noncompact line, and iii) the same potential on the compact circle, which matches the molecular torsion topology. We numerically show that the topological description can significantly alter relevant observables. We realize the compact torsion via both a wrapped Gottesman-Kitaev-Preskill (GKP) encoding and a two-qumode circle encoding and introduce two methods that reduce the hardware resources of both encodings. Spreading an encoded coordinate over several qumodes lowers the average occupation number of each mode, and adding counterterms to the Hamiltonian reduces the required squeezing amount. We expect that these methods make percent-level simulations of the considered molecules promising targets for hybrid qubit-qumode hardware. For more complex molecules, each additional nuclear coordinate is represented directly as a single qumode or an encoded register using the same native operations, so the framework extends naturally to larger systems.

Introduction

In photochemistry, light absorption prepares an excited molecule that can explore configurations far from equilibrium before forming a product or returning to its initial structure. Predicting this evolution connects molecular interactions to reaction pathways, branching fractions and the time scales of chemical change. In photoisomerization, the electronic excitation drives a large-amplitude rearrangement of molecular geometry, with applications ranging from reversible molecular photoswitches to the primary photochemical event in vision [1, 2, 3].

Near conical intersections, where electronic potential-energy surfaces become degenerate, nonadiabatic couplings between electrons and nuclei govern the reaction. A nuclear wavepacket can transfer between electronic states, split into different branches and accumulate phases that affect its subsequent evolution [1, 4, 5, 6]. Retinal photoisomerization, the rotation from the cis to the trans geometry about a carbon-carbon double bond, illustrates this connection between electronic transitions and structural change, with femtosecond spectroscopy establishing ultrafast photoproduct formation [7] and passage through a conical intersection [2].

Among classical methods, trajectory approaches combine electronic evolution with approximate nuclear motion [8, 9], while wavepacket methods, such as the multiconfiguration time-dependent Hartree (MCTDH) algorithm [10], retain the correlated quantum dynamics of selected nuclear coordinates. Multilayer MCTDH [11] and tensor-network methods [12, 13, 14] extend wavepacket propagation to tens and, in favorable cases, hundreds of modes. However, as the number of active coordinates grows, resolving their correlations and maintaining an accurate global description of the potential becomes increasingly demanding. Reduced vibronic models provide a controlled setting in which to study these questions and to compare simulation methods [15, 16]. Reduced models of retinal range from the two-mode Hahn-Stock model [17, 18] to the three-mode PSB3 model [19, 20], which retain the torsional coordinate and conical intersections that drive photoisomerization. For a large-amplitude reaction, even a small model must retain the features of the coordinate that govern its global motion. A molecular torsion is a continuous compact coordinate on the circle: a full rotation returns to the same molecular geometry, while a local displacement or an oscillator quadrature is continuous and noncompact, so distinct points remain distinct however far apart they lie. As we show in this work, this topological distinction alters the reaction dynamics even when the local forces are reproduced exactly.

Quantum simulation offers an alternative route to retaining the electronic and nuclear degrees of freedom and their correlations [21, 22, 23]. Digital qubit algorithms discretize nuclear coordinates on grids or in truncated vibrational bases, with the Hamiltonian structure determining the circuit cost [24, 25, 26, 27, 28, 29]. Recent block encodings of vibrational Hamiltonians for fault-tolerant energy estimation expand the potential about the equilibrium geometry and leave large-amplitude torsion outside their scope [30]. Hybrid qubit-qumode architectures offer a more direct physical representation: qubits encode electronic states and qumodes, quantum harmonic oscillators with continuous canonical quadratures, encode nuclear motion [31, 32, 33]. This representation makes nuclear displacements and harmonic evolution native operations. Bosonic devices have been used for vibronic spectroscopy through boson sampling [34], quantum optical emulation of molecular spectra with trapped ions [35], photonic simulation of vibrational dynamics [36], anharmonic vibrational dynamics on bosonic devices [37] and time-domain simulation of vibronic spectra with a single trapped ion [38].

Coupling to a two-level system enables the preparation of arbitrary field states [39] and universal control of a cavity qumode [40, 41]. Neutral atoms in optical tweezers couple electronic and motional degrees of freedom [42, 43]. Trapped ions combine coherent control of internal states and collective motional qumodes [44, 45] and have simulated conical intersections, their geometric-phase interference and the chemical dynamics of vibronic Hamiltonians [46, 47, 48]. Gaussian operations generate Hamiltonians at most quadratic in the quadratures, so universal continuous-variable dynamics requires a non-Gaussian resource [32, 49, 50]. Qubit-assisted signal processing provides constructions for anharmonic vibronic potentials [51]. Trigonometric gate constructions provide direct access to periodic potentials and multimode couplings through programmable Fourier synthesis [52, 53] and have been realized in trapped ions together with programmable anharmonic dynamics [54, 55]. Their application to photochemistry raises the central question of how to represent the compact torsional coordinate on which those periodic interactions act.

We give an overview of the problem and our approach in Figure 1 using the Hahn-Stock model of photoisomerization, connecting the molecular potential-energy surfaces and conical intersections to a hybrid qubit-qumode representation and to reaction observables that separate local, noncompact and compact descriptions. Specifically, we distinguish three descriptions throughout this work:

Figure 1. Quantum simulation of torsional photochemistry. (a) The Hahn-Stock adiabatic surfaces S0S_0 and S1S_1 over one full rotation, indicating the cis and trans geometries and the two conical intersections. (b) Electronic and nuclear degrees of freedom are represented by qubits and qumodes. The illustration shows trapped ion collective motional modes as qumodes. The compact torsion can be encoded through a finite-energy Gottesman-Kitaev-Preskill (GKP) construction or a circle constraint. (c) The upper panel compares electronic populations for LVC and the full trigonometric potential on the continuous noncompact line and the physical compact circle. The lower panels show the total torsional probability density ρ(θ,t)\rho(\theta,t), summed over electronic states and integrated over the vibration, for LVC, noncompact cosine and compact cosine, from left to right. The color scale is logarithmic. We show the noncompact density only in the central cell [−π,π][-\pi,\pi], so the probability that has moved beyond ±π\pm\pi does not contribute.

  1. Local models. Linear and quadratic vibronic coupling models, denoted LVC and QVC, expand the potential about a reference geometry and treat the torsional displacement as a coordinate on a line.

  1. Trigonometric line. The full trigonometric potential on a noncompact coordinate restores the global forces but treat θ\theta and θ+2π\theta+ 2\pi as distinct configurations.

  1. Trigonometric circle. The same potential is placed on the circle, the physical configuration space of the torsion, where θ\theta and θ+2π\theta+ 2\pi coincide.

Comparing (i) with (ii) tests the approximation to the potential and comparing (ii) with (iii) tests the topology of the configuration space. The true model requires (iii) to correctly describe the torsion, but (ii) is the cheaper option since qumodes are naturally noncompact. It can be accurate until amplitudes that explore different angular cells produce an observable difference.

On the circle, amplitudes that reach the same molecular geometry after different winding histories can interfere. On the line, those amplitudes remain in separate translated cells, so electronic populations and product probabilities potentially differ despite identical periodic forces. We illustrate the two physical effects in Figure 1(c): local confinement can suppress torsional motion, while the full-potential line and circle track each other initially before separating at later times as winding interference alters electronic populations and product probabilities.

Continuous compact variables also occur in lattice gauge theory, where an angular field and its conjugate electric field obey the same rotor algebra as molecular torsion. Hybrid qubit-qumode proposals [45, 50, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66] and experiments [67] have motivated two bosonic encodings that distinguish this algebra from that of a noncompact qumode. A Gottesman-Kitaev-Preskill (GKP) translation stabilizer [68] selects a compact rotor sector within the Bloch decomposition of a qumode [64]. A second construction constrains two qumode coordinates to a circle and uses their angular motion for the compact variable [62]. In the molecular setting of this work, these two constructions must preserve nonadiabatic transfer and reaction observables while coupling the torsion to additional nuclear degrees of freedom. We develop and test both realizations under those requirements.

Our primary application is the two-mode Hahn-Stock model of retinal photoisomerization [17, 18], which has the form of the photoisomerization models introduced by Seidner and Domcke [69]. They solved these models in a direct-product basis of diabatic electronic states, free-rotor states and harmonic-oscillator states, the classical counterpart of the qubit, rotor and qumode registers used here. Its two electronic states couple to the torsional reaction coordinate and one vibration, giving conical intersections and transfer between cis (θ=0\theta= 0) and trans (θ=π\theta= \pi) regions in a model that admits accurate reference propagation. Extensions with further vibrational modes have also served as benchmarks for quantum and quantum-classical dynamics [70]. Versions of the model with additional bath modes have served as benchmarks for tensor-network and open-system methods [71, 72]. We compare the local vibronic models, the full trigonometric potential on the line, and the physical compact model on the circle through their effects on population transfer and on the time-dependent branching between cis and trans basins. Additionally, we study the three-mode PSB3 model [19, 20], which provides a more detailed chromophore application, with additional mode coupling, coherent motion and nuclear recurrences [73].

We map each degree of freedom to the resource that describes it most directly. The electronic states naturally map onto a qubit and the coupling vibration, a noncompact continuous variable, maps directly onto a qumode. The compact torsion requires more care because it is not native. For that we adopt two bosonic encodings: a wrapped GKP-style encoding on a qumode or a circle embedded in two qumodes. In practice, the states corresponding to these two encodings are replaced by finite-energy GKP states and finite-width circle states. We quantify the Fock-space occupation each encoding requires and show how to exploit two freedoms that make the encodings cheaper without changing the simulated dynamics. Passive mode mixing [32, 74] distributes an encoding over several physical qumodes, conserving the total excitation number while lowering the occupation of each mode. A similar uniform redistribution fixes the local cutoff in qubit simulations of boson sampling [75]. Second, counterterms added to the Hamiltonian, leave the encoded dynamics unchanged and reduce the sensitivity of the dynamics to the finite width of the encoding. This relaxes the required squeezing and lowers the total occupation that redistribution then shares among the modes.

We also build the time step from elementary gates and compare it with a qubit implementation of the same Hamiltonian, in the broader context of bosonic encodings, planar rotors and vibronic simulation [28, 76, 77]. Finally, we estimate the resources these simulations need under stated assumptions. The representation also extends to larger molecules. Each additional vibration adds one qumode and each additional torsion one encoded register, with the same native operations. While classical wavepacket methods, such as MCTDH, can treat vibronic models with about two dozen modes, such as the 24-mode model of pyrazine [10, 78], their cost grows exponentially with the number of strongly correlated coordinates [11], and for tensor networks with the entanglement that builds up during the dynamics [13]. Resource estimates for vibronic dynamics on quantum hardware [79, 80] indicate where quantum simulations can extend beyond this regime.

The rest of the paper is organized as follows. In Sec. II we introduce the Hahn-Stock Hamiltonian and relevant observables. In Sec. III we develop the compact encodings, their extension over several qumodes and the Hamiltonian counterterms. In Sec. IV we present the Hahn-Stock dynamics and finite-resource benchmarks and in § 2.4 we develop the PSB3 application. In Sec. VI we compare qubit and qumode requirements and discuss the experimental tradeoffs of the two encodings.

Molecular torsion and the Hahn-Stock model

A qubit coupled to a rotor and a vibration

The minimal Hahn-Stock model of photoisomerization couples two electronic states (a qubit) to two active nuclear coordinates: a torsion θ\theta (a planar rotor) and a coupling vibration xx (a harmonic oscillator) [17, 18]. Its reduced configuration space retains the two ingredients that are central to the present work. Rotations along θ\theta change the electronic energy landscape globally, while the vibration xx permits nonadiabatic population transfer.

We work in the diabatic electronic basis {∣0⟩, ∣1⟩}\{\lvert0\rangle,\ \lvert1\rangle\}, whose states are independent of the nuclear coordinates and are chosen such that the potential matrix is a smooth function of them. In this basis, the nuclear kinetic terms are diagonal and the electronic coupling is carried entirely by the potential. The physical torsion θ\theta is a compact coordinate on the circle S1\mathbb{S}^{1} and the vibration xx is described by a dimensionless quadrature on R\mathbb{R} with [x^,p^x]=i[\hat{x},\hat{p}_{x}]=i. Therefore, the molecular Hilbert space is

HHS=C2⊗L2(S1)⊗L2(R).(1)\mathcal{H}_{\mathrm{HS}}=\mathbb{C}^{2}\otimes L^{2}(\mathbb{S}^{1})\otimes L^{2}(\mathbb{R}). \tag*{(1)}

For simplicity, we use ℏ=1\hbar=1 in operator expressions and quote molecular energies in eV and times in fs, unless otherwise stated. The molecular Hamiltonian is

H^HS=G2L^2+ω2(p^x2+x^2)+V^(θ^,x^),(2)\hat{H}_{\mathrm{HS}}=\frac{G}{2}\hat{L}^{2}+\frac{\omega}{2}\left(\hat{p}_{x}^{2}+\hat{x}^{2}\right)+\hat{V}(\hat{\theta},\hat{x}), \tag*{(2)}
V(θ,x)=(W02(1−cos⁡θ)λxλxE1−W12(1−cos⁡θ)+κx).(3)V(\theta,x)=\begin{pmatrix} \frac{W_{0}}{2}(1-\cos\theta) & \lambda x \\ \lambda x & E_{1}-\frac{W_{1}}{2}(1-\cos\theta)+\kappa x \end{pmatrix}. \tag*{(3)}

Here V(θ,x)V(\theta,x) is the coordinate-space matrix of the potential operator V^\hat{V} in the {∣0⟩, ∣1⟩}\{\lvert0\rangle,\ \lvert1\rangle\} basis while the kinetic terms act as the identity in the electronic space. The parameters of this model are (E1,W0,W1,ω,κ,λ)=(2.48,3.60,1.09,0.19,0.10,0.19) eV(E_{1},W_{0},W_{1},\omega,\kappa,\lambda)=(2.48,3.60,1.09,0.19,0.10,0.19)\ \mathrm{eV} and the inverse torsional moment of inertia G=4.84×10−4 eVG=4.84\times10^{-4}\ \mathrm{eV}. This two-coordinate parametrization is also implemented in QuantumModelLib [81]. Because the cosine terms carry opposite signs in the two diabats, the cis geometry (θ=0\theta=0) is a minimum of the ground diabat ∣0⟩\lvert0\rangle and a maximum of the excited diabat ∣1⟩\lvert1\rangle. The diagonal coupling κx\kappa x displaces the excited potential along the vibration, while λx\lambda x drives electronic transitions.

On S1\mathbb{S}^{1}, periodic boundary conditions restrict the angular momentum operator L^=−i∂θ\hat{L}=-i\partial_{\theta} to integer eigenvalues. Together with the unitary shift operator U^=eiθ^\hat{U}=e^{i\hat{\theta}}, the rotor algebra reads

[L^,U^]=U^,L^∣ℓ⟩=ℓ∣ℓ⟩,ℓ∈Z.(4)[\hat{L},\hat{U}]=\hat{U},\qquad\hat{L}\lvert\ell\rangle=\ell\lvert\ell\rangle,\quad\ell\in\mathbb{Z}. \tag*{(4)}

If S1\mathbb{S}^{1} is unwrapped onto the real line R\mathbb{R}, the operators L^=−i∂θ\hat{L}=-i\partial_{\theta} and U^=eiθ^\hat{U}=e^{i\hat{\theta}} satisfy the same commutation relation [L^,U^]=U^[\hat{L},\hat{U}]=\hat{U}, but L^\hat{L} acquires a continuous spectrum on R\mathbb{R}. Although wavepackets on S1\mathbb{S}^{1} and R\mathbb{R} experience identical local forces under Eq. (3), their global boundary conditions differ. Trajectories that wind around the torsion by different multiples of 2π2\pi recombine and interfere on S1\mathbb{S}^{1} while on R\mathbb{R} they propagate into distinct spatial cells. To make this explicit, let ψm(θ,x,t)\psi_{m}(\theta,x,t) be the two-component line wavefunction in the cell at θ+2πm\theta+2\pi m, with the line packet initially in the central cell and its coherent periodization as the circle state. The compact amplitude is then the sum over translated line amplitudes. The two reduced angular densities are

pcirc(θ,t)=∫dx∥∑mψm(θ,x,t)∥2,p_{\mathrm{circ}}(\theta,t)=\int dx\left\Vert\sum_{m}\psi_{m}(\theta,x,t)\right\Vert^{2},
pfold(θ,t)=∫dx∑m∥ψm(θ,x,t)∥2.(5)p_{\mathrm{fold}}(\theta,t)=\int dx\sum_{m}\left\Vert\psi_{m}(\theta,x,t)\right\Vert^{2}. \tag*{(5)}

Compactness retains the cross terms between winding histories, while the folded line readout omits them. The remaining degrees of freedom determine which cross terms survive the partial trace. If two winding contributions factorize as ψm=cmξm\psi_m = c_m \xi_m, their cross term is proportional to cmcm′∗⟨ξm′∣ξm⟩c_m c_{m'}^{*}\langle\xi_{m'} \mid\xi_m\rangle. Orthogonal vibrational states distinguish the paths and suppress their interference, while identical ones preserve it. At early times the winding amplitudes are small, so the line reproduces the compact observables until they grow.

We quantify the effect of this topological distinction on the reaction observables in Sec. IV.

Potential-energy surfaces and conical intersections

In this work we will simulate the dynamics of the Hahn-Stock model in the diabatic {∣0⟩,∣1⟩}\{\lvert0\rangle, \lvert1\rangle\} basis, but the adiabatic potential-energy surfaces make the reaction pathways transparent: the descent from the cis geometry, the branching at the conical intersections and the arrival in the trans basin or the return to the cis geometry. Denoting the Pauli matrices by σ^i\hat{\sigma}_i with i∈{x,y,z}i \in\{x,y,z\}, we write

V^=vˉ(θ^,x^)+dz(θ^,x^)σ^z+λx^σ^x,\hat{V} = \bar{v}(\hat{\theta},\hat{x}) + d_z(\hat{\theta},\hat{x})\hat{\sigma}_z + \lambda\hat{x}\hat{\sigma}_x,
vˉ(θ,x)=E12+W0−W14(1−cos⁡θ)+κx2,\bar{v}(\theta,x) = \frac{E_1}{2} + \frac{W_0-W_1}{4}(1-\cos\theta) + \frac{\kappa x}{2},
dz(θ,x)=W0+W14(1−cos⁡θ)−E12−κx2.(6)d_z(\theta,x) = \frac{W_0+W_1}{4}(1-\cos\theta) - \frac{E_1}{2} - \frac{\kappa x}{2}. \tag*{(6)}

Diagonalizing the electronic structure at fixed nuclear coordinates yields the 2π2\pi-periodic adiabatic potential-energy surfaces S0(θ,x)S_0(\theta,x) and S1(θ,x)S_1(\theta,x) shown in Figure 1(a):

S0,1(θ,x)=ωx22+vˉ(θ,x)∓dz(θ,x)2+λ2x2.(7)S_{0,1}(\theta,x) = \frac{\omega x^2}{2} + \bar{v}(\theta,x) \mp\sqrt{d_z(\theta,x)^2+\lambda^2x^2}. \tag*{(7)}

The adiabatic gap closes when dz(θ,x)d_z(\theta,x) and λx\lambda x vanish simultaneously. For λ≠0\lambda\ne0, the intersections occur at x=0x=0 and

cos⁡θc=1−2E1W0+W1,(8)\cos\theta_c = 1-\frac{2E_1}{W_0+W_1}, \tag*{(8)}

giving θ=±θc\theta=\pm\theta_c, where the subscript cc labels the conical intersection and θc=93.30∘\theta_c=93.30^\circ. The two nuclear coordinates resolve the conical geometry around each crossing. For example, near the intersection at θ=+θc\theta=+\theta_c, let δθ=θ−θc\delta\theta=\theta-\theta_c and ac=(W0+W1)sin⁡θc/4a_c=(W_0+W_1)\sin\theta_c/4. The leading gap is

S1−S0=2(acδθ−κx/2)2+λ2x2+O(δθ2,x2).(9)S_1-S_0=2\sqrt{(a_c\delta\theta-\kappa x/2)^2+\lambda^2x^2}+O(\delta\theta^2,x^2). \tag*{(9)}

The tuning coordinate acδθ−κx/2a_c\delta\theta-\kappa x/2 and the coupling coordinate λx\lambda x linearly span the branching plane of the intersection.

The energetic ordering of the diabatic states ⟨0∣V^∣0⟩\langle0\lvert\hat{V}\rvert0\rangle and ⟨1∣V^∣1⟩\langle1\lvert\hat{V}\rvert1\rangle flips across the intersection. At the cis geometry (θ,x)=(0,0)(\theta,x)=(0,0) their energies are equal to 00 and E1=2.48E_1=2.48 eV, while at the trans geometry (θ,x)=(π,0)(\theta,x)=(\pi,0) they are equal to W0=3.60W_0=3.60 eV and E1−W1=1.39E_1-W_1=1.39 eV, respectively. Therefore, the lower adiabatic surface S0S_0 is dominated by ∣0⟩\lvert0\rangle near cis and by ∣1⟩\lvert1\rangle near trans.

Photoexcitation and molecular scales

Before photoexcitation, the molecule occupies the nuclear ground state of the ∣0⟩\lvert0\rangle diabat, ∣0⟩⊗∣χcis⟩⊗∣n=0⟩x\lvert0\rangle\otimes\lvert\chi_{\mathrm{cis}}\rangle\otimes\lvert n=0\rangle_x. The vibrational factor ∣n=0⟩x\lvert n=0\rangle_x is the zero Fock state associated with the vibrational degree of freedom and ∣χcis⟩\lvert\chi_{\mathrm{cis}}\rangle is the ground state of Gℓ^2/2+W0(1−cos⁡θ)/2G\hat{\ell}^2/2+W_0(1-\cos\theta)/2 on the circle. Expanding around θ=0\theta=0 (the cis minimum) we obtain the harmonic oscillator approximation

W02(1−cos⁡θ)=W04θ2+O(θ4).(10)\frac{W_0}{2}(1-\cos\theta)=\frac{W_0}{4}\theta^2+O(\theta^4). \tag*{(10)}

Its frequency and ground-state position width are

Ωθ=GW02=0.0295 eV,\Omega_\theta=\sqrt{\frac{GW_0}{2}}=0.0295\ \mathrm{eV},
σθ,0=(G2W0)1/4=0.0905 rad.(11)\sigma_{\theta,0}=\left(\frac{G}{2W_0}\right)^{1/4}=0.0905\ \mathrm{rad}. \tag*{(11)}

Because σθ,0≪π\sigma_{\theta,0}\ll\pi, the initial packet is tightly localized within the cis basin where the harmonic approximation is accurate.

Photoexcitation enters as a sudden electronic transition at fixed nuclear coordinates, ∣1⟩⟨0∣⊗I\lvert1\rangle\langle0\rvert\otimes\mathbb{I}. This description is valid for pulses much shorter than the two characteristic nuclear periods, 2πℏ/ω=21.82\pi\hbar/\omega=21.8 fs for the coupling vibration and 2πℏ/Ωθ=1402\pi\hbar/\Omega_\theta=140 fs for the torsional well. The resulting photoexcited state is the initial state of the dynamics

∣Ψ(0)⟩=∣1⟩⊗∣χcis⟩⊗∣n=0⟩x.(12)\lvert\Psi(0)\rangle=\lvert1\rangle\otimes\lvert\chi_{\mathrm{cis}}\rangle\otimes\lvert n=0\rangle_x. \tag*{(12)}

This specifies a product Franck-Condon preparation, which deposits in the molecule the vertical excitation energy

Evert=[⟨1∣V^∣1⟩−⟨0∣V^∣0⟩](θ,x)=(0,0)=E1.(13)E_{\mathrm{vert}}=\left[\langle1\lvert\hat{V}\rvert1\rangle-\langle0\lvert\hat{V}\rvert0\rangle\right]_{(\theta,x)=(0,0)}=E_1. \tag*{(13)}

At cis, the excited diabat is at a local maximum of the torsional potential, so the wavepacket accelerates away from the cis region toward the intersection region. If the entire excitation energy E1E_1 is converted into torsional rotation, GℓE2/2=E1G\ell_E^2/2=E_1, the angular momentum reaches its energetic bound ℓE\ell_E corresponding to an angular velocity GℓE/ℏG\ell_E/\hbar and rotation period TET_E:

ℓE=2E1G≃101,TE=2πℏGℓE≃84.4 fs.(14)\ell_E=\sqrt{\frac{2E_1}{G}}\simeq101,\qquad T_E=\frac{2\pi\hbar}{G\ell_E}\simeq84.4\ \mathrm{fs}. \tag*{(14)}

A packet initially 0.090.09 rad wide spreads over angular momenta up to ℓ∼∣ℓE∣\ell\sim\lvert\ell_E\rvert, requiring the rotor register to resolve roughly 200 angular momentum states. A propagation window of several hundred femtoseconds covers several full windings at TE≃84T_E \simeq84 fs and ten to twenty vibrational periods, so it probes the global configuration space. For simulations on R\mathbb{R}, the localized packet is initialized in the central cell [−π,π][-\pi,\pi].

Local vibronic approximations

Standard vibronic models expand the diabatic potential around a reference geometry [1, 15] on the line R\mathbb{R} with momentum p^θ=−i∂θ\hat{p}_{\theta}=-i\partial_{\theta}. The linear vibronic coupling (LVC) approximation retains a state-independent potential along with linear displacements.

H^LVC=G2p^θ2+W04θ^2+ω2(p^x2+x^2)+(E1+κx^)I^2−σ^z2+λx^σ^x.(15)\hat{H}_{\mathrm{LVC}}=\frac{G}{2}\hat{p}_{\theta}^{2}+\frac{W_{0}}{4}\hat{\theta}^{2}+\frac{\omega}{2}\left(\hat{p}_{x}^{2}+\hat{x}^{2}\right)+\frac{\left(E_{1}+\kappa\hat{x}\right)\hat{I}_{2}-\hat{\sigma}_{z}}{2}+\lambda\hat{x}\hat{\sigma}_{x}. \tag*{(15)}

In Eq. (15), the torsion appears only through its own harmonic Hamiltonian, which commutes with every other term. The torsion therefore decouples and the electronic transitions are only driven by xx alone, while the tails of the localized torsional state give an exponentially suppressed probability in the trans region.

The quadratic vibronic coupling (QVC) model restores the state-dependent quadratic torsional curvature,

H^QVC=G2p^θ2+ω2(p^x2+x^2)+(W0θ^24λx^λx^E1−W1θ^24+κx^).(16)\hat{H}_{\mathrm{QVC}}=\frac{G}{2}\hat{p}_{\theta}^{2}+\frac{\omega}{2}\left(\hat{p}_{x}^{2}+\hat{x}^{2}\right)+\begin{pmatrix}\dfrac{W_{0}\hat{\theta}^{2}}{4} & \lambda\hat{x}\\ \lambda\hat{x} & E_{1}-\dfrac{W_{1}\hat{\theta}^{2}}{4}+\kappa\hat{x}\end{pmatrix}. \tag*{(16)}

While H^QVC\hat{H}_{\mathrm{QVC}} captures the initial acceleration and the first nonadiabatic passage, its excited diabatic surface is an inverted parabola that is unbounded from below, lacking the global restoring forces of the cosine potential.

Observables and reference calculations

We characterize the time-evolved state ∣Ψ(t)⟩=e−iH^t/ℏ∣Ψ(0)⟩|\Psi(t)\rangle=e^{-i\hat{H}t/\hbar}|\Psi(0)\rangle via expectation values ⟨O^⟩t=⟨Ψ(t)∣O^∣Ψ(t)⟩\langle\hat{O}\rangle_t=\langle\Psi(t)|\hat{O}|\Psi(t)\rangle. For this paper, we mainly focus on three physical questions. Whether electronic population has transferred, whether the torsion has reached the trans basin and whether a photoproduct has formed, that is, whether the torsion is in the trans basin while the electronic state is on the lower adiabatic surface S0S_0.

To calculate the electronic population transfer, we define the excited-configuration population

P1(t)=⟨1∣⟨1∣⊗I^θ⊗I^x⟩t=Tr⁡[ρ^el(t)∣1⟩⟨1∣].(17)P_{1}(t)=\langle1|\langle1|\otimes\hat{I}_{\theta}\otimes\hat{I}_{x}\rangle_t=\operatorname{Tr}\left[\hat{\rho}_{\mathrm{el}}(t)|1\rangle\langle1|\right]. \tag*{(17)}

where I^θ\hat{I}_{\theta} and I^x\hat{I}_{x} are the identities on the torsional and vibrational factors and ρ^el(t)=Tr⁡θ,x∣Ψ(t)⟩⟨Ψ(t)∣\hat{\rho}_{\mathrm{el}}(t)=\operatorname{Tr}_{\theta,x}|\Psi(t)\rangle\langle\Psi(t)| is the reduced electronic density matrix obtained by tracing out those nuclear degrees of freedom.

Nuclear arrival in the trans basin is defined by the trans indicator χtr(θ)\chi_{\mathrm{tr}}(\theta) for ∣θ mod 2π∣>π/2|\theta\bmod2\pi|>\pi/2, corresponding to the operator χ^tr=I^2⊗χtr(θ^)⊗I^x\hat{\chi}_{\mathrm{tr}}=\hat{I}_{2}\otimes\chi_{\mathrm{tr}}(\hat{\theta})\otimes\hat{I}_{x}. Since ∣θ mod 2π∣>π/2|\theta\bmod2\pi|>\pi/2 is equivalent to cos⁡θ<0\cos\theta<0, the indicator is a function of the globally defined rotor operator alone

χtr(θ^)=Θ(−cos⁡θ^),cos⁡θ^=U^+U^†2,(18)\chi_{\mathrm{tr}}(\hat{\theta})=\Theta(-\cos\hat{\theta}),\qquad\cos\hat{\theta}=\frac{\hat{U}+\hat{U}^{\dagger}}{2}, \tag*{(18)}

with Θ\Theta the Heaviside step function.

Resolving chemical product formation requires projecting onto the adiabatic basis. After the vertical excitation, the packet descends adiabatically on S1S_1. Near the conical intersection, where the gap closes, adiabatic evolution breaks down and the packet branches. The component that flips to S0S_0 continues into the trans basin. Conversely, the branch that remains in S1S_1 faces a potential barrier rising toward W0=3.60W_0=3.60 eV at trans, above the available energy E1E_1, so it reflects back to the intersection region where it branches again. The part that further remains on S1S_1 returns toward the cis region before descending to the intersection again. Product formation at a given time therefore requires two conditions jointly: arrival in the trans basin and support on S0S_0. Each condition alone is reversible, since trans amplitude on the upper surface can swing back through the intersection and lower-surface amplitude at cis is recovered reactant. The potential matrix has the form vˉI^2+h⃗⋅σ⃗^\bar{v}\hat{I}_{2}+\vec{h}\cdot\hat{\vec{\sigma}} with h⃗=(λx,0,dz)\vec{h}=(\lambda x,0,d_z), so its eigenvalues are vˉ±∣h⃗∣\bar{v}\pm|\vec{h}| and its eigenvectors depend only on the direction of h⃗\vec{h}. Away from the conical intersections, where h⃗=0\vec{h}=0, the projector onto the lower surface is

Π^−(θ,x)=12[I^2−dzσ^z+λxσ^xdz2+λ2x2]=12[I^2−h⃗∣h⃗∣⋅σ⃗^].(19)\hat{\Pi}_{-}(\theta,x)=\frac{1}{2}\left[\hat{I}_{2}-\frac{d_z\hat{\sigma}_{z}+\lambda x\hat{\sigma}_{x}}{\sqrt{d_z^{2}+\lambda^{2}x^{2}}}\right]=\frac{1}{2}\left[\hat{I}_{2}-\frac{\vec{h}}{|\vec{h}|}\cdot\hat{\vec{\sigma}}\right]. \tag*{(19)}

which defines the finite-time product probability Q−Q_{-} and the conditional product fraction Y−(t)Y_{-}(t), the trans probability given that the molecule is on the lower surface,

Q−(t)=⟨χ^trΠ^−(θ^,x^)⟩t=P(trans,S0),Q_{-}(t)=\langle\hat{\chi}_{\mathrm{tr}}\hat{\Pi}_{-}(\hat{\theta},\hat{x})\rangle_t=P(\mathrm{trans},S_0),
Y−(t)=Q−(t)⟨Π^−⟩t=P(trans,S0)P(S0)=P(trans∣S0).(20)Y_{-}(t)=\frac{Q_{-}(t)}{\langle\hat{\Pi}_{-}\rangle_t}=\frac{P(\mathrm{trans},S_0)}{P(S_0)}=P(\mathrm{trans}\mid S_0). \tag*{(20)}

Without an environment the dynamics is closed and Q−Q_{-} keeps fluctuating rather than settling to a steady-state quantum yield. The observable Y−Y_{-} measures the branching fraction of the S0S_0 population between trans product and cis reactant.

We also consider the return probability

R(t)=∣⟨Ψ(0)∣Ψ(t)⟩∣2,(21)\mathcal{R}(t)=|\langle\Psi(0)|\Psi(t)\rangle|^{2}, \tag*{(21)}

defined relative to the complete initial state of each model or encoding.

Throughout the paper, we compare the curves associated with the various models and encodings to a converged reference of the observable OO and demand that the largest absolute error over the time window T\mathcal{T},

ϵO=max⁡t∈T∣Oenc(t)−Oref(t)∣,(22)\epsilon_O = \max_{t \in\mathcal{T}} \left|O_{\mathrm{enc}}(t)-O_{\mathrm{ref}}(t)\right|, \tag*{(22)}

is less than a given tolerance. For most of the simulations, the tolerance we set is 0.01. However, for experiments that only need to detect the effect of compactness a potentially larger tolerance can suffice, which we discuss in Sec. VI B. Since Y−Y_- is only well defined when the wavepacket has a non-zero weight in S0S_0, we quote its value only at times when ⟨Π^−⟩t≥0.05\langle\hat{\Pi}_-\rangle_t \ge0.05. Numerical details can be found in Appendix A.

Compact coordinates with qubits and qumodes

A qumode is a harmonic oscillator with the standard canonical pair [q^,p^]=iI^[\hat{q},\hat{p}]=i\hat{I} on L2(R)L^2(\mathbb{R}), with continuous noncompact position and momentum coordinates. Therefore, it naturally encodes the molecular vibration. However, the torsion obeys the rotor algebra of Eq. (4), with integer angular momentum on a compact coordinate. The molecular potential of our model is periodic in θ^\hat{\theta} and carries over to the qumode under θ^↦q^\hat{\theta}\mapsto\hat{q}, but the angular identification of the compact coordinate must be encoded. We use two encodings of a rotor in qumodes. The first one wraps the qumode line onto the circle by selecting a translation sector of one qumode via a GKP construction [64, 68]. The second one embeds the circle as a constrained subspace of two qumodes [62]. Figure 2 sketches both in their finite-energy form. Their finite-energy errors have a common description in terms of conserved sectors, which also provides systematic ways to reduce them. Both constructions extend to collective encodings distributed over several qumodes and admit Hamiltonian counterterms.

Figure 2. Finite-energy compact encodings, with widths exaggerated for visibility. (a) Wrapped GKP encoding: momentum distribution of a finite-energy comb, Eq. (29). Each integer angular momentum ℓ\ell carries a tooth of width σν\sigma_\nu of Eq. (30) and the molecular coefficients ∣cℓ∣2|c_\ell|^2 set the envelope. (b) Two-qumode circle encoding: position distribution of a finite-width annulus, Eq. (37), of radius RR and radial width σr≃R/(4A)\sigma_r \simeq R/(4A) in the plane of q0q_0 and q1q_1. The torsion θ\theta is the polar angle.

GKP encoding and the direct integral of rotors

The logic of this encoding is to make the qumode coordinate periodic by construction. Define the unwrapped angle θ^R≡sq^\hat{\theta}_{R}\equiv s\hat{q}, where s>0s>0 and identify it modulo 2π2\pi, θ^R≃θ^R+2πj\hat{\theta}_{R}\simeq\hat{\theta}_{R}+2\pi j with j∈Zj\in\mathbb{Z}, equivalently q^≃q^+2πj/s\hat{q}\simeq\hat{q}+2\pi j/s. Once the identification holds, the rotor operators carry over through the formal replacements

θ^↦θ^R=sq^,U^=eiθ^R,L^R=p^/s.(23)\hat{\theta}\mapsto\hat{\theta}_{R}=s\hat{q},\qquad\hat{U}=e^{i\hat{\theta}_{R}},\qquad\hat{L}_{R}=\hat{p}/s. \tag*{(23)}

The identification itself is enforced by a stabilizer. The idea is that we only work within the +1+1 eigenspace of the translation

S^=e2πiL^R,(24)\hat{S}=e^{2\pi i\hat{L}_{R}}, \tag*{(24)}

which shifts the unwrapped angle θ^R\hat{\theta}_{R} by one period 2π2\pi. Equivalently, the stabilizer in Eq. (24) acts as the identity on the space it stabilizes and that is exactly where

Spec⁡(L^R)≃Z\operatorname{Spec}(\hat{L}_{R})\simeq\mathbb{Z}, namely, the rescaled momentum has integer values. Within this sector, the most general state is

∑ℓ∈Zcℓ∣p=sℓ⟩.(25)\sum_{\ell\in\mathbb{Z}}c_{\ell}\lvert p=s\ell\rangle. \tag*{(25)}

The above can equivalently be seen as a superposition of comb states ∣θ⟩≡∑ℓ∈Ze−iℓθ∣p=sℓ⟩\lvert\theta\rangle\equiv\sum_{\ell\in\mathbb{Z}}e^{-i\ell\theta}\lvert p=s\ell\rangle, which correspond exactly to the encoded angular states of the rotor in the qumode Hilbert space. These belong to the GKP family. Indeed, each encoded ∣θ⟩\lvert\theta\rangle state is a so-called qunaught state, i.e. a grid state that encodes no logical information, unlike the usual GKP state that encodes a logical qubit. Thus the GKP identification makes the momentum spectrum discrete while retaining the full angular wavefunction. The amplitudes cℓc_{\ell}, which can themselves be states of the other molecular coordinates, are determined by the molecular wavefunction. A momentum eigenstate occupies a single tooth of the comb, while a localized angular wavepacket requires a coherent superposition across many teeth.

In order to formally understand why we can wrap the real line onto the circle, and thus connect the rotor with the oscillator, we make use of the Zak transform. Every eigenvalue of L^R\hat{L}_{R} on the line splits uniquely as LR=ℓ+νL_{R}=\ell+\nu with an integer part ℓ∈Z\ell\in\mathbb{Z} and a fractional part −1/2≤ν<1/2-1/2\le\nu<1/2. The integer part plays the role of a rotor angular momentum. The fractional part is read out by the stabilizer, since S^\hat{S} has eigenvalue e2πiνe^{2\pi i\nu} on such a state and the encoded sector of Eq. (25) corresponds to ν=0\nu=0. The Zak transform organizes the whole line by this fractional momentum and gives the unitary decomposition [64]

L2(R)≃∫−1/21/2 ⁣ ⁣⊕dν Lν2(S1).(26)L^2(\mathbb{R})\simeq\int_{-1/2}^{1/2}\!\!{}^{\oplus}d\nu\,L^2_\nu(\mathbb{S}^1). \tag*{(26)}

Each integrand Lν2(S1)L^2_\nu(\mathbb{S}^1) of the direct integral is called a fiber. Its wavefunctions satisfy the twisted periodicity condition ψ(θ+2π)=e2πiνψ(θ)\psi(\theta+2\pi)=e^{2\pi i\nu}\psi(\theta). Multiplying them by e−iνθe^{-i\nu\theta} makes them strictly 2π2\pi periodic and on these periodic wavefunctions the Hamiltonian of the fiber is

H^ν=G2(L^+νI^)2+ω2(p^x2+x^2)+V^(θ^,x^).(27)\hat{H}_{\nu}=\frac{G}{2}(\hat{L}+\nu\hat{I})^2+\frac{\omega}{2}(\hat{p}_{x}^{2}+\hat{x}^{2})+\hat{V}(\hat{\theta},\hat{x}). \tag*{(27)}

Here, L^=−i∂θ\hat{L}=-i\partial_{\theta} has integer eigenvalues on the periodic fiber. These fiber Hamiltonians assemble into the line Hamiltonian

H^R=G2L^R2+ω2(p^x2+x^2)+V^(sq^,x^),(28)\hat{H}_{R}=\frac{G}{2}\hat{L}_{R}^{2}+\frac{\omega}{2}(\hat{p}_{x}^{2}+\hat{x}^{2})+\hat{V}(s\hat{q},\hat{x}), \tag*{(28)}

which is obtained by substituting Eq. (23) into Eq. (2). Since [H^R,S^]=0[\hat{H}_{R},\hat{S}]=0, the fiber label ν\nu is conserved and the physical compact molecular Hamiltonian is H^0\hat{H}_0, i.e. the one with vanishing fiber label, or equivalently, where the angular momentum spectrum is integer. The periodic potential acts identically in every fiber, so the fibers differ only through the kinetic shift visible in Eq. (27).

This description is the ideal limit of the encoding. In reality, a single fiber is spanned by momentum eigenstates and is non-normalizable in L2(R)L^2(\mathbb{R}). A finite-energy state carries a distribution over fiber labels concentrated near ν=0\nu=0. Because ν\nu is conserved, this distribution is fixed in time and only the state within each fiber evolves.

A normalizable state follows by giving each tooth of the ideal comb a finite width. Writing the periodic molecular state through its angular Fourier coefficients, ψmol(θ,x)=∑ℓcℓ(x)eiℓθ\psi_{\mathrm{mol}}(\theta,x)=\sum_\ell c_\ell(x)e^{i\ell\theta}, we replace the momentum delta function at each integer ℓ\ell by the same narrow Gaussian

Ψ~r(LR,x)=N∑ℓe−(LR−ℓ)2/(4σν2)cℓ(x),(29)\widetilde{\Psi}_r(L_R,x)=\mathcal{N}\sum_\ell e^{-(L_R-\ell)^2/(4\sigma_\nu^2)}c_\ell(x), \tag*{(29)}
σν=e−reff4π.(30)\sigma_\nu=\frac{e^{-r_{\mathrm{eff}}}}{\sqrt{4\pi}}. \tag*{(30)}

No overall envelope over ℓ\ell is needed, since the coefficients cℓc_\ell of a normalizable molecular state already decay beyond the populated momenta. For σν≪1/2\sigma_\nu\ll1/2 the teeth do not overlap, so every fiber carries the same molecular state with the weight w(ν)=e−ν2/(2σν2)/2πσν2w(\nu)=e^{-\nu^2/(2\sigma_\nu^2)}/\sqrt{2\pi\sigma_\nu^2}, concentrated near the physical fiber ν=0\nu=0.

The effective squeezing reffr_{\mathrm{eff}} sets the tooth width in rotor units. Since L^R=p^/s\hat{L}_R=\hat{p}/s, the physical momentum width is sσνs\sigma_\nu. Matching it to the width e−rphys/2e^{-r_{\mathrm{phys}}}/\sqrt{2} of a momentum-squeezed vacuum gives the physical squeezing and its decibel value,

rphys=reff+log⁡2πs,SdB=20rphyslog⁡10.(31)r_{\mathrm{phys}}=r_{\mathrm{eff}}+\log\frac{\sqrt{2\pi}}{s},\qquad S_{\mathrm{dB}}=\frac{20r_{\mathrm{phys}}}{\log10}. \tag*{(31)}

Because the fiber label is conserved, a measurement returns the average over the fiber distribution. Thus, for a periodic observable O^\hat{O}, we have

⟨O^⟩(t)=∫−1/21/2dν w(ν)⟨ψν(t)∣O^∣ψν(t)⟩.(32)\langle\hat{O}\rangle(t)=\int_{-1/2}^{1/2}d\nu\,w(\nu)\langle\psi_\nu(t)|\hat{O}|\psi_\nu(t)\rangle. \tag*{(32)}

A conditional probability such as Y−Y_- is the ratio of the averaged numerator and denominator. The fibers differ only through the kinetic shift of Eq. (27), so the average dephases the rotor. For a free rotor the coherence between ℓ\ell and ℓ′\ell' acquires the factor

exp⁡[−12G2σν2t2(ℓ−ℓ′)2].(33)\exp\left[-\frac{1}{2}G^2\sigma_\nu^2t^2(\ell-\ell')^2\right]. \tag*{(33)}

A finite tooth width therefore biases late-time observables. Narrowing the teeth suppresses this bias but costs squeezing.

Two-qumode circle and radial sectors

The second encoding we consider realizes the circle inside the plane of two qumodes, with canonical pairs [q^μ,p^η]=iδμηI^[\hat{q}_\mu,\hat{p}_\eta]=i\delta_{\mu\eta}\hat{I}, with μ,η=0,1\mu,\eta=0,1. The logic parallels the GKP construction. The plane contains the circle as the constraint surface q^02+q^12=R2I^\hat{q}_0^2+\hat{q}_1^2=R^2\hat{I} of radius RR and we only work with states concentrated on that surface [62]. The angular momentum and the angular phase are represented by

J^=q^0p^1−q^1p^0,U^R=q^0+iq^1R.(34)\hat{J}=\hat{q}_0\hat{p}_1-\hat{q}_1\hat{p}_0,\qquad\hat{U}_R=\frac{\hat{q}_0+i\hat{q}_1}{R}. \tag*{(34)}

The exact unitary phase operator on the whole plane, (q^0+iq^1)/q^02+q^12(\hat{q}_0+i\hat{q}_1)/\sqrt{\hat{q}_0^2+\hat{q}_1^2}, requires reciprocals and square roots of quadratures, which are not available gates. We use U^R\hat{U}_R and absorb the constant 1/R1/R into the Hamiltonian parameters through L^→J^\hat{L}\to\hat{J} and cos⁡θ^→q^0/R\cos\hat{\theta}\to\hat{q}_0/R. The commutator [J^,U^R]=U^R[\hat{J},\hat{U}_R] = \hat{U}_R holds on the whole plane, while U^R†U^R=(q^02+q^12)/R2=!I^\hat{U}_R^\dagger\hat{U}_R=(\hat{q}_0^2+\hat{q}_1^2)/R^2\overset{!}{=}\hat{I} holds only on the constraint surface, where the substitutions reproduce the rotor Hamiltonian.

With a^±=(a^0∓ia^1)/2\hat{a}_{\pm}=(\hat{a}_0\mp i\hat{a}_1)/\sqrt{2}, one obtains

J^=i(a^1†a^0−a^0†a^1)=n^+−n^−,\hat{J}=i(\hat{a}_1^\dagger\hat{a}_0-\hat{a}_0^\dagger\hat{a}_1)=\hat{n}_+-\hat{n}_-,

with n^±=a^±†a^±\hat{n}_{\pm}=\hat{a}_{\pm}^\dagger\hat{a}_{\pm}.

{#eq:35}

Integer angular momentum is therefore built into the planar representation. The operator inequality ∣J^∣≤n^++n^−|\hat{J}| \le\hat{n}_{+}+\hat{n}_{-} gives a minimum occupation for each represented angular-momentum sector.

The linear substitution needs one refinement because if we naively insert it into Eq. (3), it produces terms linear in q^0\hat{q}_{0}, so the ground-d iabat well, whose cosine enters with the positive coefficient W0/2W_{0}/2, becomes unbounded from below on the plane. As in Ref. [62], we complete it to a form quadratic in the quadratures,

1−cos⁡θ^⟼(q^0−RI^)2+q^122R2,(35)1-\cos\hat{\theta}\longmapsto\frac{(\hat{q}_{0}-R\hat{I})^{2}+\hat{q}_{1}^{2}}{2R^{2}}, \tag*{(35)}

applied to the ground diabat. The excited diabat carries the cosine with the opposite sign, so we write it about its own minimum at trans, −W12(1−cos⁡θ)=−W1+W12(1+cos⁡θ)-\frac{W_{1}}{2}(1-\cos\theta)=-W_{1}+\frac{W_{1}}{2}(1+\cos\theta), and complete 1+cos⁡θ^⟼[(q^0+RI^)2+q^12]/(2R2)1+\cos\hat{\theta}\longmapsto[(\hat{q}_{0}+R\hat{I})^{2}+\hat{q}_{1}^{2}]/(2R^{2}). Both right-hand sides are nonnegative on the whole plane, so both diabats are bounded from below, and on the constraint surface they agree exactly with the linear substitution.

The exact structure is again a decomposition into conserved sectors since the Hamiltonian conserves the radius. In the coordinate representation, write q0=Rρcos⁡θq_{0}=R\rho\cos\theta and q1=Rρsin⁡θq_{1}=R\rho\sin\theta. The pair (R,ρ)(R,\rho) mirrors the pair (s,ν)(s,\nu) of the GKP encoding. The fixed radius RR, like the scale ss, is a chosen parameter of the encoding: it never appears as an operator and it sets physical costs such as the minimum occupation. The dimensionless ρ\rho measures the radius in units of RR, so the constraint surface is ρ=1\rho=1 and ρ\rho serves as the conserved fiber label, with the offset ρ−1\rho-1 playing the part of ν\nu. No radial kinetic energy is added to the molecular Hamiltonian, so the radius is conserved. Writing H^x=ω(p^x2+x^2)/2\hat{H}_{x}=\omega(\hat{p}_{x}^{2}+\hat{x}^{2})/2,

L2(R2)≃∫0∞⊕ρ dρ Lρ2(S1),L^{2}(\mathbb{R}^{2})\simeq\int_{0}^{\infty}{}^{\oplus}\rho\,\mathrm{d}\rho\,L_{\rho}^{2}(S^{1}),
H^ρ=G2J^2+H^x+V^(1∓cos⁡θ^⟼1+ρ22∓ρcos⁡θ^,x^),(36)\hat{H}_{\rho}=\frac{G}{2}\hat{J}^{2}+\hat{H}_{x} +\hat{V}\left(1\mp\cos\hat{\theta}\longmapsto\frac{1+\rho^{2}}{2}\mp\rho\cos\hat{\theta},\hat{x}\right), \tag*{(36)}

where J^\hat{J} acts as −i∂θ-i\partial_{\theta} on each fiber and the potential rule is Eq. (35) and its mirror evaluated at radius ρ\rho, with the upper sign in the ground diabat and the lower sign in the excited diabat. Each radius ρ\rho is a fiber carrying a rotor Hamiltonian. Compared with Eq. (27), the kinetic term is now the same in every fiber and the potential varies with ρ\rho. Because the constant (1+ρ2)/2(1+\rho^{2})/2 enters the two diabats with the coefficients W0/2W_{0}/2 and W1/2W_{1}/2, it shifts their energies differently, so the electronic gap and the conical intersections move at first order in ρ−1\rho-1. As in the GKP case, the states of a single fiber are non-normalizable with respect to the radial measure, each radial label is conserved independently and a normalizable state carries a conserved distribution over ρ\rho concentrated near the physical fiber.

For the finite-energy preparation we replace the exact constraint by a narrow annulus. Its radial amplitude is

fA(ρ)=NAexp⁡[−A2(ρ2−1)2],(37)f_{A}(\rho)=\mathcal{N}_{A}\exp[-A^{2}(\rho^{2}-1)^{2}], \tag*{(37)}

normalized with the measure ρ dρ\rho\,\mathrm{d}\rho. This is the action of the non-Gaussian filter e−A2(ρ^2−1)2e^{-A^{2}(\hat{\rho}^{2}-1)^{2}} on a smooth input state, with ρ^2=(q^02+q^12)/R2\hat{\rho}^{2}=(\hat{q}_{0}^{2}+\hat{q}_{1}^{2})/R^{2}. For A≫1A\gg1 the annulus has fractional radial width σρ≃1/(4A)\sigma_{\rho}\simeq1/(4A) and physical width σr≃R/(4A)\sigma_{r}\simeq R/(4A), which play the role of the tooth width σν\sigma_{\nu} of the GKP encoding.

The constraint filter has a direct ancilla construction. Let C^=q^02+q^12−R2I^2\hat{C}=\hat{q}_{0}^{2}+\hat{q}_{1}^{2}-R^{2}\hat{I}^{2}, prepare an ancilla with momentum variance e−2ra/2e^{-2r_{a}}/2 and apply

V^a=e−iγC^q^a.(38)\hat{V}_{a}=e^{-i\gamma\hat{C}\hat{q}_{a}}. \tag*{(38)}

Conditioning on the central homodyne outcome gives the Kraus density

⟨pa=0∣V^a∣ψa⟩∝e−γ2e2raC^2/2,γera=2AR2.(39)\langle p_{a}=0|\hat{V}_{a}|\psi_{a}\rangle\propto e^{-\gamma^{2}e^{2r_{a}}\hat{C}^{2}/2},\qquad\gamma e^{r_{a}}=\frac{\sqrt{2}A}{R^{2}}. \tag*{(39)}

For details of the circuit that builds this filter, see Ref. [62]. Interaction strength and ancilla squeezing therefore enter through one known combination, which sets the annular width. A physical acceptance window has finite probability and slightly broadens the annulus. Applied to an input carrying the molecular state ψmol(θ,x)\psi_{\mathrm{mol}}(\theta,x), the filter produces Eq. (37).

For either encoding, the finite-width bias is an average of fiber trajectories over the conserved distribution w(ν)w(\nu) or w(ρ)w(\rho). Conditioning the readout on the fiber label removes the bias at the cost of discarded shots. Sections III C and III D give two ways to reduce the cost of a narrow distribution, which can be combined.

Collective encodings across multiple qumodes

The first freedom is to distribute an encoded qumode over several physical modes. For kk canonical pairs, define

X^=1k∑j=1kq^j,P^=1k∑j=1kp^j,[X^,P^]=iI^.(40)\hat{X}=\frac{1}{\sqrt{k}}\sum_{j=1}^{k}\hat{q}_{j},\qquad \hat{P}=\frac{1}{\sqrt{k}}\sum_{j=1}^{k}\hat{p}_{j},\qquad [\hat{X},\hat{P}]=i\hat{I}. \tag*{(40)}

The unwrapped GKP coordinate is θ^R=sX^\hat{\theta}_{\mathrm{R}}=s\hat{X}, with L^R=P^/s\hat{L}_{\mathrm{R}}=\hat{P}/s, or

θ^R=sk∑jq^j,L^R=1sk∑jp^j.(41)\hat{\theta}_{\mathrm{R}}=\frac{s}{\sqrt{k}}\sum_{j}\hat{q}_{j},\qquad \hat{L}_{\mathrm{R}}=\frac{1}{s\sqrt{k}}\sum_{j}\hat{p}_{j}. \tag*{(41)}

An orthogonal passive transformation completes (X^,P^)(\hat{X},\hat{P}) with k−1k-1 difference modes [74]. Preparing these modes in vacuum gives one collective encoded state with no additional occupied degrees of freedom. Spreading a mode uniformly over many vacuum modes in this way bounds the occupation that each mode must represent [75].

Hamiltonians that depend only on X^\hat{X} and P^\hat{P} preserve this sector.

For the circle, we distribute each Cartesian pair in the same way

q^μ=1k∑jq^μj,p^μ=1k∑jp^μj,μ=0,1.(42)\hat{q}_{\mu}=\frac{1}{\sqrt{k}}\sum_j\hat{q}_{\mu j},\qquad\hat{p}_{\mu}=\frac{1}{\sqrt{k}}\sum_j\hat{p}_{\mu j},\qquad\mu=0,1. \tag*{(42)}

The constraint acts on q^02+q^12\hat{q}_0^2+\hat{q}_1^2 and there are 2k2k physical torsion modes. One can equivalently distribute the two circular modes, provided the state and all operators are transformed consistently. The construction adds empty difference modes and changes the mode basis. It preserves both the molecular dynamics and the total number distribution, while reducing the occupation required in each output mode. The quantitative tradeoff and its gate implementation are derived in Sec. VI.

Hamiltonian counterterms

The second freedom lies in the Hamiltonian itself. The molecule fixes the encoded Hamiltonian only on the physical fiber. Away from it we may add terms, which we call counterterms. They must commute with the fiber label, S^\hat{S} of Eq. (24) for the GKP encoding and ρ^2\hat{\rho}^2 for the circle, so that the label stays conserved and the state still splits into independent fibers. They must also leave the physical fiber unchanged, use the available gates and keep the Hamiltonian bounded from below. Chosen to cancel the leading dependence on the fiber label, they let a broader fiber distribution reach the same accuracy. The encoding then needs less squeezing and fewer photons.

For the GKP encoding the fiber label enters only through the kinetic shift of Eq. (27). Replacing the encoded momentum in the kinetic energy GL^R2/2G\hat{L}_{R}^{2}/2 by

f3(L^R)=L^R−sin⁡(2πL^R)2π(43)f_3(\hat{L}_R)=\hat{L}_R-\frac{\sin(2\pi\hat{L}_R)}{2\pi} \tag*{(43)}

removes this dependence at leading order. At momentum ℓ+ν\ell+\nu, f3=ℓ+(2π2/3)ν3+O(ν5)f_3=\ell+(2\pi^2/3)\nu^3+O(\nu^5), so Gf3(L^R)2/2Gf_3(\hat{L}_R)^2/2 reproduces the kinetic energy for every integer ℓ\ell and the coherence loss of Eq. (33) scales with ⟨ν6⟩\langle\nu^6\rangle instead of σν2\sigma_\nu^2. This kinetic phase is not Gaussian and can be realized with the trigonometric gate synthesis of Refs. [52, 53]. More generally, the trigonometric polynomial

f2k+1(L^R)=L^R+1π1(2kk)∑j=1k(−1)jj(2kk−j)sin⁡(2πjL^R)(44)f_{2k+1}(\hat{L}_R)=\hat{L}_R+\frac{1}{\pi}\frac{1}{\binom{2k}{k}}\sum_{j=1}^{k}\frac{(-1)^j}{j}\binom{2k}{k-j}\sin(2\pi j\hat{L}_R) \tag*{(44)}

reduces to Eq. (43) for k=1k=1. Its derivative f2k+1′(L^R)=4ksin⁡2k(πL^R)/(2kk)f'_{2k+1}(\hat{L}_R)=4^k\sin^{2k}(\pi\hat{L}_R)/\binom{2k}{k} is nonnegative and vanishes to order 2k2k at every integer, so f2k+1f_{2k+1} leaves every integer fixed and f2k+1(ℓ+ν)=ℓ+O(ν2k+1)f_{2k+1}(\ell+\nu)=\ell+O(\nu^{2k+1}). It cancels the fiber dependence through order ν2k\nu^{2k} at the cost of kk trigonometric harmonics. As kk grows, f2k+1f_{2k+1} approaches the staircase that rounds LRL_R to the nearest integer.

For the circle embedding, the exact fiber-blind cosine would divide q^0\hat{q}_0 by the radial coordinate ρ^\hat{\rho}. Since ρ^2=(q^02+q^12)/R2\hat{\rho}^2=(\hat{q}_0^2+\hat{q}_1^2)/R^2 is a polynomial, we expand 1/ρ^=(ρ^2)−1/21/\hat{\rho}=(\hat{\rho}^2)^{-1/2} in powers of ρ^2−I^\hat{\rho}^2-\hat{I} and truncate at order kk, such that

C^2k+1=q^0∑n=0k(−1)n4n(2nn)(ρ^2−I^)n(45)\hat{C}_{2k+1}=\hat{q}_0\sum_{n=0}^{k}\frac{(-1)^n}{4^n}\binom{2n}{n}\left(\hat{\rho}^2-\hat{I}\right)^n \tag*{(45)}

is a polynomial of degree 2k+12k+1 in the quadratures. On the fiber of radius ρ=1+δ\rho=1+\delta, C^2k+1/R=g2k+1(ρ)cos⁡θ^\hat{C}_{2k+1}/R=g_{2k+1}(\rho)\cos\hat{\theta} with

g2k+1(1+δ)=1+(−1)k2k+1(2k+2k+1)δk+1+O(δk+2),(46)g_{2k+1}(1+\delta)=1+\frac{(-1)^k}{2^{k+1}}\binom{2k+2}{k+1}\delta^{k+1}+O(\delta^{k+2}), \tag*{(46)}

so the cosine strength is fiber blind through order δk\delta^k. The lowest order k=1k=1 gives C^3=32q^0−12R2q^0(q^02+q^12)\hat{C}_3=\frac{3}{2}\hat{q}_0-\frac{1}{2R^2}\hat{q}_0(\hat{q}_0^2+\hat{q}_1^2) with g3(1+δ)=1−32δ2−12δ3g_3(1+\delta)=1-\frac{3}{2}\delta^2-\frac{1}{2}\delta^3. In an experiment the coefficients can also be calibrated by tuning them until the fiber-resolved observable is stationary across the populated fibers.

The fiber constant is flattened in the same way, which also restores boundedness, since the odd tail of C^2k+1\hat{C}_{2k+1} alone would be unbounded. The completed potential is replaced by

1−cos⁡θ^⟶I^+η(ρ^2−I^)2m−C^2k+1R,(47)1-\cos\hat{\theta}\longrightarrow\hat{I}+\eta\left(\hat{\rho}^2-\hat{I}\right)^{2m}-\frac{\hat{C}_{2k+1}}{R}, \tag*{(47)}

with m=⌈(k+1)/2⌉m=\lceil(k+1)/2\rceil and η>0\eta>0. On every fiber this equals 1+η(ρ2−1)2m−g2k+1(ρ)cos⁡θ^1+\eta(\rho^2-1)^{2m}-g_{2k+1}(\rho)\cos\hat{\theta}, which matches 1−cos⁡θ^1-\cos\hat{\theta} on the physical fiber through order δk\delta^k. Since 4m≥2k+14m\ge2k+1, the radial penalty dominates the counterterm at large radius, so the ground-diabat well is bounded from below. Its strength η\eta is a free parameter that leaves the physical fiber unchanged and only sets the radius beyond which the potential is bounded. Because the radius is conserved, the penalty does not change the radial distribution of the state. In the fiber propagations of Appendix A we set k=1k=1 and m=1/4m=1/4. For the excited diabatic state we replace 1+cos⁡θ^1+\cos\hat{\theta} by I^+η(ρ^2−I^)2m+C^2k+1/R\hat{I}+\eta(\hat{\rho}^2-\hat{I})^{2m}+\hat{C}_{2k+1}/R, the same operator with the sign of C^2k+1\hat{C}_{2k+1} reversed, which is bounded for the same reason. Like the redistribution over modes, the counterterms leave the molecular dynamics on the physical fiber unchanged. Their effect on the squeezing and the occupation is quantified in Sec. VI B and their gate implementation in Sec. VI C.

Hahn-Stock dynamics and finite-resource benchmarks

Local approximations and compact dynamics

Before working with the bosonic encodings of the previous sections, we benchmark the continuum molecular dynamics across four Hamiltonian descriptions to isolate two distinct physical effects: the breakdown of local potential expansions during large-amplitude motion and the dynamical signatures of compact torsional topology. We then show that adding Fock levels to a direct qumode encoding cannot recover the compact dynamics, while the compact encodings of Sec. III do.

In Figure 3 we compare the time evolution of a wavepacket initialized in the cis basin under the linear vibronic coupling (LVC) model, the quadratic vibronic coupling (QVC) model, the full trigonometric potential on the noncompact line R\mathbb{R} and the physical rotor Hamiltonian on the compact circle S1\mathrm{S}^{1}. The torsional coordinate in the LVC model decouples from the electronic and vibrational degrees of freedom, as Eq. (15) implies. Consequently, the wavepacket exhibits weak electronic-vibrational oscillations while remaining near the cis geometry.

Figure 3. Hahn-Stock dynamics for local and global descriptions of the torsion. (a) Electronic population P1P_1. (b) Return probability R(t)\mathcal{R}(t) of Eq. (21). (c) Conditional lower-adiabatic trans probability Y−=P(trans,S0)/P(S0)Y_- = P(\mathrm{trans}, S_0)/P(S_0), with the adiabatic projector of each model’s own potential. All curves start from the same packet in the cis basin and the line readout folds probabilities over cells. The LVC and QVC controls use Eqs. (15) and (16), with QVC shown through 200 fs (Sec. IV A). The trans readout assigns half weight to the discrete basin-boundary nodes.

The QVC model restores the negative torsional curvature on the excited diabat, capturing the initial acceleration away from the cis region and the first passage through the crossing region. At later times, its inverted excited-state parabola drives the wavepacket outward without bound. QVC then departs from the full-potential calculations after about 100 fs. Because the QVC Hamiltonian is unbounded from below along the torsion, the escaping packet keeps gaining kinetic energy and spreads exponentially in time, so the number of grid points or Fock levels needed to follow it grows exponentially with the propagation time. We therefore stop QVC at 200 fs. Longer times would add no information, since by then QVC has already departed from the full potential. We propagate Eq. (16) on a wide position grid, so the deviation of QVC from the compact calculation reflects the model error alone.

Comparing the full trigonometric potential on R\mathbb{R} and S1\mathrm{S}^{1} isolates the role of the global topology in the time evolution. Because the initial wavepacket is tightly localized (σθ,0≪π\sigma_{\theta,0} \ll\pi) and both manifolds share identical local forces, the line reference accurately tracks the physical circle dynamics at early times. The two trajectories diverge once counter-propagating branches interfere: the circle retains the cross terms between amplitudes in different angular cells, while the folded line does not. Eq. (5) makes this explicit at the wavefunction level. A packet launched near the top of the excited torsional potential spreads in both angular directions. On the circle the two branches meet after traversing different paths and interfere, while on the line their copies occupy distinct cells. With all potential terms held fixed, the electronic population on the line stays within 0.01 of the circle until about 256 fs before diverging, reaching a difference of 0.10 near 400 fs, ten times the tolerance. Similarly, the conditional adiabatic product fraction Y−Y_{-} differs by up to 0.15. The return probability shows the topology most directly. On the line the copies of the packet move apart in different cells and R(t)\mathcal{R}(t) decays to zero after about 250 fs. On the circle the two branches come back to the Franck-Condon region and R(t)\mathcal{R}(t) revives to about 0.09 by 420 fs. The line reference

is numerically converged and verified across two independent implementations to far below these topological discrepancies (Appendix A).

Cutoff effects in direct qumode encodings

The simplest way to place the torsion on a qumode is to identify the angle with a quadrature and keep the full cosine potential. We call this a direct encoding. The quadrature is not periodic, so a direct encoding simulates the line and not the circle. A finite register adds a cutoff error on top of it. Here and below, NN denotes the number of Fock states {∣0⟩,…,∣N−1⟩}\{|0\rangle,\ldots,|N-1\rangle\} retained per physical qumode. These are distinct from the rotor levels ∣ℓ⟩|\ell\rangle of the encoded torsion.

In Figure 4 we compare each finite register with the line and circle references through the maximum population error ϵP1\epsilon_{P_1}. The line reference is the limit N→∞N \to\infty of the register, so the error against it measures the truncation alone. This error decreases steadily with the cutoff and reaches 3×10−43 \times10^{-4} at N=2047N = 2047. The error against the circle instead approaches the largest difference between the line and circle references, 0.10, shown by the dash-dotted line, and reaches it at N=2047N = 2047. At intermediate cutoffs it falls below this value because the truncation error and the difference between line and circle have opposite signs at late times and partly cancel. More Fock levels therefore only converge the simulation to the unwrapped line. Recovering the physical photoisomerization dynamics requires encoding the compact S1\mathrm{S}^{1} topology directly.

Figure 4. Fock cutoff of a direct qumode encoding of the torsion in Hahn-Stock dynamics. We show the largest error in P1P_1 over 420 fs when the noncompact trigonometric Hamiltonian is truncated to NN Fock levels at the qumode scale s=0.32s = 0.32 (Appendix A). The green line compares each register with the untruncated noncompact theory and shows how the noncompact theory converges onto itself as NN grows. The dashed blue line compares each register with the compact rotor and tends to the dash-dotted line, the largest gap between the noncompact and compact theories. The dotted line marks the 0.01 tolerance.

Finite energy and Fock truncation

We now check that the two compact encodings reproduce the circle with finite resources. A physical implementation has a finite fiber width, set by the squeezing or the constraint strength, and a finite Fock cutoff per qumode. With a single register per encoded mode, at reff=2.1r_{\mathrm{eff}} = 2.1 and qumode scale s=0.6s = 0.6 for the GKP comb and at A=192A = 192 and R=20R = 20 for the circle, both encodings reproduce P1P_1, Q−Q_{-}, Y−Y_{-} and the return probability R\mathcal{R} within 0.01 through 416.5 fs. For the circle the error in Y−Y_{-} includes a term for the probability missed by the readout (Appendix A). These single registers need large cutoffs and strong squeezing. In Sec. VI we show how spreading the encoding over several qumodes and adding counterterms bring them within a budget of 13 dB of squeezing and about 100 Fock levels per mode.

PSB3: a three-coordinate molecular application

To demonstrate how our hybrid framework scales to multi-coordinate reaction pathways, we apply it to protonated Schiff-base model PSB3 of the retinal chromophore [19, 20, 73]. It shares with Hahn-Stock the same three types of degrees of freedom: electronic states (qubits), noncompact vibrational coordinates (oscillators) and a compact torsion (a rotor). The encodings of Sec. III therefore carry over directly. However, PSB3 introduces three complications: the torsion couples simultaneously to two distinct vibrational modes, the potential depends on the torsion through more than one cosine term and the global minimum of the lower adiabatic surface S0S_0 lies at the trans conformer rather than the cis reactant.

Coordinates and full molecular Hamiltonian

In addition to the central torsion θ^\hat{\theta}, the model includes a bond-length alternation (BLA) coordinate r^\hat{r} and a hydrogen-out-of-plane (HOOP) coordinate ϕ^\hat{\phi}. The electronic coupling and the diagonal potentials depend on the combination τ^=θ^−ϕ^/2\hat{\tau} = \hat{\theta} - \hat{\phi}/2, so torsional motion changes the forces acting on the other coordinates. We use the refined two-state potential of Refs. [19, 20, 81]. The compact molecular Hamiltonian is

H^PSB3=T^I^2+V^PSB3(r^,θ^,ϕ^),(48)\hat{H}_{\mathrm{PSB3}} = \hat{T}\hat{I}_2 + \hat{V}_{\mathrm{PSB3}}(\hat{r},\hat{\theta},\hat{\phi}), \tag*{(48)}
T^=12(Grrp^r2+GθθL^2+Gϕϕp^ϕ2)=−12(Grr∂r2+Gθθ∂θ2+Gϕϕ∂ϕ2),(49)\hat{T} = \frac{1}{2}\left(G_{rr}\hat{p}_r^2 + G_{\theta\theta}\hat{L}^2 + G_{\phi\phi}\hat{p}_{\phi}^2\right) = -\frac{1}{2}\left(G_{rr}\partial_r^2 + G_{\theta\theta}\partial_{\theta}^2 + G_{\phi\phi}\partial_{\phi}^2\right), \tag*{(49)}

where L^=−i∂θ\hat{L} = -i\partial_\theta acts on periodic wavefunctions while p^r=−i∂r\hat{p}_r = -i\partial_r and p^ϕ=−i∂ϕ\hat{p}_\phi= -i\partial_\phi act on the two noncompact coordinates. The coordinate-space matrix of V^PSB3\hat{V}_{\mathrm{PSB3}} is denoted by V(r,θ,ϕ)V(r,\theta,\phi) below. We use the diagonal kinetic approximation of Ref. [73], with coefficients listed in Table I of Appendix B. Reference [20] also considers a torsion-HOOP kinetic cross term. Here the diagonal metric and the two noncompact coordinate ranges are held fixed across the potential and topology comparisons.

To express the potential operator V^PSB3\hat{V}_{\mathrm{PSB3}}, we denote the BLA coordinate in angstroms by rAr_A, while everywhere else rr is in bohr. We define the Morse and quadratic radial functions

MP(rA)=D[1−e−d1(rA−rTS)]2,M_P(r_A)=D\left[1-e^{-d_1(r_A-r_{\mathrm{TS}})}\right]^2,
MM(rA)=D[1−e−d4(rA−rmin⁡)]2,M_M(r_A)=D\left[1-e^{-d_4(r_A-r_{\min})}\right]^2,
C(rA)=c1rA2+c2rA+c3.(50)C(r_A)=c_1r_A^2+c_2r_A+c_3. \tag*{(50)}

In the diabatic basis {∣0⟩,∣1⟩}\{\lvert0\rangle,\lvert1\rangle\}, we write the matrix elements of V^PSB3\hat{V}_{\mathrm{PSB3}} directly in Fourier components of τ^\hat{\tau},

V^00=A0(r^,ϕ^)+d32cos⁡τ^+A2(r^)cos⁡2τ^,\hat{V}_{00}=A_0(\hat{r},\hat{\phi})+\frac{d_3}{2}\cos\hat{\tau}+A_2(\hat{r})\cos2\hat{\tau},
V^11=F0(r^,ϕ^)+F2(r^)cos⁡2τ^,\hat{V}_{11}=F_0(\hat{r},\hat{\phi})+F_2(\hat{r})\cos2\hat{\tau},
V^01=k1sin⁡2τ^,(51)\hat{V}_{01}=k_1\sin2\hat{\tau}, \tag*{(51)}

with

A0=MP+d2+MM+d32+h1sin⁡2(ϕ/4),A_0=\frac{M_P+d_2+M_M+d_3}{2}+h_1\sin^2(\phi/4),
A2=MM−MP−d22,A_2=\frac{M_M-M_P-d_2}{2},
F0=(1+c52)C+c42+h1sin⁡2(ϕ/4),F_0=\left(1+\frac{c_5}{2}\right)C+\frac{c_4}{2}+h_1\sin^2(\phi/4),
F2=c4−c5C2.(52)F_2=\frac{c_4-c_5C}{2}. \tag*{(52)}

with the parameters of Table 1. In this representation, the non-Gaussian trigonometric interactions act on linear combinations of the torsional and HOOP coordinates, such as cos⁡(2θ^−ϕ^)\cos(2\hat{\theta}-\hat{\phi}) and cos⁡(θ^−ϕ^/2)\cos(\hat{\theta}-\hat{\phi}/2). While the Morse functions are exponentials of r^\hat{r}, which no native gate can implement, the wavepacket explores a BLA interval short compared to the Morse decay lengths d1−1,d4−1d_1^{-1},d_4^{-1}. For gate synthesis, these radial factors can therefore be represented by a low-order power series in r^\hat{r}. Alternatively, one can use a trigonometric expansion with the multimode Fourier synthesis of Ref. [53]. In all numerical calculations in this section we retain the exact Morse potentials.

ParameterValueUnitParameterValueUnit
DD2000kcal mol−1^{-1}c1c_1437.068kcal mol−1^{-1} Å−2^{-2}
rTSr_{\mathrm{TS}}0.025079167Åc2c_216.7317kcal mol−1^{-1} Å−1^{-1}
rmin⁡r_{\min}0.0912615Åc3c_37.35468kcal mol−1^{-1}
d1d_13733.5/(2D)\sqrt{3733.5/(2D)}Å−1^{-1}c4c_488.517kcal mol−1^{-1}
d4d_41097.19/(2D)\sqrt{1097.19/(2D)}Å−1^{-1}c5c_55.952211
d2d_254.633kcal mol−1^{-1}h1h_1155.749kcal mol−1^{-1}
d3d_33.8008kcal mol−1^{-1}k1k_124.0358kcal mol−1^{-1}
Kinetic coefficients (atomic units)
GrrG_{rr}7.981×10−57.981 \times10^{-5}GθθG_{\theta\theta}2.599×10−52.599 \times10^{-5}GϕϕG_{\phi\phi} 40.375×10−540.375 \times10^{-5}

Table 1. PSB3 model parameters.

We show both adiabatic surfaces, S0S_0 and S1S_1, at ϕ=0\phi=0 over the full periodic domain θ∈[−π,π]\theta\in[-\pi,\pi] in Figure 5(a). Solving V01(r,θ,0)=0V_{01}(r,\theta,0)=0 and V00(r,θ,0)=V11(r,θ,0)V_{00}(r,\theta,0)=V_{11}(r,\theta,0) locates the two conical intersections at θ=±π/2\theta=\pm\pi/2 and r≃0.05139r\simeq0.05139 bohr, which enable nonadiabatic branching between the cis (θ=0\theta=0) and trans (θ=±π\theta=\pm\pi) basins.

Figure 5. Molecular geometry and dynamics of the three-coordinate PSB3 model. (a) Lower and upper adiabatic surfaces at ϕ=0\phi= 0, evaluated over the full period −π≤θ≤π-\pi\le\theta\le\pi and −0.20≤r≤0.30-0.20 \le r \le0.30 bohr. The markers labeled CI are the two conical intersections. (b) Conditional lower-adiabatic trans probability Y−=P(trans,S0)/P(S0)Y_- = P(\mathrm{trans},S_0)/P(S_0). (c) Late return probability to the initial molecular state. The physical circle dynamics are compared with an unwrapped five-cell line reference using the same periodic potential. The QVC model expands the torsional trigonometric factors through second order and preserves the HOOP dependence.

Cis preparation and comparison Hamiltonians

Unlike in the Hahn-Stock model, the global minimum of the PSB3 ground adiabatic surface S0S_0 lies in the trans basin, while the cis reactant occupies a higher-energy metastable well. Both minima lie at rA=rmin⁡r_A=r_{\min} and ϕ=0\phi=0. The cis minimum lies d3=0.165d_3=0.165 eV above the trans one. Along the minimum-energy torsional path, the two wells are separated by a barrier at θ=π/2\theta=\pi/2 that lies 2 eV above the cis minimum and keeps the cis conformer metastable.

However, in the eye, interactions with the surrounding protein keep retinal, of which PSB3 is a reduced model, in the cis form until a photon arrives. The photon then rotates the molecule to trans, and this motion starts the chain of events that produces a nerve signal [2, 7]. The cis well is therefore the physically relevant starting point.

Since the ground state is localized in the trans region, we isolate the cis-localized state that occurs in retinal by adding a confining wall

Wcis(θ)=0.50min⁡[((∣θ∣−θw)+π−θw)4,1],(53)W_{\mathrm{cis}}(\theta)=0.50\min\left[\left(\frac{(|\theta|-\theta_w)_+}{\pi-\theta_w}\right)^4,1\right], \tag*{(53)}

Hartree, with θw=0.48π\theta_w=0.48\pi and (z)+=max⁡(z,0)(z)_+=\max(z,0). This function vanishes throughout the cis well, raises the trans basin out of reach and defines the multiplication operator W^cis\hat{W}_{\mathrm{cis}}.

If ∣u−(q)⟩\lvert u_-(q)\rangle and ∣u+(q)⟩\lvert u_+(q)\rangle are the local electronic eigenvectors of V(q)V(q), a sudden vertical photoexcitation transfers the ground packet from the lower to the upper adiabatic surface at every geometry and prepares the initial wavefunction

ΨFC(q)=ND^FC(q)Ψg(q),\Psi_{\mathrm{FC}}(q)=\mathcal{N}\hat{D}_{\mathrm{FC}}(q)\Psi_g(q),
D^FC(q)=∣u+(q)⟩⟨u−(q)∣,q=(r,θ,ϕ).(54)\hat{D}_{\mathrm{FC}}(q)=\lvert u_+(q)\rangle\langle u_-(q)\rvert,\qquad q=(r,\theta,\phi). \tag*{(54)}

The wall is then removed and ΨFC\Psi_{\mathrm{FC}} is propagated under H^PSB3\hat{H}_{\mathrm{PSB3}}. Preparing this state on hardware is a nontrivial task which lies outside the scope of this work. Adaptive variational methods acting natively on qumodes, such as CV-ADAPT-VQE [82], suggest a possible route toward preparing the initial ground packet.

We benchmark the exact compact dynamics on S1\mathrm{S}^1 against two approximations sharing the same initial state ΨFC\Psi_{\mathrm{FC}} and diagonal kinetic operator T^\hat{T}:

  1. The quadratic vibronic coupling (QVC) model, the PSB3 counterpart of Eq. (16). We Taylor-expand the potential matrix V(r,θ,ϕ)V(r,\theta,\phi) to second order in θ\theta around θ=0\theta=0 while keeping the full nonpolynomial dependence on rr and ϕ\phi intact:

VQVC(r,θ,ϕ)=V(r,0,ϕ)+θ ∂θV∣θ=0+θ22 ∂θ2V∣θ=0.(55)V_{\mathrm{QVC}}(r,\theta,\phi)=V(r,0,\phi)+\theta\,\partial_\theta V\big|_{\theta=0}+\frac{\theta^2}{2}\,\partial_\theta^2V\big|_{\theta=0}. \tag*{(55)}

This isolates the breakdown of local torsional approximations without introducing BLA or HOOP truncation errors.

  1. A noncompact trigonometric line. We unwrap θ\theta onto R\mathbb{R} using a converged five-cell periodic box of length 10π10\pi, folding the resulting adiabatic probability densities modulo 2π2\pi onto [−π,π][-\pi,\pi] (with half weight assigned to the basin-boundary nodes at ∣θ∣=π/2|\theta|=\pi/2).

Dynamics and finite-resource results

In Figure 5(b) and Figure 5(c) we compare these three descriptions for the same cis-localized preparation over a time window of 200 fs. Just as in the Hahn-Stock model, the QVC model VQVCV_{\mathrm{QVC}} misses the reaction, with a product fraction that differs from the exact conditional product fraction Y−(t)Y_{-}(t) by up to 0.64. Over the 200 fs window, the noncompact line R\mathbb{R} and the compact circle S1\mathbb{S}^{1} differ by at most 0.016 in Y−(t)Y_{-}(t) and 0.019 in the adiabatic S1S_{1} population.

Compactness gives a larger signal in the late return probability R(t)\mathcal{R}(t), where the line and circle differ by 0.10 near 180 fs. The return and product observables thus probe different aspects of the motion. Conditional-state overlaps between winding paths (Eq. (5)) can resolve how the additional coordinates shape this difference.

Finally, we verify that our compact bosonic encodings accurately capture the three-mode PSB3 dynamics under finite resources. The finite-energy GKP comb at reff=1.5r_{\mathrm{eff}}=1.5 holds the product fraction within 0.003 of the compact dynamics through 200 fs. To test Fock-space truncation in the presence of all three coupled modes, we retain 3072 Fock levels in the torsional register and propagate through the first 70 fs. We read out after a normalized projection onto the 256 GKP teeth, which discards at most 0.0074 of the probability. The encoded trajectory then tracks the exact S1S_{1} reference to within 0.004 in the adiabatic S1S_{1} population and 0.008 in Y−(t)Y_{-}(t). This comparison of the product fraction uses a strict basin boundary. Extending the truncated dynamics into the late-time recurrence regime is the natural next resource test for this model. We leave this to future work.

Hardware resources and experimental prospects

In this section, we determine the qumode and gate resources required to implement the compact encodings of Sec. III on hybrid quantum hardware. While the single-register benchmarks of Sec. IV require large Fock cutoffs and high squeezing, distributing the encoding across multiple qumodes and incorporating Hamiltonian counterterms or postselection lowers the required squeezing and per-mode occupation. The numbers below are estimates for stated assumptions and initial conditions, and we use them to explain the logic of the resource count rather than to fix a hardware target.

Qumode frames, localization and occupation

For the GKP encoding with the unwrapped qumode angle θ^R=sq^\hat{\theta}_{\mathrm{R}}=s\hat{q} and L^R=p^/s\hat{L}_{\mathrm{R}}=\hat{p}/s, the occupation of a fixed encoded state is exactly

⟨n^⟩s(t)=12[Aθ(t)s2+s2Bθ(t)−1],Aθ=⟨θ^R2⟩,Bθ=⟨L^R2⟩.(56)\langle\hat{n}\rangle_{s}(t)=\frac{1}{2}\left[\frac{A_{\theta}(t)}{s^{2}}+s^{2}B_{\theta}(t)-1\right], \qquad A_{\theta}=\langle\hat{\theta}_{\mathrm{R}}^{2}\rangle,\quad B_{\theta}=\langle\hat{L}_{\mathrm{R}}^{2}\rangle. \tag*{(56)}

Balancing the two terms gives the optimal sopt4=Aθ/Bθs_{\mathrm{opt}}^{4}=A_{\theta}/B_{\theta} and the occupation AθBθ−1/2\sqrt{A_{\theta}B_{\theta}}-1/2. Because a device uses one choice of ss for the whole trajectory, the register capacity is dictated by the largest occupation over the time window. Furthermore, an upper bound rphys≤rmax⁡r_{\mathrm{phys}}\le r_{\max} on the physical squeezing of the teeth imposes, through Eq. (31),

s≥smin⁡=2π ereff−rmax⁡.(57)s\ge s_{\min}=\sqrt{2\pi}\,e^{r_{\mathrm{eff}}-r_{\max}}. \tag*{(57)}

Less squeezing at fixed tooth width σν\sigma_{\nu} of Eq. (30) therefore requires a larger ss and pushes the momentum teeth p=sℓp=s\ell farther from the origin. Once smin⁡s_{\min} exceeds sopts_{\mathrm{opt}}, this raises the boson occupation through the term s2Bθs^{2}B_{\theta} of Eq. (56).

For the two-qumode circle encoding at radius RR and physical radial width σr≃R/(4A)\sigma_{r}\simeq R/(4A), we obtain the mean occupation number

⟨n^0+n^1⟩≃R22+18σr2+⟨J^2⟩2R2−1.(58)\langle\hat{n}_{0}+\hat{n}_{1}\rangle\simeq\frac{R^{2}}{2}+\frac{1}{8\sigma_{r}^{2}}+\frac{\langle\hat{J}^{2}\rangle}{2R^{2}}-1. \tag*{(58)}

where the three positive terms correspond to the radial displacement, the radial momentum and the angular momentum. A visually narrow annulus can therefore be expensive in the mean occupation number. At fixed radial width the first and third terms compete. Minimizing over R2R^{2} gives the photon-optimal radius R2=⟨J^2⟩R^{2}=\sqrt{\langle\hat{J}^{2}\rangle} and the occupation ⟨J^2⟩+1/(8σr2)−1\sqrt{\langle\hat{J}^{2}\rangle}+1/(8\sigma_{r}^{2})-1. The radius cannot always take this value, because the dynamics constrains the fractional width σρ=σr/R\sigma_{\rho}=\sigma_{r}/R, which sets the spread of the fiber label. A tolerance σρ≤σρmax⁡\sigma_{\rho}\le\sigma_{\rho}^{\max} requires R≥σr/σρmax⁡R\ge\sigma_{r}/\sigma_{\rho}^{\max}, so the radius is

R=max⁡(⟨J^2⟩1/4, σr/σρmax⁡).(59)R=\max\left(\langle\hat{J}^{2}\rangle^{1/4},\ \sigma_{r}/\sigma_{\rho}^{\max}\right). \tag*{(59)}

Relaxing σρmax⁡\sigma_{\rho}^{\max}, as the counterterms of Sec. III D do, lets the circle reach its photon-optimal radius (Figure 6).

Figure 6. Minimum torsion qumodes for a given number of Fock levels per qumode at a 13 dB squeezing cap, with fiber tolerances that hold P1P_1 within 0.03 through 420 fs, without (solid) and with (dashed) the counterterms (ct) of Sec. III D. At 101 levels (dotted line) the GKP comb needs 80 qumodes for about 5300 photons and the kinetic counterterm reduces this to 12 qumodes for about 770 photons. The circle needs a radius R=14.2R = 14.2 without the counterterm and R=8.45R = 8.45 with it, and in both cases fits on its two native qumodes.

Trading qumode count for qumode occupation

Even at the optimal frame a single qumode may hold hundreds of photons for the trajectories of Sec. IV. The first freedom of Sec. III C is the number of physical modes we use. We can organize each encoded coordinate in a collective mode a^coll=k−1/2∑ja^j\hat{a}_{\mathrm{coll}}=k^{-1/2}\sum_j\hat{a}_j of kk qumodes, with all orthogonal modes in vacuum. The wrapped encoding uses one such mode and the circle a pair, so the circle uses 2k2k physical modes. This is simply a passive mixing that conserves the total occupation. If the collective mode has number distribution pnp_n, each of its photons lands in a given physical mode with probability 1/k1/k, so that mode holds mm photons with probability

qk(m)=∑n=m∞pn(nm)k−m(1−k−1)n−m,(60)q_k(m)=\sum_{n=m}^{\infty}p_n\binom{n}{m}k^{-m}(1-k^{-1})^{n-m}, \tag*{(60)}

also when the collective mode is entangled with the molecule. The same binomial splitting underlies multiplexed photon-number-resolving detection [83, 84]. Its mean is nˉ/k\bar{n}/k for a collective mean nˉ\bar{n}. Sharing a tail budget ϵ\epsilon among the kk marginals, so that each mode misses at most ϵ/k\epsilon/k, the Fock levels needed per mode follow for a narrow collective distribution from the normal approximation

N−1≃nˉk+znˉk(1−1k),(61)N-1 \simeq\frac{\bar{n}}{k}+z\sqrt{\frac{\bar{n}}{k}\left(1-\frac{1}{k}\right)}, \tag*{(61)}

with zz the upper-tail quantile of ϵ/k\epsilon/k. The exact mixture of Eq. (60) applies when the distribution is broad.

To illustrate the resources we require the electronic population P1P_1, the direct qubit readout, to stay within 0.03 of the compact dynamics through 420 fs for both encodings. We take 13 dB of squeezing and 101 Fock levels per qumode as an example budget. We choose an odd number of Fock states so that the zero eigenvalue for both quadratures exists.

At the single-register operating points of Sec. IV C the mean occupation peaks at 837 for the GKP comb along its 420 fs trajectory and at 406 in total for the circle, 203 in each circular mode. Under the 13 dB cap the occupation is instead set by the fiber tolerance each encoding must

meet. Figure 6 shows how many qumodes each encoding then needs for a given number of Fock levels per qumode, without and with its counterterm.

The GKP curves of Figure 6 follow from Eqs. (56) and (57) at the fiber tolerances of the comb without and with the kinetic counterterm (Appendix A). They use the moments of the propagated GKP trajectory of Sec. IV C with one frame for the whole window and are split over modes by Eq. (61). The circle curves are analytic displaced-squeezed distributions at a 13 dB cap, with the radii fixed by the fiber tolerances of Appendix A and the angular term of Eq. (58) at ⟨J^2⟩=ℓE2/2\langle\hat{J}^{2}\rangle= \ell_{E}^{2}/2, the energetic bound, which exceeds the largest ⟨L^2⟩≃2840\langle\hat{L}^{2}\rangle\simeq2840 of that trajectory. Every plotted cutoff keeps the probability outside the truncated register below 10−310^{-3}, shared equally among its modes. These curves are occupation estimates for the stated tolerance, squeezing cap, initial state and time window. They change with these assumptions. A truncated register conserves the fiber label only approximately, so we propagate the truncated two-mode circle register with its counterterm. At 101 levels per mode it holds P1P_{1} within the 0.03 target through 420 fs.

We can also always postselect the readout on the conserved fiber label. Keeping only the shots near the physical fiber removes most of the finite-width bias, so a broader fiber distribution, with less squeezing and fewer photons, reaches the same accuracy at the cost of repeated shots.

Redistribution leaves the tooth width, the radial resolution and the total occupation unchanged, so it replaces one demanding qumode with several less occupied ones at equal mean loss exposure. The gate constructions below act directly on the distributed coordinates and preserve this advantage during the nonlinear evolution.

Gate decomposition of the hybrid time evolution

We now construct the time evolution explicitly for the Hahn-Stock Hamiltonian. We use the hybrid gate set of [45]: qubit rotations Ri(ϕ)=eiϕσ^i/2R_i(\phi)=e^{i\phi\hat{\sigma}_i/2}, the qumode rotation R(ϕ)=eiϕn^R(\phi)=e^{i\phi\hat{n}}, the displacement D(z)=eza^†−z∗a^D(z)=e^{z\hat{a}^{\dagger}-z^{*}\hat{a}}, the beam splitter BS⁡(z)=eza^†b^−z∗a^b^†\operatorname{BS}(z)=e^{z\hat{a}^{\dagger}\hat{b}-z^{*}\hat{a}\hat{b}^{\dagger}}, the Kerr and cross-Kerr phases K(ϕ)=eiϕn^2K(\phi)=e^{i\phi\hat{n}^{2}} and CK⁡(ϕ)=eiϕn^an^b\operatorname{CK}(\phi)=e^{i\phi\hat{n}_{a}\hat{n}_{b}} and the conditioned displacement

CD⁡(z)=eσ^z(za^†−z∗a^),(62)\operatorname{CD}(z)=e^{\hat{\sigma}_{z}(z\hat{a}^{\dagger}-z^{*}\hat{a})}, \tag*{(62)}

with z∈Cz \in\mathbb{C} and ϕ∈R\phi\in\mathbb{R}. We add the quadratic phase P(κ)=eiκq^2/2P(\kappa)=e^{i\kappa\hat{q}^{2}/2}, its conditional form CP⁡(κ)=eiκσ^zq^2/2\operatorname{CP}(\kappa)=e^{i\kappa\hat{\sigma}_{z}\hat{q}^{2}/2} [32] and the conditional cosine gate Cos⁡(γ)=e−iγcos⁡(sq^)⊗σ^z\operatorname{Cos}(\gamma)=e^{-i\gamma\cos(s\hat{q})\otimes\hat{\sigma}_{z}} of [52, 53, 54]. Their pulse realizations depend on the platform [41, 49, 54, 55].

To expose the required blocks, write H^=K^+D^+M^\hat{H}=\hat{K}+\hat{D}+\hat{M}, where

K^=G2L^2,\hat{K}=\frac{G}{2}\hat{L}^{2},
D^=ω(n^x+I^/2)+E1I^−σ^z2+(c0I^+czσ^z)(I^−cos⁡θ^),\hat{D}=\omega(\hat{n}_{x}+\hat{I}/2)+E_{1}\frac{\hat{I}-\hat{\sigma}_{z}}{2}+(c_{0}\hat{I}+c_{z}\hat{\sigma}_{z})(\hat{I}-\cos\hat{\theta}),
M^=x^(κ2I^−κ2σ^z+λσ^x),(63)\hat{M}=\hat{x}\left(\frac{\kappa}{2}\hat{I}-\frac{\kappa}{2}\hat{\sigma}_{z}+\lambda\hat{\sigma}_{x}\right), \tag*{(63)}

with c0=(W0−W1)/4c_{0}=(W_{0}-W_{1})/4 and cz=(W0+W1)/4c_{z}=(W_{0}+W_{1})/4. The kinetic and coupling blocks act on different factors and commute, so a symmetric second-order step is

S^2(Δt)=e−iK^Δt/2e−iM^Δt/2e−iD^Δte−iM^Δt/2e−iK^Δt/2.(64)\hat{S}_{2}(\Delta t)=e^{-i\hat{K}\Delta t/2}e^{-i\hat{M}\Delta t/2}e^{-i\hat{D}\Delta t}e^{-i\hat{M}\Delta t/2}e^{-i\hat{K}\Delta t/2}. \tag*{(64)}

The half steps of K^+M^\hat{K}+\hat{M} merge between consecutive steps. We draw one step for each register in Figure 7 and identify its gates block by block, with τ\tau the duration of each factor.

Quantum circuit implementations for wrapped encoding, all-qubit, and circle embedding

Figure 7. One time step of Eq. (64) as a circuit, with the kinetic and coupling blocks first and the potential block second. Thin and thick lines represent qubits and qumodes, respectively, with labels indicating their degree of freedom. A dot marks the qubit whose σ^z\hat{\sigma}_z conditions the gate. Gates follow Eqs. (62) to (71). (a) Wrapped encoding: three physical lines and one ancilla. The kinetic block is the momentum quadratic phase PpP_p, the coupling block is Eq. (65), and the potential block contains the vibration rotation RR, the qubit rotation RzR_z, and the two conditional cosine gates of Eq. (66). On kk qumodes the conditional displacements of the cosine gates are repeated on each mode and PpP_p splits into the one- and two-mode phases of Eq. (68). (b) All-qubit register of Sec. VII with eight torsion and four vibration qubits. The quantum Fourier transform (QFT) brings the torsion to momentum for the kinetic phases, the basis change B^x\hat{B}_x brings the vibration to position for the coupling block and the uniformly controlled rotation UCRz\mathrm{UCR}_z applies both cosine phases. (c) Circle embedding on its native pair: four physical lines and no ancilla. Kerr and cross-Kerr phases give the kinetic block of Eq. (70). The passive transformation BS maps the circular modes ±\pm to the Cartesian modes 0 and 1, which the two torsion lines carry until BS†\mathrm{BS}^{\dagger} maps them back. There, the completed potential acts through the gates of Eq. (71) and its Gaussian scalar part.

Two gates of e−iτD^e^{-i\tau\hat{D}} are common to every register. The harmonic term is the qumode rotation R(−ωτ)R(-\omega\tau) of the vibration and the constant part is a qubit rotation RzR_{z}, up to a global phase. Its angle is τ(E1−2cz)\tau(E_{1}-2c_{z}) in the wrapped and qubit registers and τ(E1−2cz+c0)\tau(E_{1}-2c_{z}+c_{0}) for the circle, whose completed potentials carry part of the constant. The coupling block is exactly reducible to linear operations. With v=λ2+(κ/2)2v=\sqrt{\lambda^{2}+(\kappa/2)^{2}}, sin⁡β=λ/v\sin\beta=\lambda/v and cos⁡β=−κ/(2v)\cos\beta=-\kappa/(2v),

e−iτM^=D(−iτκ22)Ry(−β)CD⁡(−iτv2)Ry(β),(65)e^{-i\tau\hat{M}}=D\left(-\frac{i\tau\kappa}{2\sqrt{2}}\right)R_{y}(-\beta)\operatorname{CD}\left(-\frac{i\tau v}{\sqrt{2}}\right)R_{y}(\beta), \tag*{(65)}

one displacement of the vibration, one conditional displacement and two qubit rotations in every hybrid register. The remaining factors are the cosine phases eiτc0cos⁡θ^e^{i\tau c_{0}\cos\hat{\theta}} and eiτczσ^zcos⁡θ^e^{i\tau c_{z}\hat{\sigma}_{z}\cos\hat{\theta}} and the kinetic factor e−iτK^e^{-i\tau\hat{K}}, which depend on the encoding.

In the wrapped encoding cos⁡θ^→cos⁡(sq^)\cos\hat{\theta}\to\cos(s\hat{q}), so the two cosine phases are conditional cosine gates,

eiτczσ^zcos⁡θ^=Cos⁡(−τcz)e,eiτc0cos⁡θ^=Cos⁡(−τc0)a,(66)e^{i\tau c_{z}\hat{\sigma}_{z}\cos\hat{\theta}}=\operatorname{Cos}(-\tau c_{z})_{e},\qquad e^{i\tau c_{0}\cos\hat{\theta}}=\operatorname{Cos}(-\tau c_{0})_{a}, \tag*{(66)}

where the subscript names the gate qubit. The electronic qubit ee provides σ^z\hat{\sigma}_{z} for the first. An ancilla aa prepared in ∣0⟩\lvert0\rangle provides it for the second, which then acts as a scalar phase and leaves the ancilla in ∣0⟩\lvert0\rangle up to the synthesis error, so one ancilla serves every step. The construction with a single gate qubit [53, 54] realizes Cos⁡(γ)\operatorname{Cos}(\gamma) as a sequence of ncn_{c} first-order steps, with 2nc+12n_{c}+1 conditional displacements and 2nc2n_{c} rotations of the gate qubit. Already the lowest order nc=1n_{c}=1, three conditional displacements and two rotations, reproduces the gate up to O(γ2)\mathcal{O}(\gamma^{2}), and the sequence converges to the exact gate as ncn_{c} grows. We count gates at nc=1n_{c}=1. The kinetic factor is a quadratic phase in momentum, e−iτGp^2/(2s2)=R(π/2)P(−τG/s2)R(−π/2)e^{-i\tau G\hat{p}^{2}/(2s^{2})}=R(\pi/2)P(-\tau G/s^{2})R(-\pi/2), written PpP_{p} in Figure 7.

For a collective coordinate every conditional displacement factorizes exactly

e−iβσ^zx^=∏j=1kCD⁡(−iβ2k)q^j,(67)e^{-i\beta\hat{\sigma}_{z}\hat{x}}=\prod_{j=1}^{k}\operatorname{CD}\left(-\frac{i\beta}{\sqrt{2k}}\right)_{\hat{q}_{j}}, \tag*{(67)}

where the subscript names the mode acted on. A construction with MCDM_{\mathrm{CD}} displacements therefore uses kMCDkM_{\mathrm{CD}} after distribution, with the same approximation order. The kinetic block also factorizes,

G2s2P^2=G2ks2(∑jp^j2+2∑i<jp^ip^j).(68)\frac{G}{2s^{2}}\hat{P}^{2}=\frac{G}{2ks^{2}}\left(\sum_{j}\hat{p}_{j}^{2}+2\sum_{i<j}\hat{p}_{i}\hat{p}_{j}\right). \tag*{(68)}

into kk one-mode and (k2)\binom{k}{2} two-mode commuting Gaussian phases. The kinetic counterterm of Eq. (43) replaces this phase by e−iτGf3(L^R)2/2e^{-i\tau G f_3(\hat{L}_R)^2/2}, with

f3(L^R)2=L^R2−L^Rsin⁡(2πL^R)π+I^−cos⁡(4πL^R)8π2.(69)f_3(\hat{L}_R)^2=\hat{L}_R^2-\frac{\hat{L}_R\sin(2\pi\hat{L}_R)}{\pi}+\frac{\hat{I}-\cos(4\pi\hat{L}_R)}{8\pi^2}. \tag*{(69)}

The first term is the Gaussian phase above and the last is a cosine gate on the momentum quadrature. Only the middle term mixes polynomial and trigonometric functions. It can be synthesized from the same native gates by universality [49, 52]. The simplest construction is a group commutator of a conditional displacement and a conditional sine gate along the momentum quadrature, which reproduces it to leading order in the gate angles. We do not include these gates in the counts.

For the circle in its two circular modes the kinetic unitary is exactly

e−iτGJ^2/2=K(−τG2)+K(−τG2)−CK⁡(τG)+−.(70)e^{-i\tau G\hat{J}^2/2}=K\left(-\frac{\tau G}{2}\right)_{+}K\left(-\frac{\tau G}{2}\right)_{-}\operatorname{CK}(\tau G)_{+-}. \tag*{(70)}

On 2k2k distributed modes the total angular momentum ∑j(n^j,+−n^j,−)\sum_j(\hat{n}_{j,+}-\hat{n}_{j,-}) acts as J^\hat{J} on the encoded sector, so its square requires 2k2k Kerr and (2k2)\binom{2k}{2} cross-Kerr phases. The completed potential of Eq. (35) acts on the Cartesian modes, which the passive transformation BS=R(−π/2)−BS(π/4)\mathrm{BS}=\mathrm{R}(-\pi/2)_{-}\mathrm{BS}(\pi/4), a 50:50 beam splitter followed by a quarter rotation of one mode, obtains from the circular ones by inverting Eq. (35). With each diabat completed about its own minimum, the potential block is q^02+q^122R2(czI^+c0σ^z)−q^0R(c0I^+czσ^z)\frac{\hat{q}_0^2+\hat{q}_1^2}{2R^2}(c_z\hat{I}+c_0\hat{\sigma}_z)-\frac{\hat{q}_0}{R}(c_0\hat{I}+c_z\hat{\sigma}_z) up to constants. Its σ^z\hat{\sigma}_z part gives

CD⁡(iτcz2R)0CP⁡(−τc0R2)0CP⁡(−τc0R2)1,(71)\operatorname{CD}\left(\frac{i\tau c_z}{\sqrt{2R}}\right)_0\operatorname{CP}\left(-\frac{\tau c_0}{R^2}\right)_0\operatorname{CP}\left(-\frac{\tau c_0}{R^2}\right)_1, \tag*{(71)}

while its scalar part gives the Gaussian gates D⁡(iτc0/(2R))0P⁡(−τcz/R2)0P⁡(−τcz/R2)1\operatorname{D}(i\tau c_0/(\sqrt{2}R))_0\operatorname{P}(-\tau c_z/R^2)_0\operatorname{P}(-\tau c_z/R^2)_1 and the constants join RzR_z. The radial counterterm of Eq. (47) adds cubic and quartic position phases, and polarization identities reduce them to one-mode phases in Gaussian-transformed coordinates.

Per step of Eq. (64), with the half steps of consecutive steps merged, these constructions give the following counts of qubit-qumode operations. The wrapped encoding on one qumode uses seven conditional displacements, one in the coupling block and three in each of the two cosine gates, with one Gaussian shear and one ancilla qubit. On kk qumodes the cosine displacements are repeated on each mode, which gives 6k+16k+1 conditional displacements and (k2)\binom{k}{2} two-mode Gaussian phases, 73 and 66 on 12 qumodes. The circle on its native pair uses four conditioned operations and one cross-Kerr phase with the completed potential, or eleven with the counterterm. Its potential acts on the Cartesian modes and its kinetic phases on the circular modes, so each step adds two passive transformations between them. The counterterm further adds the Gaussian coordinate changes of its polarization.

Gaussian coordinate changes of its polarization. Single-mode non-Gaussian phases are not counted. These are the two Kerr phases of the circle and, with the counterterm, the seven unconditioned cubic and quartic phases of its scalar part. Conditioned displacements are demonstrated primitives, while conditioned quadratic, cubic and quartic phases require a platform-dependent synthesis. Each counted operation is compiled into native gates with a sequence length set by the platform and the target precision.

The two compact encodings offer complementary experimental tradeoffs. With its counterterm, the circle encoding requires only two qumodes and modest squeezing, making it well suited for platforms with native Kerr and higher-order polynomial interactions. Conversely, the GKP encoding is built exclusively from conditional displacements and the demonstrated trigonometric gates [54, 55], at the cost of distributing the unwrapped angular spread across many qumodes.

All-qubit construction

A fair comparison with qubits should retain the structure of the molecular Hamiltonian, since encoding and synthesis choices can substantially change the cost of bosonic and vibronic problems [28, 76, 77]. We follow the split-operator approach to chemical dynamics [24, 26], with a Fourier grid of dθ=2mθd_\theta=2^{m_\theta} points for the torsion and dx=2mxd_x=2^{m_x} Fock levels for the coupling vibration.

The kinetic phase is quadratic in the binary angular momentum and costs (mθ2)\binom{m_\theta}{2} controlled phases. An exact quantum Fourier transform pair, QFT and QFT†^\dagger, between momentum and angle costs mθ(mθ−1)m_\theta(m_\theta-1) more [85], with two CNOTs per controlled phase. On the angular grid the cosine factorizes as cos⁡θ^=Z^θ⊗F^\cos\hat{\theta}=\hat{Z}_\theta\otimes\hat{F}, because a shift by π\pi flips the top bit and the sign. A uniformly controlled rotation UCR⁡z\operatorname{UCR}_z on the top bit then implements the scalar phase with dθ/2d_\theta/2 CNOTs [86]. A second one between two CNOTs from the electronic qubit gives the electronic phase, so the cosine block costs dθ+2d_\theta+2 CNOTs. The coupling block is diagonal in the eigenbasis of x^\hat{x} and costs 2dx2d_x CNOTs, together with a basis change B^x\hat{B}_x and its inverse. The basis changes recur at every step because the harmonic term is diagonal in the Fock basis instead. That term is linear in the binary vibrational number and the constant part acts on the electronic qubit, so both are one-qubit rotations RzR_z. For rr steps, with adjacent half steps merged,

Ccircuit≤r[dθ+2+2dx+3mθ(mθ−1)+2C(B^x)]+3mθ(mθ−1)+2dx+2C(B^x).(72)C_{\mathrm{circuit}}\le r\left[d_\theta+2+2d_x+3m_\theta(m_\theta-1)+2\mathcal{C}(\hat{B}_x)\right] +3m_\theta(m_\theta-1)+2d_x+2\mathcal{C}(\hat{B}_x). \tag*{(72)}

In our case, eight torsional qubits suffice to hold the angular momenta of the reference calculations and four qubits hold 16 vibrational levels, so the register has thirteen qubits including the electronic one. Therefore, the bound in Eq. (72) with (mθ,mx)=(8,4)(m_\theta,m_x)=(8,4) gives 658r+400658r+400 CNOTs. Merging diagonal operations or arithmetic phase evaluation can lower this baseline, while routing on hardware with limited connectivity raises it. One step thus costs 658 CNOTs on qubits, or 1058 for an isolated step. The circle uses four conditioned operations and one cross-Kerr phase per step, or eleven with its counterterm. The wrapped encoding uses seven conditional displacements on a single qumode and, distributed over the twelve qumodes of Figure 6, 73 conditional displacements and the 66 two-mode Gaussian phases of Eq. (68).

The two counts rest on very different assumptions. The qubit count assumes error-free logical gates and omits rotation synthesis, routing and error correction, and other qubit encodings might prove more suitable for realistic hardware constraints [28]. On the other hand, the qumode count treats conditioned polynomial, Kerr and cross-Kerr phases as single gates.

The more robust difference is how the two costs grow. On qubits, the gate count is directly related to the cutoff. The cosine block grows linearly with the number of grid points dθd_\theta, a general basis change of the vibration grows quadratically with dxd_x, and a generic function of several coordinates grows with the product of their grid sizes. Arithmetic evaluation of the phases reduces this growth at the price of ancillas and a larger constant. For qumodes, the gate parameters do not depend on the cutoff, so a larger Fock range changes only the hardware requirement but not the circuit. When the potential is a sum of few-body terms of bounded polynomial and trigonometric degree, each additional nuclear coordinate adds one qumode, each bilinear coupling adds one Gaussian gate and polynomial terms of fixed degree add a number of gates polynomial in the number of modes, each compiled with a sequence whose length is set by the degree and the target precision. The kinetic, harmonic and coupling blocks are native or exact for qumodes, while the qubit register pays for them through Fourier transforms, multiplexed rotations and basis changes at every step. Calibrated gate durations and errors on a given platform turn these counts into run times, a natural next step.

Conclusions and outlook

Quantum simulations of photochemical reactions have largely focused on local expansions around equilibrium geometries, leaving large-amplitude reactive coordinates, such as molecular torsions, outside of their natural scope. Photoisomerization reactions strongly depend on both the global anharmonicity of the potential as well as the global topology of the nuclear configuration space. As a result, treating the reaction coordinate via a local vibronic expansion, a noncompact trigonometric expansion on the line, or on the physical circle leads to qualitatively different dynamics. In this work, we have developed a hybrid qubit-qumode framework that maps each molecular degree of freedom to its native hardware counterpart: electronic states to qubits, vibrations to quantum oscillators and the periodic torsion to a compactly encoded bosonic register.

The two-mode Hahn-Stock model of retinal photoisomerization makes this distinction especially clear. While linear vibronic approximations can trap the wavepacket near the cis well, restoring the full trigonometric potential allows the wavepacket to accelerate towards the conical-intersection region and branch into the two adiabatic surfaces. Over early timescales, the unwrapped line provides a useful approximation, as the wavepacket explores only the first angular period, but at later times counter-propagating branches that follow different winding paths recombine and interfere on the circle, while on the unwrapped line they remain separated. Under the same potential, the compact topology alters the late-time electronic populations, return probabilities and adiabatic cis/trans branching fractions by 10–15% after the first angular period.

Extending this analysis to the more complicated three-mode PSB3 model shows how the physics changes as we increase the number of nuclear coordinates. Compactness affects the product fraction less than in Hahn-Stock, but the late return dynamics remains sensitive to it, so the accuracy of the noncompact approximation strongly depends on the observable of interest. Correlations with the other nuclear coordinates can change how the winding paths overlap and interfere, showing the need for a diverse and complementary set of molecular observables to probe the dynamical role of global nuclear geometry.

To realize the compact torsion on naturally noncompact bosonic hardware, we analyzed and benchmarked two compact qumode encodings: a wrapped Gottesman-Kitaev-Preskill stabilizer code and a two-qumode constrained-circle embedding. Both split the Hilbert space into conserved sectors, labeled by the fractional momentum for the GKP comb and by the radius for the circle. On realistic hardware the state spreads over a narrow distribution of these sectors, which gradually dephases the encoded dynamics. We introduced two independent and scalable methods that reduce the hardware requirements of both encodings without changing the simulated dynamics. Distributing an encoded coordinate across several qumodes lowers the boson occupation of each mode, at the price of more couplings during the time evolution. Hamiltonian counterterms that leave the physical sector unchanged reduce the dependence of the encoded dynamics on the sector label, which relaxes the required squeezing and occupation.

Compiled into elementary gates, the qumode constructions are compact. The kinetic, harmonic and vibronic coupling terms are native or exact on qumodes and their gate count does not depend on the Fock cutoff, while a qubit register pays for the same terms with a gate count that grows with the number of grid points. This structure favors qumodes for potentials built from few-body terms of bounded polynomial and trigonometric degree, the structure of the vibronic coupling models widely used in photochemistry and of the torsional harmonics of PSB3.

Recent advances in trapped-ion and superconducting platforms, including trigonometric gates and conditional displacements on the qumodes, make these photoisomerization simulations promising targets for near-term experiments. A dedicated hardware study will refine these estimates with platform-specific gate errors and noise. A natural first hardware milestone is to prepare a localized cis wavepacket, follow its nonadiabatic evolution through a conical intersection and resolve the changes produced by the full potential and the compact encoding. The electronic population alone carries this signal, so a direct qubit readout is enough. The molecular benchmarks developed here set the accuracy needed to distinguish these effects. The circle embedding with its radial counterterms runs on its native pair of qumodes with modest squeezing, which makes it a first possible target for such an experiment. The wrapped encoding uses only demonstrated gate types and distributes its occupation over more qumodes as the torsion completes full turns.

Progress toward utility in quantum chemistry will require extending this connection between molecular observables and physical resources. Additional reactive coordinates, couplings to an environment and potentials informed by electronic-structure calculations are natural next steps. They will allow studies of how nuclear correlations and environmental relaxation alter photochemical branching and recurrences. Comparison with established classical dynamics methods can identify which regimes benefit most from retaining the coupled quantum motion. By treating compact reaction coordinates within the same qubit-qumode description as electronic transitions and noncompact vibrations, the present framework provides the compact coordinates that quantum simulations of larger photochemical systems require.

Acknowledgments

We thank Zohreh Davoudi, Matt Grau, Norbert Linke, and Peter Love for helpful discussions. F.H. is supported by the Austrian Science Fund (FWF), project 10.55776/J4854. T.R. and F.R. are supported by the U.S. Department of Energy, Office of Science, Office of Nuclear Physics, Early Career Program under contract No. DE-SC0025881. This research used resources of the National Energy Research Scientific Computing Center (NERSC), a U.S. Department of Energy Office of Science User Facility, under NERSC award NP-ERCAP0037787. We thank Stony Brook Research Computing and Cyber-infrastructure and the Institute for Advanced Computational Science at Stony Brook University for access to SeaWulf. The system was funded by National Science Foundation award 1531492 and Major Instrumentation award 2215987, with matching funds from Empire State Development’s Division of Science, Technology and Innovation (NYSTAR) under contract C210148.

Numerical conventions and convergence

All reference calculations represent each degree of freedom in a basis with a cutoff, a Fourier basis for the torsion and a coordinate or harmonic-oscillator basis for the coupling vibration. Every cutoff is increased until the reported observables are converged well below the simulation tolerance. On position grids we propagate with the split-operator method. The kinetic factor is applied exactly in momentum space by fast Fourier transforms and the potential factor, a 2×22 \times2 matrix in the electronic states at each grid point, is exponentiated exactly point by point. In a Fock or angular-momentum basis we instead use a Chebyshev expansion of the propagator. We compare trajectories at identical sample times, since P1P_1 and the return probability oscillate on the 1.7 fs electronic period and interpolation between samples would add spurious differences.

The noncompact line is represented by joining KK copies of the angular cell into one periodic box of circumference 2πK2\pi K, with KK large enough that the outer cells remain empty over the reported times. Two independent implementations agree far below the tolerance. In the basin readouts, boundary nodes receive half weight.

For the circle encoding the radial measure is ρ dρ\rho\,\mathrm{d}\rho and finite-width reference dynamics are assembled from radial sectors. The reference ansatz fA(ρ)ψmol(θ,x)f_A(\rho)\psi_{\mathrm{mol}}(\theta,x) gives every angular harmonic the same radial profile. A physical planar state vanishes smoothly at the origin, which this ansatz satisfies to numerical precision because the prepared annulus carries no resolvable probability there.

The fiber tolerances used in Sec. VI B follow from exact propagation of single fibers of the circle and the comb, Eqs. (36) and (27), with and without the counterterms of Eqs. (47) and (43), with each circle diabatic completed about its own minimum. A Gaussian fiber distribution of width σ\sigma weights the recorded trajectories, wrapped onto −1/2≤ν<1/2-1/2 \leq\nu< 1/2 for the comb. The tolerance is the largest σ\sigma for which P1P_1 stays within the target of its physical-fiber value, σν≤0.082\sigma_\nu\leq0.082 and 0.22 for the comb without and with the kinetic counterterm and 1.1% and 5.5% for the circle without and with the radial counterterm. The same propagations also record Q−Q_-, Y−Y_- and the return. Holding these within the target as well narrows the deterministic tolerances to σν≤0.068\sigma_\nu\leq0.068 and 0.19 for the comb and 0.12% and 4.3% for the circle. The most sensitive observable is Y−Y_- for the comb and the coherent return for the bare circle.

The finite-cutoff benchmarks propagate the projected Hamiltonian PH^P\mathcal{P}\hat{H}\mathcal{P}, with P\mathcal{P} the projector onto the retained Fock states, after projecting and normalizing the initial state. Projecting an operator and applying a function to the projected operator are different operations. With PN\mathcal{P}_N the projector onto ∣0⟩,…,∣N−1⟩\lvert0\rangle,\ldots,\lvert N-1\rangle of one mode, for example,

PNp^2PN=(PNp^PN)2+N2∣N−1⟩⟨N−1∣.(73)\mathcal{P}_N\hat{p}^{2}\mathcal{P}_N=(\mathcal{P}_N\hat{p}\mathcal{P}_N)^2+\frac{N}{2}\lvert N-1\rangle\langle N-1\rvert. \tag*{(73)}

In all Fock-basis calculations we truncate the Hamiltonian itself. We compute its matrix elements in the full Fock basis and keep the block of the retained levels. The circle benchmark propagates the projected polynomial Hamiltonian with independent cutoffs on n+n_+ and n−n_-. These runs use the linear substitution cos⁡θ^→q0/R\cos\hat{\theta}\rightarrow q_0/R, which coincides with the completed form of Eq. (35) on the constraint surface. We check the completed form fiber by fiber. At A=192A=192 these fiber propagations give a finite-width error in P1P_1 of 0.0030 for the linear substitution and 0.0017 for the completion.

The electronic populations and return probabilities are evaluated in the evolution basis, while the adiabatic and basin observables require a coordinate-space quadrature. If that readout misses probability δ\delta, with captured lower-surface and trans probabilities L−L_{-} and T−T_{-}, then

T−L−+δ≤Y−≤T−+δL−+δ.(74)\frac{T_{-}}{L_{-}+\delta} \le Y_{-} \le\frac{T_{-}+\delta}{L_{-}+\delta}. \tag*{(74)}

These intervals give the bound on the circle error in Y−Y_{-} used in Sec. IV C.

PSB3 parameters

Table I lists the parameters of the PSB3 model of § 2.4 A. The radial functions of Eq. (50) take the BLA coordinate rAr_{\mathrm{A}} in angstroms and the potential energies are converted from kcal/mol to Hartree by the factor 1/627.5094740631 before they are combined with the kinetic operator.

References

References

  1. [1]W. Domcke and D. R. Yarkony, Role of Conical Intersections in Molecular Spectroscopy and Photoinduced Chemical Dynamics, Annu. Rev. Phys. Chem. 63, 325 (2012).DOI
  2. [2]D. Polli, P. Altoè, O. Weingart, K. M. Spillane, C. Manzoni, D. Brida, G. Tomasello, G. Orlandi, P. Kukura, R. A. Mathies, M. Garavelli, and G. Cerullo, Conical intersection dynamics of the primary photoisomerization event in vision, Nature 467, 440 (2010).
  3. [3]T. Schultz, J. Quenneville, B. Levine, A. Toniolo, T. J. Martínez, S. Lochbrunner, M. Schmitt, J. P. Shaffer, M. Z. Zgierski, and A. Stolow, Mechanism and dynamics of azobenzene photoisomerization, Journal of the American Chemical Society 125, 8098 (2003).
  4. [4]D. R. Yarkony, Diabocal conical intersections, Rev. Mod. Phys. 68, 985 (1996).
  5. [5]G. A. Worth and L. S. Cederbaum, Beyond Born-Oppenheimer: Molecular Dynamics Through a Conical Intersection, Annu. Rev. Phys. Chem. 55, 127 (2004).
  6. [6]B. G. Levine and T. J. Martínez, Isomerization through conical intersections, Annual Review of Physical Chemistry 58, 613 (2007).pubmed.ncbi.nlm.nih.gov/17291184
  7. [7]R. W. Schoenlein, L. A. Peteanu, R. A. Mathies, and C. V. Shank, The first step in vision: femtosecond isomerization of rhodopsin, Science 254, 412 (1991).pubmed.ncbi.nlm.nih.gov/1925597
  8. [8]J. C. Tully, Molecular dynamics with electronic transitions, J. Chem. Phys. 93, 1061 (1990).DOI
  9. [9]R. Crespo-Otero and M. Barbatti, Recent Advances and Perspectives on Nonadiabatic Mixed Quantum-Classical Dynamics, Chem. Rev. 118, 7026 (2018).
  10. [10]M. H. Beck, A. Jäckle, G. A. Worth, and H.-D. Meyer, The multiconfiguration time-dependent Hartree (MCTDH) method: a highly efficient algorithm for propagating wavepackets, Phys. Rep. 324, 1 (2000).
  11. [11]O. Vendrell and H.-D. Meyer, Multilayer multiconfiguration time-dependent Hartree method: Implementation and applications to a Henon–Heiles Hamiltonian and to pyrazine, J. Chem. Phys. 134, 044135 (2011), arXiv:1012.4625.
  12. [12]S. M. Greene and V. S. Batista, Tensor-Train Split-Operator Fourier Transform (TT-SOFT) Method: Multidimensional Nonadiabatic Quantum Dynamics, J. Chem. Theory Comput. 13, 4034 (2017).
  13. [13]A. Baiardi and M. Reiher, Large-Scale Quantum Dynamics with Matrix Product States, J. Chem. Theory Comput. 15, 3481 (2019), arXiv:1903.10622.DOI
  14. [14]X. Xie, Y. Liu, Y. Yao, U. Schollwöck, C. Liu, and H. Ma, Time-dependent density matrix renormalization group quantum dynamics for realistic chemical systems, J. Chem. Phys. 151, 224101 (2019), arXiv:1908.10588.
  15. [15]H. Köppel, W. Domcke, and L. S. Cederbaum, Multimode Molecular Dynamics Beyond the Born-Oppenheimer Approximation, Adv. Chem. Phys. 57, 59 (1984).
  16. [16]L. L. E. Cigrang et al., Roadmap for Molecular Benchmarks in Nonadiabatic Dynamics, J. Phys. Chem. A 129, 7023 (2025), arXiv:2502.14569 [physics.chem-ph].DOI
  17. [17]S. Hahn and G. Stock, Quantum-Mechanical Modeling of the Femtosecond Isomerization in Rhodopsin, J. Phys. Chem. B 104, 1146 (2000).DOI
  18. [18]S. Hahn and G. Stock, Femtosecond secondary emission arising from the nonadiabatic photoisomerization in rhodopsin, Chem. Phys. 259, 297 (2000).DOI
  19. [19]E. Marsili, M. H. Farag, X. Yang, L. De Vico, and M. Olivucci, Two-State, Three-Mode Parametrization of the Force Field of a Retinal Chromophore Model, J. Phys. Chem. A 123, 1710 (2019).DOI
  20. [20]E. Marsili, M. Olivucci, D. Lauvergnat, and F. Agostini, Quantum and Quantum-Classical Studies of the Photoisomerization of a Retinal Chromophore Model, J. Chem. Theory Comput. 16, 6032 (2020).DOI
  21. [21]I. Kassal, J. D. Whitfield, A. Perdomo-Ortiz, M.-H. Yung, and A. Aspuru-Guzik, Simulating Chemistry Using Quantum Computers, Annu. Rev. Phys. Chem. 62, 185 (2011), arXiv:1007.2648 [quant-ph].DOI
  22. [22]P. J. Ollitrault, G. Mazzola, and I. Tavernelli, Nonadiabatic Molecular Quantum Dynamics with Quantum Computers, Phys. Rev. Lett. 125, 260511 (2020), arXiv:2006.09405 [quant-ph].DOI
  23. [23]A. B. Magann, M. D. Grace, H. A. Rabitz, and M. Sarovar, Digital quantum simulation of molecular dynamics and control, Phys. Rev. Res. 3, 023165 (2021), arXiv:2002.12497 [quant-ph].DOI
  24. [24]C. Zalka, Simulating quantum systems on a quantum computer, Proc. R. Soc. Lond. A 454, 313 (1998), arXiv:quant-ph/9603026 [quant-ph].
  25. [25]E. C. Eklund, A. Tikku, P. Sinnott, W. J. Huggins, G. H. Low, D. W. Berry, and I. Kassal, End-to-End Simulation of Chemical Dynamics on a Quantum Computer (2026), arXiv:2603.19007 [quant-ph].
  26. [26]I. Kassal, S. P. Jordan, P. J. Love, M. Mohseni, and A. Aspuru-Guzik, Polynomial-time quantum algorithm for the simulation of chemical dynamics, Proc. Natl. Acad. Sci. U.S.A. 105, 18681 (2008), arXiv:0801.2986 [quant-ph].DOI
  27. [27]A. Macridin, P. Spentzouris, J. Amundson, and R. Harnik, Electron-Phonon Systems on a Universal Quantum Computer, Phys. Rev. Lett. 121, 110504 (2018), arXiv:1802.07347 [quant-ph].DOI
  28. [28]N. P. Sawaya, T. Menke, T. H. Kyaw, S. Johri, A. Aspuru-Guzik, and G. G. Guerreschi, Resource-efficient digital quantum simulation of d-level systems for photonic, vibrational, and spin-s Hamiltonians, npj Quantum Inf. 6, 49 (2020), arXiv:1909.12847 [quant-ph].DOI
  29. [29]P. J. Ollitrault, A. Baiardi, M. Reiher, and I. Tavernelli, Hardware efficient quantum algorithms for vibrational structure calculations, Chem. Sci. 11, 6842 (2020), arXiv:2003.12578 [quant-ph].
  30. [30]S. G. Mehendale, I. Loaiza, S. Malpathak, S. Fomichev, D. Motlagh, and A. F. Izmaylov, Efficient Block Encodings of Vibrational Hamiltonians (2026), arXiv:2610.33800 [quant-ph].
  31. [31]S. L. Braunstein and P. van Loock, Quantum information with continuous variables, Rev. Mod. Phys. 77, 513 (2005), arXiv:quant-ph/0410100 [quant-ph].DOI
  32. [32]C. Weedbrook, S. Pirandola, R. García-Patrón, N. J. Cerf, T. C. Ralph, J. H. Shapiro, and S. Lloyd, Gaussian quantum information, Rev. Mod. Phys. 84, 621 (2012), arXiv:1110.3234 [quant-ph].DOI
  33. [33]R. J. MacDonell, C. E. Dickerson, C. J. T. Birch, A. Kumar, C. L. Edmunds, M. J. Biercuk, C. Hempel, and I. Kassal, Analog quantum simulation of chemical dynamics, Chem. Sci. 12, 9794 (2021), arXiv:2012.01852 [quant-ph].
  34. [34]J. Huh, G. G. Guerreschi, B. Peropadre, J. R. McClean, and A. Aspuru-Guzik, Boson sampling for molecular vibronic spectra, Nat. Photonics 9, 615 (2015), arXiv:1412.8427 [quant-ph].DOI
  35. [35]Y. Shen, Y. Lu, K. Zhang, J. Zhang, S. Zhang, J. Huh, and K. Kim, Quantum optical emulation of molecular vibronic spectroscopy using a trapped-ion device, Chem. Sci. 9, 836 (2018), arXiv:1702.04859 [quant-ph].
  36. [36]C. Sparrow et al., Simulating the vibrational quantum dynamics of molecules using photonics, Nature 557, 660 (2018).DOI
  37. [37]S. Malpathak, S. D. Kallullathil, and A. F. Izmaylov, Simulating Vibrational Dynamics on Bosonic Quantum Devices, J. Phys. Chem. Lett. 16, 1855 (2025), arXiv:2411.17950 [quant-ph].DOI
  38. [38]R. J. MacDonell, T. Navickas, T. F. Wohlers-Reichel, C. H. Valahu, A. D. Rao, M. J. Millican, M. A. Currington, M. J. Biercuk, T. R. Tan, C. Hempel, and I. Kassal, Predicting molecular vibronic spectra using time-domain analog quantum simulation, Chem. Sci. 14, 9439 (2023), arXiv:2209.06558 [quant-ph].
  39. [39]C. K. Law and J. H. Eberly, Arbitrary Control of a Quantum Electromagnetic Field, Phys. Rev. Lett. 76, 1055 (1996).pubmed.ncbi.nlm.nih.gov/10061622
  40. [40]S. Krastanov, V. V. Albert, C. Shen, C.-L. Zou, R. W. Heeres, B. Vlastakis, R. J. Schoelkopf, and L. Jiang, Universal control of an oscillator with dispersive coupling to a qubit, Phys. Rev. A 92, 040303 (2015), arXiv:1502.08015 [quant-ph].DOI
  41. [41]A. Eickbusch, V. Sivak, A. Z. Ding, S. S. Elder, S. R. Jha, J. Venkatraman, B. Royer, S. Girvin, R. J. Schoelkopf, and M. H. Devoret, Fast universal control of an oscillator with weak dispersive coupling to a qubit, Nat. Phys. 18, 1464 (2022), arXiv:2111.06414 [quant-ph].DOI
  42. [42]A. L. Shaw, P. Scholl, R. Finkelstein, R. B.-S. Tsai, J. Choi, and M. Endres, Erasure cooling, control, and hyperentanglement of motion in optical tweezers, Science 388, 845 (2025), arXiv:2311.15580 [quant-ph].
  43. [43]J. Hwang, T. Xu, S. R. Muleady, S. K. P. Hampel, G. Lubin, D. P. Hewatt, C. A. Regal, and A. M. Rey, Hybrid Qubit-Oscillator Module from Motional States of Two Interacting Atoms, Phys. Rev. Lett. 137, 113403 (2026), arXiv:2512.06429 [quant-ph].DOI
  44. [44]D. Leibfried, R. Blatt, C. Monroe, and D. Wineland, Quantum dynamics of single trapped ions, Rev. Mod. Phys. 75, 281 (2003).DOI
  45. [45]J. Y. Araz, M. Grau, J. Montgomery, and F. Ringer, Hybrid quantum simulations with qubits and qumodes on trapped-ion platforms, Phys. Rev. A 112, 012620 (2025), arXiv:2410.07346 [quant-ph].DOI
  46. [46]J. Whitlow, Z. Jia, Y. Wang, C. Fang, J. Kim, and K. R. Brown, Quantum simulation of conical intersections using trapped ions, Nat. Chem. 15, 1509 (2023), arXiv:2211.07319 [quant-ph].
  47. [47]C. H. Valahu, V. C. Olaya-Agudelo, R. J. MacDonell, T. Navickas, A. D. Rao, M. J. Millican, J. B. Pérez-Sánchez, J. Yuen-Zhou, M. J. Biercuk, C. Hempel, T. R. Tan, and I. Kassal, Direct observation of geometric-phase interference in dynamics around a conical intersection, Nat. Chem. 15, 1503 (2023), arXiv:2211.07320 [quant-ph].DOI
  48. [48]T. Navickas, R. J. MacDonell, C. H. Valahu, V. C. Olaya-Agudelo, F. Scuccimarra, M. J. Millican, V. G. Matsos, H. L. Nourse, A. D. Rao, M. J. Biercuk, C. Hempel, I. Kassal, and T. R. Tan, Experimental Quantum Simulation of Chemical Dynamics, J. Am. Chem. Soc. 147, 23566 (2025), arXiv:2409.04044 [quant-ph].DOI
  49. [49]S. Lloyd and S. L. Braunstein, Quantum computation over continuous variables, Phys. Rev. Lett. 82, 1784 (1999), arXiv:quant-ph/9810082 [quant-ph].DOI
  50. [50]Y. Liu et al., Hybrid Oscillator-Qubit Quantum Process: Instruction Set Architectures, Abstract Machine Models, and Applications, PRX Quantum 7, 010201 (2026), arXiv:2407.10381 [quant-ph].DOI
  51. [51]J. Hong, S. H. Kim, S. K. Min, and J. Huh, Oscillator-qubit generalized quantum signal processing: a case study of the uracil cation vibronic model, Chem. Sci. 17, 11911 (2026), arXiv:2510.10495 [quant-ph].DOI
  52. [52]T. Rainaldi, V. Ale, M. Grau, D. Kharzeev, E. Rico, F. Ringer, P. Shome, and G. Siopsis, Trigonometric continuous-variable gates and hybrid quantum simulations of the sine-Gordon model, JHEP 03, 125, arXiv:2512.19582 [quant-ph].arxiv.org/abs/2512.19582
  53. [53]T. Chalertmputisarak, K. Schwennicke, I. Kassal, and T. R. Tan, Programmable generation of arbitrary continuous-variable anharmonicities and nonlinear couplings (2025), arXiv:2511.22286 [quant-ph].
  54. [54]T. Rainaldi, J. Montgomery, C. G. Yale, B. K. McFarland, M. C. Revelle, D. Lobser, E. C. Tortorici, S. M. Clark, G. Siopsis, M. Grau, and F. Ringer, Trigonometric Continuous-Variable Quantum Gates: Realization with Trapped Ions and Nonperturbative Wigner Negativity (2026), arXiv:2607.14085 [quant-ph].arxiv.org/abs/2607.14085
  55. [55]C. McGarry, T. Chalermpusitarak, K. Schwennicke, F. Succimarra, M. J. Millican, V. G. Matsos, C. H. Valahu, P. Nagpal, H.-K. Chan, H. L. Nourse, I. Kassal, and T. R. Tan, Programmable quantum simulation of anharmonic dynamics (2026), arXiv:2603.04744 [quant-ph].arxiv.org/abs/2603.04744
  56. [56]K. Marshall, R. Pooser, G. Siopsis, and C. Weedbrook, Quantum simulation of quantum field theory using continuous variables, Phys. Rev. A 92, 063825 (2015), arXiv:1503.08121 [quant-ph].DOI
  57. [57]Z. Davoudi, N. M. Linke, and G. Pagano, Toward simulating quantum field theories with controlled phonon-ion dynamics: A hybrid analog-digital approach, Phys. Rev. Research 3, 043072 (2021), arXiv:2104.09346 [quant-ph].DOI
  58. [58]R. G. Jha, F. Ringer, G. Siopsis, and S. Thomas, Continuous-variable quantum computation of the O(3) model in 1+1 dimensions, Phys. Rev. A 109, 052412 (2024), arXiv:2310.12512 [quant-ph].DOI
  59. [59]E. Crane et al., Hybrid Oscillator–Qubit Quantum Processors: Simulating Fermions, Bosons, and Gauge Fields, Phys. Rev. X 16, 041008 (2026), arXiv:2409.03747 [quant-ph].DOI
  60. [60]R. A. Briceño, R. G. Edwards, M. Eaton, C. González-Arciniegas, O. Pfister, and G. Siopsis, Toward coherent quantum computation of scattering amplitudes with a measurement-based photonic quantum processor, Phys. Rev. Res. 6, 043065 (2024), arXiv:2312.12613 [quant-ph].arxiv.org/abs/2312.12613
  61. [61]V. Ale, N. M. Bauer, R. G. Jha, F. Ringer, and G. Siopsis, Quantum computation of SU(2) lattice gauge theory with continuous variables, JHEP 06, 084, arXiv:2410.14580 [hep-lat].DOI
  62. [62]V. Ale, T. Rainaldi, E. Rico, F. Ringer, and G. Siopsis, Simulating quantum electrodynamics in 2+1 dimensions with qubits and qumodes, JHEP 04, 122, arXiv:2511.14506 [quant-ph].arxiv.org/abs/2511.14506
  63. [63]S. Abel, M. Spannowsky, and S. Williams, Real-time scattering processes with continuous-variable quantum computers, Phys. Rev. A 112, 012614 (2025), arXiv:2502.01767 [quant-ph].DOI
  64. [64]V. Ale, T. Rainaldi, E. Rico, F. Ringer, and G. Siopsis, Encoding Compact U(1) Gauge Fields in Bosonic Modes with GKP Stabilization (2026), arXiv:2609.00167 [quant-ph].arxiv.org/abs/2609.00167
  65. [65]K. Shinbrough, D. J. Webb, I. R. Øvergaard, O. Băzăvan, S. Saner, C. J. Ballance, and R. Srinivas, Applications of Spin-Dependent Generalized Squeezing in Hybrid Spin-Oscillator Quantum Processors (2026), arXiv:2608.04197 [quant-ph].arxiv.org/abs/2608.04197
  66. [66]Q. Wang, C. G. Yale, E. C. Tortorici, B. K. McFarland, D. Lobser, M. C. Revelle, A. J. Landahl, S. M. Clark, and P. J. Love, Quantum Simulation of the Static Yukawa Model via Effective Particles on a Trapped-Ion Quantum Simulator (2026), arXiv:2610.28453 [quant-ph].arxiv.org/abs/2610.28453
  67. [67]A. T. Than, S. V. Kadam, V. Vikramaditya, N. H. Nguyen, X. Liu, Z. Davoudi, A. M. Green, and N. M. Linke, Quantum field theory dynamics on a spin-phonon quantum computer, Nature Phys. 22, 1557 (2026), arXiv:2509.11477 [quant-ph].arxiv.org/abs/2509.11477
  68. [68]D. Gottesman, A. Kitaev, and J. Preskill, Encoding a qubit in an oscillator, Phys. Rev. A 64, 012310 (2001), arXiv:quant-ph/0008040.DOI
  69. [69]L. Leidner and W. Domcke, Microscopic modelling of photoisomerization and internal-conversion dynamics, Chem. Phys. 186, 27 (1994).
  70. [70]A. Pereira, J. Knapik, A. Chen, D. Lauvergnat, and F. Agostini, Quantum molecular dynamics simulations of the effect of secondary modes on the photoisomerization of a retinal chromophore model, Eur. Phys. J. Spec. Top. 232, 1917 (2023).DOI
  71. [71]N. Lyu, M. B. Soley, and V. S. Batista, Tensor-Train Split-Operator KSL (TT-SOKSL) Method for Quantum Dynamics Simulations, J. Chem. Theory Comput. 18, 3327 (2022), arXiv:2203.00527.DOI
  72. [72]A. J. Schile and D. T. Limmer, Simulating conical intersection dynamics in the condensed phase with hybrid quantum master equations, J. Chem. Phys. 151, 014106 (2019), arXiv:1905.00029.DOI
  73. [73]M. Aarabi, E. Marsili, M. Oliviucci, D. Lauvergnat, F. Agostini, M. Garavelli, and F. Santoro, Quantum Dynamics Predicts Coherent Oscillations in the Early Times of a Biological Photoisomerization, J. Phys. Chem. Lett. 16, 8486 (2025).DOI
  74. [74]M. Reck, A. Zeilinger, H. J. Bernstein, and P. Bertani, Experimental realization of any discrete unitary operator, Phys. Rev. Lett. 73, 58 (1994).pubmed.ncbi.nlm.nih.gov/10056719
  75. [75]C. Oh, Logarithmic-depth quantum simulation of boson sampling (2026), arXiv:2609.18907 [quant-ph].arxiv.org/abs/2609.18907
  76. [76]M. S. Moeed, J. Brown, A. Ibrahim, E. V. B. De Oliveira, and P.-N. Roy, Qubit encodings for lattices of dipolar planar rotors, J. Chem. Phys. 163, 174103 (2025), arXiv:2507.17952 [quant-ph].DOI
  77. [77]H. L. Nourse, V. C. Olaya-Agudelo, and I. Kassal, Using bosons to improve resource efficiency of quantum simulation of vibronic molecular dynamics (2025), arXiv:2512.20828 [quant-ph].arxiv.org/abs/2512.20828
  78. [78]A. Raab, G. A. Worth, H.-D. Meyer, and L. S. Cederbaum, Molecular dynamics of pyrazine after excitation to the S2 electronic state using a realistic 24-mode model Hamiltonian, J. Chem. Phys. 110, 936 (1999).DOI
  79. [79]D. Motlagh, R. A. Lang, P. Jain, J. A. Campos-Gonzalez-Angulo, W. Maxwell, T. Zeng, A. Aspuru-Guzik, and J. M. Arrazola, Quantum algorithm for vibronic dynamics: case study on singlet fission solar cell design, Quantum Sci. Technol. 10, 045048 (2025), arXiv:2411.13669 [quant-ph].DOI
  80. [80]M. Kang, H. Nuomin, S. N. Chowdhury, J. L. Yuly, K. Sun, J. Whitlow, J. Valdiviezo, Z. Zhang, P. Zhang, D. N. Beratan, and K. R. Brown, Seeking a quantum advantage with trapped-ion quantum simulations of condensed-phase chemical dynamics, Nat. Rev. Chem. 8, 340 (2024), arXiv:2305.03156 [quant-ph].DOI
  81. [81]D. Lauvergnat, QuantumModelLib, Software repository, https://github.com/lauvergn/QuantumModelLib, accessed October 14, 2026.
  82. [82]D. Athanassakos, G. Tejedor-García, J. Y. Araz, M. Ramôa, B. Sambasivam, S. E. Economou, and F. Ringer, Continuous-variable ADAPT-VQE for bosonic lattice models (2026), arXiv:2606.05297 [quant-ph].
  83. [83]M. J. Fitch, B. C. Jacobs, T. B. Pittman, and J. D. Franson, Photon-number resolution using time-multiplexed single-photon detectors, Phys. Rev. A 68, 043814 (2003), arXiv:quant-ph/0305193.DOI
  84. [84]D. Achilles, C. Silberhorn, C. Śliwa, K. Banaszek, and I. A. Walmsley, Fiber-assisted detection with photon number resolution, Opt. Lett. 28, 2387 (2003), arXiv:quant-ph/0305191.DOI
  85. [85]A. Barenco, A. Ekert, K.-A. Suominen, and P. Törmä, Approximate quantum Fourier transform and decoherence, Phys. Rev. A 54, 139 (1996), arXiv:quant-ph/9601018 [86] M. Möttönen, J. J. Vartiainen, V. Bergholm, and M. M. Salomaa, Quantum Circuits for General Multiqubit Gates, Phys. Rev. Lett. 93, 130502 (2004), arXiv:quant-ph/0404089 [quant-ph].

Paper details

vlnc:W_awqem9qd
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 9

downloads · 0

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

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents