betaFeedback

WorksPhysics and Astronomy

Nonperturbative string dynamics in a (2+1)D lattice gauge theory on a superconducting quantum processor

Abstract

Electric flux strings underpin confinement in gauge theories. In two spatial dimensions, their evolution involves competing processes of charge motion, pair creation, and transverse deformation. Accessing these dynamics beyond perturbative limits challenges classical approaches, as entanglement growth restricts the times accessible to tensor-network simulations. Here we use quantum simulation to investigate a (2+1)D Z2 lattice gauge theory with dynamical matter and an explicit magnetic plaquette interaction, spanning strongly confined and nonperturbative deconfined regimes. Using 63 superconducting qubits, including 12 ancillas, we evolve initially localized strings with static or dynamical endpoint charges. Spatially resolved measurements reveal rapid charge and flux spreading in the deconfined regime, enhanced matter occupation in the string interior near the pair-creation resonance, and suppressed spreading at strong electric fields. String-path occupations further probe transverse rearrangements between configurations connected by plaquette operations. Retaining separate matter and gauge degrees of freedom makes local Gauss-law constraints directly accessible for error detection. Combining these constraints with ancilla checks and noise rescaling, we measure dynamics in circuits containing up to 15 second-order Trotter steps. Our results establish quantum simulation as a practical approach to exploring nonperturbative string dynamics in two spatial dimensions, with gauge redundancy serving as an experimental resource.

Introduction

Quantum simulation offers a route to probe real-time dynamics of gauge theories by encoding the Hamiltonian directly into controllable quantum degrees of freedom [1, 2, 3]. Most experimental progress so far has been in one spatial dimension, where models ranging from the Schwinger model to Z2\mathbb{Z}_2 and non-Abelian LGTs have been realized on trapped ions [4, 5, 6], ultracold atoms [7, 8, 9], and superconducting qubits [10, 11, 12, 13, 14, 15, 16], observing various phenomena ranging from string breaking and particle scattering to thermalization and its absence.

In two spatial dimensions (2+12+1D), the transverse spatial extent allows the flux tube to fluctuate and broaden, enabling roughening and transverse string dynamics [17, 18, 19, 20, 21] and providing a more faithful picture of confinement than one-dimensional settings. Observing these effects at experimentally accessible time scales requires an explicit magnetic plaquette term [22], which deforms the string transversally by flipping the electric fluxes around a single plaquette.

Existing 2+12+1D Z2\mathbb{Z}_2 experiments fall into two classes. A first group realizes the gauge–matter model without a plaquette term [23, 24]. A second group includes the plaquette term but works in the gauge-fixed (dual) formulation, in which the matter fields are eliminated by a local unitary and the model maps onto a toric code in external fields [25, 26]. Beyond Z2\mathbb{Z}_2, a closely related confinement and string-breaking phenomenology has been observed for a dynamically generated U(1) gauge theory on a Rydberg-atom simulator [27], as well as exactly on a trapped-ion platform [28], indicating that the string phenomenology targeted here is not specific to a single gauge group or hardware platform. Gauge fixing is economical in qubits, but it discards precisely the structure exploited here: once the Gauss-law constraints have been solved, the local symmetry generators are no longer operators that can be measured on the device. Gauge-redundant encodings have been argued to outperform gauge-fixed ones below a finite physical error threshold [29].

Here, matter and gauge degrees of freedom are represented by distinct physical qubits, and the local Gauss operators GiG_i survive as measurable stabilizers of the simulated state. The physical system we realize is a 5×45\times4 matter-site lattice occupying 63 qubits of the superconducting processor ibm_phoenix, time-evolved for up to 15 Trotter steps. It supports 32 Z2\mathbb{Z}_2 symmetries available for error detection: 20 originating from Gauss’ law and 12 from the ancillas.

We focus on the time-dependent expectation values of diagonal observables that can be measured in the same basis as the system’s symmetries. For observables that are additionally local, we show empirically that post-selection schemes based on marginal distributions, together with marginal error mitigation [14, 30] can replace global schemes. Over the time scales accessible experimentally, these incur a finite sampling overhead, which is offset by the fast 100 kHz sampling rate of the IBM Nighthawk R2 devices.

The paper is organized as follows. Section II introduces duces the model and the targeted string phenomenology, together with the digital implementation of its Trotter dynamics. Section III presents the results of quantum-quench experiments across different regimes of the static phase diagram. Section IV discusses the outlook.

Model and quantum simulation setup

Model— We consider the Z2\mathbb{Z}_2 lattice gauge theory with dynamical Ising matter, introduced by Fradkin and Shenker [31], on a two-dimensional square lattice. Matter degrees of freedom τi\tau_i reside on the vertices ii and gauge degrees of freedom σij\sigma_{ij} on the links ⟨ij⟩\langle ij\rangle; both are two-level systems, so each maps onto a single qubit. The Hamiltonian reads

H=−Jp∑pBp−hz∑⟨ij⟩σijz−hx∑⟨ij⟩τixσijxτjx−Jv∑iτiz.(1)H=-J_p\sum_p B_p-h_z\sum_{\langle ij\rangle}\sigma^z_{ij}-h_x\sum_{\langle ij\rangle}\tau^x_i\sigma^x_{ij}\tau^x_j-J_v\sum_i\tau^z_i. \tag*{(1)}

The first term is the plaquette operator Bp=∏⟨ij⟩∈pσijxB_p=\prod_{\langle ij\rangle\in p}\sigma^x_{ij}, measuring the magnetic Z2\mathbb{Z}_2 flux through an elementary plaquette, in analogy with ∮A⋅dl=∫B⋅dS\oint\mathbf{A}\cdot d\mathbf{l}=\int\mathbf{B}\cdot d\mathbf{S} in electrodynamics. The second is the electric field energy, with the two eigenstates ∣0⟩\lvert0\rangle and ∣1⟩\lvert1\rangle indicating the absence and presence of Z2\mathbb{Z}_2 electric flux. The third term describes the minimal coupling between the gauge and matter degrees of freedom. It allows charges to move between neighboring sites and charge pairs to be created or annihilated, while simultaneously updating the connecting electric flux in accordance with Gauss’s law, thereby rendering the matter dynamical. The last term gives a mass to the matter fields.

The local Z2\mathbb{Z}_2 symmetry is generated by the Gauss operators

Gi=τiz∏j∈nn(i)σijz,[H,Gi]=0 ∀i,(2)G_i=\tau^z_i\prod_{j\in\mathrm{nn}(i)}\sigma^z_{ij},\qquad[H,G_i]=0\ \forall i, \tag*{(2)}

where nn(i)\mathrm{nn}(i) enumerates the links incident on site ii. Equation (2) is the lattice analogue of Gauss’s law in electrodynamics: τiz\tau^z_i measures the Z2\mathbb{Z}_2 charge at site ii, σijz\sigma^z_{ij} the electric field on the adjacent links, and the constraint ties the product of the fields emanating from a site to the local charge. We consider sectors defined by Gi∣ψ⟩=qi∣ψ⟩G_i\lvert\psi\rangle=q_i\lvert\psi\rangle, where qi=±1q_i=\pm1 specifies the static background-charge configuration. For strings with dynamical endpoints, qi=+1q_i=+1 at every site. For strings connecting two static sources, qi=−1q_i=-1 at the source positions and qi=+1q_i=+1 elsewhere. Since [H,Gi]=0[H,G_i]=0, these sector labels are conserved under the ideal dynamics. The lattice, its interaction terms and Gauss operators are shown in Figure 1(A).

Figure 1. The 2+12+1D Z2Z_2 gauge theory on a square lattice with Ising matter. (A) The physical lattice is composed of matter qubits on the vertices and gauge qubits on the links. Additional ancilla qubits are introduced to implement the four-body plaquette term (Sec. B), resulting in a square qubit connectivity. The different highlighted regions mark the four Hamiltonian terms of Eq. (1) and the Gauss-law constraint of Eq. (2). (B) Quench experiments are realized in the different phases of the model. In the confined phase, a localized string typically survives over long time scales, while it quickly delocalizes in the deconfined regime. (C) Initial string state in the strongly confined regime Jv/hx=∞J_v/h_x=\infty with both static and dynamical endpoint charges. (D) Second-order Trotter dynamics as a depth-8 unitary evolution implementing the toric-code (TC) duality.

The phase diagram of the model was studied analytically by Fradkin and Shenker [31] and is summarized in Figure 1(B). It displays three regimes: a confined regime dominated by the electric field, a Higgs regime dominated by matter, and a deconfined regime dominated by the magnetic plaquette term. As detailed in Appendix B3, the model maps onto the toric code in external fields under a local unitary transformation that integrates out the matter fields [32, 33].

Initial string states In the strongly confined regime, Jv/hx=∞J_v/h_x=\infty, the vacuum is the classical state ∣Ω⟩=∣00…⟩\lvert\Omega\rangle=\lvert00\ldots\rangle, with no matter and no electric field (assuming a periodic lattice). The lowest-lying excitations above this classical vacuum are string-like excitations along paths γ\gamma connecting to matter endpoints (i,j)(i,j) and created by gauge-invariant operators of the form

Oγ=τixSγτjxwithSγ=∏⟨kl⟩∈γσ⟨kl⟩x(3)\mathcal{O}_{\gamma}=\tau^x_i S_{\gamma}\tau^x_j\quad\text{with}\quad S_{\gamma}=\prod_{\langle kl\rangle\in\gamma}\sigma^x_{\langle kl\rangle} \tag*{(3)}

The energy of this electric flux string configuration is

V(R)≈4μ+2λR,(4)V(R)\approx4\mu+2\lambda R, \tag*{(4)}

which is linear in the charge separation RR, so that in an infinite volume the charges cannot be separated at finite energy cost. With dynamical matter, κ≠0\kappa\ne0, the string can instead break through the creation of a new charge pair, the lattice analogue of string breaking in QCD, where the color flux tube between heavy quark sources fragments through dynamical qqˉq\bar q pair creation into color-neutral hadrons. Figure 1(C) illustrates the two string processes considered in this study.

Quench dynamics setup— We study the dynamical aspects of the confined-deconfined phase transition in the Fradkin-Shenker model by repeating quench dynamics in different regimes of the phase diagram, as presented in Table I.

hzh_zJvJ_vhz/Jvh_z/J_vhxh_xJpJ_phx/Jph_x/J_p
deconfined0.001.000.000.251.00.25
deconfined0.251.000.250.251.00.25
weakly confined0.501.000.500.251.00.25
confined1.001.001.000.251.00.25
on resonance2.001.002.000.251.00.25
strongly confined4.001.004.000.251.00.25

Table I. Parameter regimes

The quench dynamics consist of up to 15 second-order Trotter steps with a time step of dt=0.3dt=0.3, for a total evolution time t=4.5t=4.5 (in units of JvJ_v), at which point the digital evolution remains an accurate approximation to the continuous-time dynamics. Each Trotter step comprises 8 CNOT layers. For each evolution time, 10610^6 shots are collected and distributed over Pauli-twirling randomizations, which convert coherent gate errors into an effective stochastic Pauli channel without changing the logical circuit [34].

Post-selection and rescaling— Owing to hardware-level noise, the time-evolved quantum state accumulates errors in proportion to the total gate count and circuit depth. Viewed as distance-1 codes, Z2\mathbb{Z}_2 gauge theories possess error-detection capabilities that can be used to reduce the systematic bias of common observables, at the cost of an increased statistical variance. In

In addition, we apply a Clifford perturbation theory (CPT) rescaling of expectation values, based on noise factors learned by setting the non-Clifford angles to zero in the compiled circuits; this carries a sampling overhead proportional to the 1-norm of the estimated noise matrix, ∥A−1∥1\lVert A^{-1}\rVert_1 [14]. To limit the exponential sampling overhead of post-selection, the analysis is restricted to local, point-like observables and to approximations of string-like observables.

Reference calculations. — Out-of-equilibrium dynamics on a 2D square lattice are difficult to approximate classically, owing to the fast growth of the entanglement entropy [35]. For small global quenches, i.e., in the regime of large electric field, any deformation or splitting of the initial string is energetically penalized, and one can work at first order in perturbation theory in the reduced basis of minimal-length strings, for which exact diagonalization is available at the lattice sizes considered here. In the weakly confined and deconfined regimes, string elongation and splitting can no longer be neglected, and other classical approximations are required. Here we use a 2D PEPS tensor network [36] to approximate the time-evolved wavefunction. Because of the high computational cost of contracting 2D tensor networks, we work at low bond dimension and report only the time steps for which

convergence with respect to the bond dimension is satisfied, this is discussed further in Appendix § 3.4.

Results and string phenomenology

Matter sector

String with static endpoints. — Starting from the product state ∣Ω⟩=∣0…0⟩|\Omega\rangle=|0\ldots0\rangle, we prepare an electric string along the path γ\gamma shown in Figure 2(B) by applying the gauge-variant SγS_\gamma operator of Eq. (3). The resulting state contains electric flux along γ\gamma and no dynamical matter excitations. It belongs to the sector with qi=−1q_i=-1 at the two endpoints and qi=+1q_i=+1 elsewhere, corresponding to two static background charges connected by the string. The source positions remain fixed throughout the evolution. This preparation removes the charge-displacement channel present when the endpoints are dynamical, while still allowing subsequent matter-pair creation, string breaking, and motion of the created charges. The gauge–matter coupling can create dynamical charge pairs both on occupied string links and on initially empty links. The corresponding energy changes follow from the diagonal part of Eq. (1). Acting with τijxσixτijx\tau_{ij}^{x}\sigma_{i}^{x}\tau_{ij}^{x} on an occupied link whose two adjacent matter sites are empty creates two matter excitations, costing 4Jv4J_{v}, while removing one unit 2hz2h_{z} of electric flux. The net energy change is therefore ΔE=4Jv−2hz\Delta E = 4J_{v} - 2h_{z}. This process breaks the string and has a bare resonance at hz=2Jvh_{z} = 2J_{v}, irrespective of whether its endpoints are static sources or dynamical charges. By contrast, creating a pair on an empty link also creates electric flux and costs ΔE=4Jv+2hz\Delta E = 4J_{v} + 2h_{z}, which becomes energetically more expensive deeper in the confined regime.

Figure 2. Matter occupation numbers across the initial string. (A) Channels acting on the matter occupation numbers. At lowest order in perturbation theory, these channels are weighted by the inverse energy difference of the initial and final configuration. (B) Time dynamics of the total matter occupation along a gauge-violating initial string with endpoint static charges. The raw and mitigated hardware data are shown in comparison with the PEPS reference result, when converged and available. (C) Same experiment with dynamical endpoint charges. In this case, the dynamics along the string in the deconfined regime are dominated by the displacement of the initial charges.

Figure 2(B) shows the resulting total matter occupation Nγ=∑i∈γniN_{\gamma} = \sum_{i\in\gamma} n_{i} as a function of time for the six ratios hz/Jvh_{z}/J_{v} of Table I, together with the PEPS reference where the bond-dimension scan has converged. In every regime the occupation grows monotonically from zero, confirming that pair creation along the string remains active throughout, consistent with the detuning argument above: the deconfined points (hz/Jv=0h_{z}/J_{v} = 0 and 0.250.25) show the fastest matter buildup, while the strongly confined point (hz/Jv=4h_{z}/J_{v} = 4) is markedly slower. The PEPS curves are regarded as reliable only over the time window in which convergence has been checked (Sec. II). In the regime where it has converged, the PEPS prediction tracks the raw and mitigated hardware data at early times. At later times, the hardware data effectively converges to an almost fully mixed state and mitigation leads to large error bars. Overall, the comparison shows that the gauge-violating string geometry isolates the pair-creation channel cleanly, giving a monotonic, regime-dependent growth rate that is complementary to the non-monotonic, resonance-driven dynamics of the dynamical-

String with dynamical endpoints. — The same experiment with a gauge-invariant path γ\gamma with dynamical charges at its endpoints is reported in Figure 1(C), prepared by acting with the operator CγC_{\gamma} of Eq. (3) on the vacuum state of a 5×45 \times4 lattice of matter sites. In the deconfined regime, the behavior differs markedly: the initially inserted charges eventually disperse throughout the lattice, leading to a drop in the string occupation NγN_{\gamma}.

Characterizing the string dynamics when the endpoints are not fixed, as in the phenomenology outlined in Figure 3(A), requires observables defined on the full 5×45 \times4 square lattice. Panel (B) displays, for a subset of four values of the electric field hz∈[0,0.5,2,4]h_{z} \in[0, 0.5, 2, 4], the measured onsite occupation of every matter and gauge qubit of the lattice at each of these times, with color encoding the local occupation number from 0 (blue) to 1 (red); the spreading of this occupation away from the initial string directly visualizes the widening and delocalization of the initial flux tube.

Figure 3. Dynamics of a string system with dynamical endpoints. (A) The initial string on the 5×45 \times4 lattice and the main phenomenological effects occurring across the phase diagram. In delocalization, new flux is created all across the lattice; in splitting, a charge pair is created on the string, breaking it; in displacement, an endpoint charge moves to a neighboring site. (B) Mitigated onsite occupation of all matter sites and gauge links across the four parameter regimes, for the subset of time steps [0, 0.6, 1.2, 1.8, 2.4, 3.0, 3.6, 4.2]. (C) Total gauge occupation of the lattice. (D) Mean matter occupation at the two initial endpoints, at the two central sites of the initial string (interior), and at the four sites adjacent to the endpoints but off the string (outside).

Deep in the deconfined regime (hz=0h_z = 0), the string does not remain a well-defined object: additional electric flux is created, its occupation spreads rapidly over the lattice and is subject to strong oscillations until late times (Figure 3(C)). In the deep confined regime, hz=4h_z = 4, the string instead remains strongly localized, and its widening is limited to a very small extent, staying visually close to its initial shape throughout the evolution shown in panel B. The chosen initial string shape is degenerate, in the strongly confined regime, with 5 strings of equal length, which gives the string a small transversal extent at equilibrium, even in the confined phase, and hence small oscillations out of equilibrium.

At the resonance hz=2Jvh_z = 2J_v, the principal channel for matter creation is string splitting along the initial path γ\gamma, due to the energy difference scaling as 1/∣ΔEsplitting∣∝1/∣hz−2Jv∣1/\lvert\Delta E_{\mathrm{splitting}}\rvert\propto1/\lvert h_z - 2J_v\rvert. This process is suppressed away from resonance. Instead, displacement of the initial charges be-

comes dominant at hz=0h_z = 0, with a scaling ∣1/ΔEdisp∣∝1/∣hz∣\lvert1/\Delta E_{\mathrm{disp}}\rvert\propto1/\lvert h_z\rvert. To quantify this effect, we measure the matter occupation of the string’s initial endpoints, of its two most central sites, as well as of the sites outside the string that neighbor its endpoints.

nˉend=12(ni+nj),(5)\bar{n}_{\mathrm{end}} = \frac{1}{2}(n_i + n_j), \tag*{(5)}
nˉint=12∑k∈Cnk,nˉout=14∑k∈Nnk,(6)\bar{n}_{\mathrm{int}} = \frac{1}{2}\sum_{k \in\mathcal{C}} n_k,\qquad\bar{n}_{\mathrm{out}} = \frac{1}{4}\sum_{k \in\mathcal{N}} n_k, \tag*{(6)}

where i=(0,1)i = (0,1) and j=(4,2)j = (4,2) are the initial endpoints of the string, C={(2,1),(2,2)}\mathcal{C} = \{(2,1),(2,2)\} are its two central sites, and N={(0,0),(0,2),(4,1),(4,3)}\mathcal{N} = \{(0,0),(0,2),(4,1),(4,3)\} are the sites outside the string that neighbor its endpoints. The raw and mitigated hardware data for both observables are shown in Figure 3(D). The endpoint occupation nˉend\bar{n}_{\mathrm{end}} decays monotonically from its initial value of 1 in all four regimes, fastest at hz=0h_z = 0 and hz=0.5h_z = 0.5 and markedly slower at hz=2h_z = 2 and hz=4h_z = 4, while the interior occupation nˉint\bar{n}_{\mathrm{int}} grows from 0, most rapidly at the resonance hz=2h_z = 2.

This is consistent with the perturbation-theory argument above: the resonant case is characterized by fast matter creation in the interior of the string, i.e., proportional to its length, while the confined phase away from resonance is dominated by point-like effects at the endpoints. In the deconfined regime, the decay of the endpoint occupation is accompanied by growth of both nˉint\bar{n}_{\mathrm{int}} and nˉout\bar{n}_{\mathrm{out}}, which reach comparable values at late times, indicating that the dominant process is displacement of the initial charges, which spread over the lattice with no preference for the sites of the initial string. The two channels for matter creation, charge displacement and string splitting, thus leave distinct and separately measurable signatures in the local matter occupation.

Gauge sector

Delocalization and string widening.— During the quench dynamics, initial string configurations oscillate owing to the magnetic plaquette term. The amplitude of each string configuration is typically quantified with Loschmidt echoes, which are global projection operators. To keep the observables compatible with the local post-selection and rescaling scheme, we instead consider partial string-like projectors

Pγ=∏ij∈γ1−σijz2(7)P_{\gamma} = \prod_{ij\in\gamma} \frac{1-\sigma^z_{ij}}{2} \tag*{(7)}

for various paths γ\gamma connecting the initial charge locations. In the strongly confined regime, the vanishing of the initial string probability observed in Figure 3(B) is followed by an increase in the amplitudes of the other equal-length configurations, as shown in Figure 4(B). The total probability of equal-length strings instead is expected to reach a steady value that is independent of the particular choice of string configuration.

Figure 4. Partial string-like projectors, grouped according to perturbation theory. (A) Overview of minimal-length string oscillations driven by the plaquette term. (B) Measured string projector ∏k∈γnk\prod_{k\in\gamma} n_k on the links of each string γ\gamma, for the initial string (left), summed over the two strings 1 plaquette (center) and 2 plaquettes (right) away, for the four values of hzh_z.

Discussion and outlook

We have implemented the real-time dynamics of a 2+12+1D Z2\mathbb{Z}_2 lattice gauge theory with dynamical Ising matter and an explicit plaquette term on 63 qubits of a superconducting processor, retaining the gauge-redundant encoding so that the local Gauss-law generators remain measurable. The matter creation channels of the model were probed using the total matter occupation along the initial string, revealing distinct string dynamics in the deconfined, weakly confined and strongly confined regimes. In the confined regime, matter creation is mostly present at resonances and concentrated along the initial string. In the deconfined regime, we observe both matter creation and displacement of the initial charges across the full extent of the lattice, largely independently of the electric field strength. The gauge content of the theory was probed separately. In the confined regime, the electric flux was observed to remain qualitatively localized close to the subset of minimal-length strings connecting the initial charges, with only a small transverse extent at early times, whereas deep in the deconfined regime it was observed to spread rapidly across the lattice and to oscillate strongly. The oscillations of the confined string toward the other minimal-length configurations, driven by the plaquette term, were further probed quantitatively through partial string-like projectors.

Methodologically, local post-selection combined with a rescaling of hardware expectation values is found to accurately reproduce PEPS results converged up to 7 Trotter steps. At longer times, convergence of the PEPS, and of the resulting expectation values, becomes slow for bond dimensions larger than 8; this finite-bond-dimension error is especially large in the deconfined regime, i.e., for large quantum quenches, owing to the rapid growth of entanglement [35]. The rescaling approach used here instead relies on Clifford perturbation theory, and is better suited to quantum circuits with small electric field strength and correspondingly small non-Clifford Trotter angles.

Gauge redundancy—usually treated as overhead to be removed by gauge fixing—supplies local, measurement-basis-diagonal stabilizers evaluated for free from the samples already collected. Notably, the post-selection rate is found to be largely independent of the quench parameters at the present system size. Mitigation and detection are complementary: mitigation can in principle be unbiased, but its overhead grows rapidly with depth, whereas detection is cheaper and non-exponential for local constraints [37]. The circuit structure enables further extensions: mid-circuit syndrome extraction, measurement-based gauge protection via a predicted quantum Zeno transition [38], and measurement-free alternatives (energy penalties [24, 39, 40, 41, 42, 43]) are all available at no additional syndrome-extraction overhead, since the Gauss and ancilla stabilizers are already encoded into single qubits at various points of the Trotter step (Appendix D).

Pairwise Gauss-stabilizer violations can also be attributed to a single gauge-qubit bit flip with high confidence, so correcting rather than discarding such samples would extend the accessible evolution time. This remains relevant in the era of partial fault tolerance, and points toward larger gauge groups ZN\mathbb{Z}_N and richer matter content, in particular fermions.

Acknowledgments

This work was carried out at IBM Research Europe – Zurich. We thank the IBM Quantum team for access to and support with the ibm_phoenix processor. Zurich. J.C.H. acknowledges funding by the Max Planck Society, the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany’s Excellence Strategy – EXC-2111 – 390814868, and the European Research Council (ERC) under the European Union’s Horizon Europe research and innovation program (Grant Agreement No. 101165667)—ERC Starting Grant QuSi-Gauge. J.C.H. acknowledges the hospitality of the Aspen Center for Physics, which is supported by National Science Foundation grant PHY-2210452, while this paper was being completed. This work is part of the Quantum Computing for High-Energy Physics (QC4HEP) working group. The authors acknowledge the usage of ChatGPT and Claude to improve the readability of the manuscript.

References

Quantum hardware

For the present study we used a 9×79 \times7 sublattice of the full 10×1210 \times12 ibm_phoenix processor. The selected qubits as well as the calibration data for the gate errors are all reported in Figure A1. For the experiments ran in the main text, the median error rates are: 6.7×10−36.7 \times10^{-3} for the readout errors, 1.8×10−31.8 \times10^{-3} for the single qubit gate errors, and 1.5×10−31.5 \times10^{-3} for the CZ entangling gate errors.

Figure A1. Calibration property of the quantum processor ibm_phoenix. (A) Embedding of the 5×45 \times4 physical system in the square lattice qubit hardware. (B) Cumulative distribution of error rates for single qubit gates, entangling CZ gates and measurement errors. (C) Visualization of error rates per qubits and per CZ gate.

Digital implementation and its stabilizer structure

In this section we discuss in more details the digital implementation of the Trotter dynamics in the Fradkin and Shenker model.

Trotter circuit

We split the Hamiltonian into two non-commuting parts, H=HA+HBH = H_A + H_B, with

HA=−hz∑⟨ij⟩σijz−Jv∑iτiz,(8)H_A = -h_z \sum_{\langle ij \rangle} \sigma^z_{ij} - J_v \sum_i \tau^z_i, \tag*{(8)}
HB=−Jp∑p∏⟨ij⟩∈pσijx−hx∑⟨ij⟩τixσijxτjx.(9)H_B = -J_p \sum_{p} \prod_{\langle ij \rangle\in p} \sigma^x_{ij} - h_x \sum_{\langle ij \rangle} \tau^x_i \sigma^x_{ij} \tau^x_j. \tag*{(9)}

and approximate the evolution operator by a second-order Suzuki–Trotter product

e−iHt≈(UA(dt/2)UB(dt)UA(dt/2))N,Δt=t/N,(10)e^{-iHt} \approx\left(U_A(dt/2) U_B(dt) U_A(dt/2)\right)^N,\quad\Delta t = t/N, \tag*{(10)}

with UA(dt/2)=e−iHAdt/2U_A(dt/2)=e^{-iH_A dt/2} and UB(dt)=e−iHBdtU_B(dt)=e^{-iH_B dt}

UAU_A consists of single-qubit ZZ rotations on all gauge and matter qubits and executes in parallel across the lattice. UBU_B contains the four-body plaquette and three-body gauge–matter interactions. Following the ancilla-mediated construction of Ref. [44], the plaquette term is implemented with one ancilla qubit at the center of each plaquette: each of the four surrounding gauge qubits is entangled with the ancilla by an HH–CNOT–HH sequence (gauge qubit as control, ancilla as target), a single-qubit RzR_z on the ancilla generates the plaquette phase, and the sequence is reversed to disentangle the ancilla. The hopping term is implemented by CNOTs from the matter qubits onto the gauge qubit, a single-qubit RxR_x on the gauge qubit, and the reversed CNOT sequence. The two terms are implemented in parallel. The full step is shown in Figure 1(A).

Apart from the single-qubit rotations proportional to the Trotter time step dtdt in Figure 1, every operation in the circuit belongs to the Clifford group. Removing the rotations leaves a purely Clifford circuit which, in our construction, reduces trivially to the identity, since all entangling gates appear in forward/reverse pairs. This provides a classically trivial noise-estimation circuit, exploited in Sec. C.

Operator propagation and stabilizers

To see how the circuit encodes the local interactions, it is useful to track Pauli operators through the entangling Clifford layers in the Heisenberg picture. For a CNOT with control cc and target tt,

Xc→XcXt,Zc→Zc,Xt→Xt,Zt→ZcZt(11)X_c \to X_cX_t,\quad Z_c \to Z_c,\quad X_t \to X_t,\quad Z_t \to Z_cZ_t \tag*{(11)}

while for the dressed gate U=(Hc⊗It) CNOT (Hc⊗It)U=(H_c\otimes I_t)\,\mathrm{CNOT}\,(H_c\otimes I_t),

Xc→Xc,Zc→ZcXt,Xt→Xt,Zt→XcZt(12)X_c \to X_c,\quad Z_c \to Z_cX_t,\quad X_t \to X_t,\quad Z_t \to X_cZ_t \tag*{(12)}

Applying these rules for the Toric code unitary disentangler of Figure 1, we find that all gauge stabilizer take a single qubit form at the middle of each Trotter step circuit.

Additionally, the ancillas are prepared as ∣0⟩\lvert0\rangle, i.e. as trival stabilizer states and in the absence of errors they return to ∣0⟩\lvert0\rangle after each step. The total state then has for stabilizer group

S={∣ψ⟩∣Gi∣ψ⟩=∣ψ⟩, Za∣ψ⟩=∣ψ⟩, ∀i,a}.(13)\mathcal{S}=\left\{\lvert\psi\rangle\mid G_i\lvert\psi\rangle=\lvert\psi\rangle,\ Z_a\lvert\psi\rangle=\lvert\psi\rangle,\ \forall i,a\right\}. \tag*{(13)}

At the midpoint of the step the same stabilizer group is represented differently: the Gauss stabilizers become single-qubit ZZ operators on the matter qubits, and the ancilla stabilizers become five-body operators consisting of ZaZ_a on the ancilla and XX on the four surrounding gauge qubits, see Figure 1(D). Both representations are local, and both are available at natural points of the circuit without any additional gates.

Crucially, the end-of-step stabilizers are diagonal in the computational basis. Since all our measurements are performed in that basis, their eigenvalues can be computed directly from each measured bitstring. The syndrome is therefore free: it requires no syndrome-extraction circuit, no extra qubits beyond the plaquette ancillas already needed for the dynamics, and no mid-circuit measurement.

Toric code duality

The Hamiltonian in Eq. (1) is unitarily related to Kitaev’s toric code in an external field [45], HTC=−Jv∑sAs−Jp∑pBp−hx∑iσix−hz∑iσizH_{\mathrm{TC}}=-J_v\sum_s A_s-J_p\sum_p B_p-h_x\sum_i\sigma_i^x-h_z\sum_i\sigma_i^z with As=∏⟨ij⟩∈star⁡(s)σijxA_s=\prod_{\langle ij\rangle\in\operatorname{star}(s)}\sigma_{ij}^x and Bp=∏⟨ij⟩∈pσijzB_p=\prod_{\langle ij\rangle\in p}\sigma_{ij}^z. With no external fields, the generators commute and generate the stabilizer group whose +1+1 eigenspace is the code space, coinciding with the ground-state manifold. Local Pauli errors create pairs of stabilizer defects: an XX error anticommutes with neighboring vertex stabilizers and creates electric excitations, a ZZ error creates magnetic ones through violated plaquettes. In the anyonic language these defects are quasiparticles, and the propagation of anyons maps onto the propagation of error strings. Since stabilizer measurements reveal only the endpoints of an error chain, decoding amounts to reconstructing the most likely chain from the syndrome—so the dynamics of excitations in Z2\mathbb{Z}_2 gauge theories is directly relevant to decoder design. The surface code is the planar realization with open boundaries, in which vertex and plaquette operators are truncated at the boundary [46]; its locality and high threshold make it a leading candidate for fault-tolerant computation [47, 48, 49].

In our un-gauge-fixed model the analogous role is played by the Gauss operators GiG_i, Eq. (2). A bit flip on a matter qubit τi\tau_i creates a single stabilizer defect at ii; a bit flip on a gauge qubit σij\sigma_{ij} creates two defects at the neighboring sites ii and jj. Because the symmetry is discrete, Gi2=1G_i^2=1 and the Hilbert space splits into binary sectors, so gauge-violating errors appear as binary syndrome defects closely resembling those of stabilizer codes. This differs qualitatively from LGTs with continuous gauge symmetry: U(1)U(1) quantum link models [50] are characterized by additive electric-flux divergence constraints, so gauge-violating errors change local charge sectors rather than flipping binary labels, and—although spin-12\frac{1}{2} quantum link models remain qubit-encodable—the resulting syndrome structure is less directly compatible with standard stabilizer decoding. Fault-tolerant protocols exploiting the native stabilizer structure of LGTs are an active direction [29, 51, 52, 53, 54, 55, 56, 57].

Local error detection and local error mitigation

Noise on current hardware drives the state out of the stabilized subspace S\mathcal{S} of Eq. (13). Full quantum error correction is out of reach, but the stabilizer structure can be exploited through post-selection, and the Clifford structure of the circuit through distribution-based mitigation. We describe both, and then their combination.

Local stabilizer post-selection

The stabilizers are sensitive to isolated single-qubit errors; simultaneous errors on neighboring qubits can leave the local eigenvalues unchanged and remain undetected. As the circuit deepens, errors accumulate and it becomes increasingly likely that several nearby qubits are affected, so the efficacy of stabilizer-based post-selection degrades with evolution time.

Given the circuit structure, the Gauss stabilizers are primarily sensitive to bit-flip errors, whereas the ancilla stabilizers provide sensitivity to phase-flip errors (Appendix D). Because all measurements are in the computational basis, phase flips are not directly visible in the final outcomes, so Gauss-law post-selection is expected to contribute more strongly to the recovery of the physical signal. We nonetheless use both.

Global post-selection. Retaining only samples that satisfy every constraint is immediately impractical at our system size. For the 5×45 \times4 lattice this means 20 Gauss constraints and 12 ancilla constraints.

Local post-selection. The locality of the stabilizers circumvents this exponential overhead. Rather than enforcing all constraints simultaneously across the lattice, we post-select on reduced local subsystems. This is physically motivated: the observables are local, and error propagation during finite-time evolution is approximately confined within an effective light cone [58], so gauge violations far from the support of an observable have little influence on its expectation value.

For each observable we define a cluster containing the observable and its neighborhood, chosen to maximize the number of locally accessible stabilizer constraints while keeping the subsystem small enough to retain an appreciable yield. For a single link occupation ⟨σij⟩\langle\sigma_{ij} \rangle we use Figure A2(B), with 2 Gauss and 2 ancilla stabilizers. For the matter occupations we use a cluster including 5 neighboring Gauss stabilizers of the local matter neighborhood.

Figure A2. Local clusters used for post-selection and error mitigation. (A) The local cluster used for the matter occupations tit_i, with 5 included Gauss conditions. (B) Cluster for a single link occupation σij\sigma_{ij}, containing 2 Gauss and 2 ancilla stabilizers. (C) Cluster for a pair projector ⟨PiPj⟩\langle P_iP_j\rangle of neighboring links, containing 3 Gauss stabilizers.

Combining post-selection with error-mitigation

Post-selection discards corrupted samples. To additionally correct the retained distribution we use distribution error mitigation (DEM) [30] in its marginal variant [14].

Because both techniques are local and act on the same marginal distributions, they can be composed. There is also a practical reason to combine them: error mitigation generally carries a sampling overhead that grows exponentially with depth as noise accumulates [37], whereas post-selection at these system sizes has a comparatively mild overhead (Figure A2(g)). Applying post-selection first makes the reconstruction act within the stabilized subspace, so the effective noise model seen by DEM is restricted to the Pauli processes that preserve the stabilizers; the post-selection then effectively tailors the channel entering the reconstruction. This conditioning does not break the block-circulant structure of AA, so the procedure remains well defined either way.

Statistical error bound for the distribution reconstruction

For a diagonal observable ⟨O⟩=∑xO(x)p(x)\langle O\rangle= \sum_x O(x)p(x) estimated from NN shots, standard Monte Carlo gives Var⁡(O^)=1N(Ep[O2]−Ep[O]2)\operatorname{Var}(\hat{O}) = \frac{1}{N}\left(\mathbb{E}_p[O^2]-\mathbb{E}_p[O]^2\right). In Clifford-perturbation based rescaling methods [12, 14], the reconstructed distribution is p~(x)=∑y(A−1)xypnoisy(y)\tilde{p}(x)=\sum_y(A^{-1})_{xy}p_{\mathrm{noisy}}(y), so

⟨O⟩DEC=∑y[∑xO(x)(A−1)xy]pnoisy(y).(14)\langle O\rangle_{\mathrm{DEC}}=\sum_y\left[\sum_x O(x)(A^{-1})_{xy}\right]p_{\mathrm{noisy}}(y). \tag*{(14)}

Defining the reweighted observable ODEC(y):=∑xO(x)(A−1)xyO_{\mathrm{DEC}}(y):=\sum_x O(x)(A^{-1})_{xy}, we have ⟨O⟩DEC=Epnoisy[ODEC]\langle O\rangle_{\mathrm{DEC}}=\mathbb{E}_{p_{\mathrm{noisy}}}[O_{\mathrm{DEC}}] and

Var⁡(O^DEC)=1N(Epnoisy[ODEC2]−Epnoisy[ODEC]2)≤1NEpnoisy[ODEC2].(15)\operatorname{Var}(\hat{O}_{\mathrm{DEC}})=\frac{1}{N}\left(\mathbb{E}_{p_{\mathrm{noisy}}}[O_{\mathrm{DEC}}^2]-\mathbb{E}_{p_{\mathrm{noisy}}}[O_{\mathrm{DEC}}]^2\right) \le\frac{1}{N}\mathbb{E}_{p_{\mathrm{noisy}}}[O_{\mathrm{DEC}}^2]. \tag*{(15)}

If ∣O(x)∣≤Omax⁡|O(x)|\le O_{\max} then ∣ODEC(y)∣≤Omax⁡∑x∣(A−1)xy∣|O_{\mathrm{DEC}}(y)|\le O_{\max}\sum_x|(A^{-1})_{xy}|, so with the matrix 1-norm ∥A−1∥1=max⁡y∑x∣(A−1)xy∣\lVert A^{-1}\rVert_1=\max_y\sum_x|(A^{-1})_{xy}| the standard deviation obeys

σDEC≤Omax⁡∥A−1∥1N.(16)\sigma_{\mathrm{DEC}}\le\frac{O_{\max}\lVert A^{-1}\rVert_1}{\sqrt{N}}. \tag*{(16)}

All observables in this work have Omax⁡=1O_{\max}=1.

Error mitigation benchmark on a 3x3 lattice

To benchmark the error mitigation against an exact reference, we run hardware experiments on a 3×33\times3 lattice of matter sites for the six ratios hz/Jvh_z/J_v of Table I, starting from an L-shaped string connecting opposite corners of the lattice. The raw and mitigated occupations along the initial string are shown in Figure A3, using the combined post-selection and DEM of Sec. C 2. In the deconfined regime, the mitigated data closely track the exact-diagonalization reference. As hz/Jvh_z/J_v increases toward the confined regime, the agreement on the gauge links degrades, which we attribute to the non-Clifford hzh_z rotations: their angle grows with hzh_z, and they are absent from the Clifford circuit used to estimate the noise (Sec. B).

Figure A3. Benchmark of the error mitigation against exact diagonalization on the 3 × 3 lattice with an L-string instead of an horizontal string. Occupation ⟨nq⟩\langle n_q\rangle of each matter site and gauge link along the initial string (columns, highlighted in the sketches), for the six values of hz/Jvh_z/J_v (rows). Black lines show exact continuous-time evolution, circles the raw hardware data, and squares the data after local post-selection and rescaling, using the gauge and matter clusters of Sec. C 1.

Local approximation of string-like projectors

The projectors of Eq. (7) are diagonal in the measurement basis and hence directly accessible. However, combining local post-selection with marginal DEM requires a cluster containing the entire string and its neighborhood. Since this cluster grows with the string length and the reconstructed marginal distribution grows exponentially with the cluster size, the direct mitigation of string projectors becomes hardly scalable for longer strings spanning larger lattices.

While the full projectors remain accessible at the present string lengths, as shown in Figure 4, longer strings and larger lattices will call for a more economical estimator. We therefore introduce an approximation that retains only two-link correlations along the path. For a string γ\gamma of LL links,

Pγapprox=∏⟨i,j⟩∈γ⟨PiPj⟩∏i∈γ\∂γ⟨Pi⟩,(17)\mathcal{P}_{\gamma}^{\mathrm{approx}} = \frac{\prod_{\langle i,j\rangle\in\gamma}\langle P_iP_j\rangle}{\prod_{i\in\gamma\backslash\partial\gamma}\langle P_i\rangle}, \tag*{(17)}

where Pi=(1−σiz)/2P_i=(1-\sigma_i^z)/2 is the occupation of link ii, ⟨i,j⟩\langle i,j\rangle denotes neighboring links along the path and ∂γ\partial\gamma its two boundary links. The denominator compensates for the overlap of adjacent two-link projectors, preserving the normalization, and reduces the measurement requirement from an LL-body projector to one- and two-body expectation values. We mitigate each two-link projector on the local cluster of Figure A2(C), which contains the 3 Gauss stabilizers of the matter sites adjacent to the two links, and each single-link occupation on the cluster of Figure A2(B).

Figure A4 benchmarks this estimator against the directly mitigated projectors for the five minimal-length strings of Figure 4. We find generally good agreement with the full projector in both the confined and deconfined regimes. Moreover, the statistical error bound of the distribution mitigation is significantly lower for the two-link projectors than for the full projector, due to their smaller marginal distributions. Since the cluster of each two-link projector does not grow with the string length, the approximation in principle allows a quantitative, mitigated measurement of string oscillations on larger lattices.

Figure A4. Comparison of the directly mitigated string projectors (Figure 4) with the two-body approximation (Eq. (17)), for the initial string (left) and summed over the two strings 1 plaquette (center) and 2 plaquettes (right) away, for hz=0h_z = 0 and hz=4h_z = 4. Error bars of the approximation are propagated assuming independent one- and two-body terms and do not include the systematic error of the approximation.

Error propagation, syndrome decoding and gauge protection

End-of-circuit syndromes

Rather than discarding every corrupted shot, one may attempt to decode the measured end-of-circuit syndromes and correct where possible, recovering a fraction of the affected samples. Such an approach has been used to improve signal recovery in a 2+12 + 1D Z2\mathbb{Z}_2 simulation without a plaquette term on a hexagonal-connectivity IBM processor [24]. We assess its potential here.

For measurement errors alone the situation is clean: assuming single-qubit readout errors, the Gauss stabilizers unambiguously detect and correct bit flips on both gauge and matter qubits, since a gauge-qubit flip flips two neighboring Gauss stabilizers and a matter-qubit flip flips one.

Errors during the Trotter evolution are less obvious, since they propagate. Table II and Table III record the propagation and end-of-step syndromes of single-qubit XX and ZZ errors on gauge qubits as a function of the substep at which they occur, with substeps labeled as in Figure A6. Each vertical gauge qubit neighbors two Gauss stabilizers (upper and lower matter qubits) and two ancilla stabilizers (left and right); each horizontal gauge qubit neighbors two Gauss stabilizers (left and right) and two ancilla stabilizers (above and below). Throughout this analysis the evolution circuit is approximated by its underlying Clifford structure.

Err.StepRLU midD midU endD endWeight
II–ZZZZZZZZZZZZZZZZZZZZZZZZZZZZ–
XX0–1ZZZZ−Z-Z−Z-Z−ZZZZZ-ZZZZZ−ZZZZZ-ZZZZZ1
XX2ZZZZZZ−Z-ZZZZZZZZZZZ−ZZZZZ-ZZZZZ2
XX3–6ZZZZZZZZZZZZZZZZZZZZZZZZZZZZ3
XX7ZZZZZZZZZZZZZZZZZZ−ZZZZZ-ZZZZZ2
XX8–10ZZZZZZZZ−ZZZZZ-ZZZZZ−ZZZZZ-ZZZZZ1
ZZ0ZZZZZZZZZZZZZZZZZZZZZZZZZZZZ1
ZZ1–3−Z-ZZZZZZZZZZZZZZZZZZZZZZZZZZZ2
ZZ4–5−Z-Z−Z-ZZZZZZZZZZZZZZZZZZZZZZZZZ3
ZZ6–8−Z-ZZZZZZZZZZZZZZZZZZZZZZZZZZZ2
ZZ9–10ZZZZZZZZZZZZZZZZZZZZZZZZZZZZ1

Table II. Propagation of single-qubit Pauli-X and Pauli-Z errors on a vertical gauge qubit, by Trotter substep (numbering as in Figure A6). Errors after the last substep may be read as measurement errors. RR, LL are the right and left ancilla stabilizers, UU, DD the upper and lower Gauss stabilizers; “mid” and “end” indicate the midpoint and end of the Trotter step. The last column gives the end-of-circuit error weight, i.e. the number of qubits the error has spread to.

Err.StepUDR midL midR endL endWeight
II–ZZZZZZZZZZZZZZZZZZZZZZZZZZZZ–
XX0ZZZZ−Z-Z−Z-Z−ZZZZZ-ZZZZZ−ZZZZZ-ZZZZZ1
XX1–3ZZZZZZ−Z-ZZZZZZZZZZZ−ZZZZZ-ZZZZZ2
XX4–5ZZZZZZZZZZZZZZZZZZZZZZZZZZZZ3
XX6–8ZZZZZZZZZZZZZZZZZZ−ZZZZZ-ZZZZZ2
XX9–10ZZZZZZZZ−ZZZZZ-ZZZZZ−ZZZZZ-ZZZZZ1
ZZ0–1ZZZZZZZZZZZZZZZZZZZZZZZZZZZZ1
ZZ2ZZ−Z-ZZZZZZZZZZZZZZZZZZZZZZZZZ2
ZZ3–6−Z-Z−Z-ZZZZZZZZZZZZZZZZZZZZZZZZZ3
ZZ7ZZ−Z-ZZZZZZZZZZZZZZZZZZZZZZZZZ2
ZZ8–10ZZZZZZZZZZZZZZZZZZZZZZZZZZZZ1

Table III. As Table II, for a horizontal gauge qubit. Here UU, DD are the upper and lower ancilla stabilizers and RR, LL the right and left Gauss stabilizers.

Figure A6. Implementation of gauge protection through single-qubit Rz(θprot)R_z(\theta_{\mathrm{prot}}) rotations on the matter qubits at the midpoint of each Trotter step, where the Gauss-law generators are encoded into single-qubit ZZ operators. Here θprot=−2α dt\theta_{\mathrm{prot}}=-2\alpha\,dt for the energy-penalty scheme and θprot=−2ϕk,n\theta_{\mathrm{prot}}=-2\phi_{k,n} for the pseudo-random gauge transformation scheme. The circuit substeps are labeled 0–10, matching Table II and Table III.

Two syndrome patterns decode unambiguously, independently of the substep:

  • simultaneous sign flips on two neighboring Gauss stabilizers correspond to a single bit-flip error on the gauge qubit between them;

  • simultaneous sign flips on two neighboring ancilla stabilizers correspond to a single phase-flip error on the gauge qubit between them.

Since ZZ errors are not directly visible in the computational basis, the first is the relevant one for end-of-circuit correction. Single-qubit bit flips on ancilla and matter qubits do not spread, since both appear only as CNOT targets; at the end of the circuit they produce single Gauss-stabilizer flips that do not interfere with the unambiguous signature above. Single-qubit phase flips on ancilla and matter qubits can, however, propagate into errors of weight up to 6, and can therefore in some cases produce patterns overlapping with the gauge-qubit signature. One nonetheless concludes that, provided there are no multiple errors within the same local region, a pair of neighboring Gauss-stabilizer flips can be decoded with high confidence.

How reliable is end-of-circuit decoding?

To test this in a more realistic setting we simulate the Clifford circuit corresponding to (hz,hx,Jv,Jv)=(0,0,0,0)(h_z, h_x, J_v, J_v) = (0, 0, 0, 0) with a Pauli noise channel inserting single-qubit Pauli errors with probability pp after each CNOT, and measure how often the observed syndrome patterns are indeed caused by the single-qubit process one would infer. We consider two patterns: a pair of neighboring Gauss-stabilizer flips, attributed to a bit flip on the gauge qubit between them; and an isolated single Gauss-stabilizer flip, attributed to a bit flip on the corresponding matter qubit. Only the first is expected to be near-unambiguous, since isolated flips can also arise from errors on neighboring gauge qubits that propagate to the same final pattern.

Figure A5 shows the results for a 5×45 \times4 lattice. At low error rates, pairwise flips are predominantly caused by single bit flips on the intervening gauge qubit even at large Trotter depth. As the error rate grows, the scheme degrades toward the regime where applying the corresponding correction introduces about as many errors as it removes. For isolated single-stabilizer flips the probability of the inferred origin sits at ≈50%\approx50\% at all depths, so correcting on that basis is as likely to introduce an error as to fix one and is not a useful strategy.

Probability of observed end-of-circuit syndrome patterns generated by single-qubit processes

Figure A5. Probability that an observed end-of-circuit syndrome pattern is generated by the single-qubit process described in the legend, versus Trotter depth, for the Clifford circuit of a 5×45 \times4 lattice, at three single-qubit Pauli error rates with errors inserted after each CNOT.

We conclude that end-of-circuit syndrome decoding reliably corrects single-qubit measurement errors, and—for the pairwise Gauss pattern and sufficiently low error rates—can also help for errors occurring during the evolution, extending the accessible times relative to post-selection alone.

Mid-circuit detection and gauge protection

Error detection can also be performed at intermediate stages, catching errors before they propagate further. Mid-circuit measurements are not yet reliably supported on current hardware, given their duration and associated errors, but our circuit accommodates them at no qubit or gate overhead: the ancilla stabilizers are encoded on single qubits at the end of each Trotter step, Figure 1(C), and the Gauss stabilizers on single matter qubits at its midpoint, Figure 1(D). Both are therefore accessible by single-qubit measurements at natural points of the circuit. Beyond detecting incoherent errors, mid-circuit measurement suppresses the coherent accumulation of unitary gate errors, which may be viewed as unwanted contributions to an effective Hamiltonian,

Heff=H−βHerr,[Herr,Gi]≠0,(18)H_{\mathrm{eff}} = H - \beta H_{\mathrm{err}}, \qquad[H_{\mathrm{err}},G_i] \ne0, \tag*{(18)}

with β\beta the error strength. Projecting the wavefunction back into the stabilized subspace acts as dynamical gauge protection; above a sufficiently high measurement rate this suppresses gauge-breaking coherent errors, with the transition occurring through a sharp quantum Zeno phase transition [38].

Energy penalty. Adding a term penalizing Gauss-law violations,

Hprot=H−βHerr−α∑iGi,(19)H_{\mathrm{prot}} = H - \beta H_{\mathrm{err}} - \alpha\sum_i G_i, \tag*{(19)}

shifts gauge-violating states by ∝α\propto\alpha and suppresses transitions out of the physical subspace; for α\alpha sufficiently large compared with β\beta the dynamics is effectively constrained [39].

Pseudo-random gauge transformations. Alternatively one inserts gauge transformations into the Trotter product of Eq. (10) [24],

U~(t)=∏k=1NUA(dt/2)UB(dt)UG({ϕk,n})UA(dt/2),(20)\widetilde{U}(t) = \prod_{k=1}^{N} U_A(dt/2) U_B(dt) U_G(\{\phi_{k,n}\}) U_A(dt/2), \tag*{(20)}

with UG({ϕk,n})=∏ne−iϕk,nGnU_G(\{\phi_{k,n}\}) = \prod_n e^{-i\phi_{k,n}G_n} and random angles ϕk,n\phi_{k,n} chosen such that ∑kϕk,n=0\sum_k \phi_{k,n} = 0, so no systematic coherent phase bias is introduced. Suppose that after a Trotter step coherent gate imperfections have generated a small gauge-violating amplitude, ∣ψ⟩=∣ψphys⟩+ϵ∣ψunphys⟩|\psi\rangle= |\psi_{\mathrm{phys}}\rangle+ \epsilon|\psi_{\mathrm{unphys}}\rangle, with Gn∣ψphys⟩=+∣ψphys⟩G_n|\psi_{\mathrm{phys}}\rangle= +|\psi_{\mathrm{phys}}\rangle and Gn∣ψunphys⟩=gn∣ψunphys⟩G_n|\psi_{\mathrm{unphys}}\rangle= g_n|\psi_{\mathrm{unphys}}\rangle, gn=±1g_n = \pm1, with gn≠1g_n \ne1 for at least one nn. Then, up to a global phase,

∣ψ~⟩=∣ψphys⟩+ϵe−iΔΦ∣ψunphys⟩,ΔΦ=∑nϕk,n(gn−1).(21)|\widetilde{\psi}\rangle=|\psi_{\mathrm{phys}}\rangle+\epsilon e^{-i\Delta\Phi}|\psi_{\mathrm{unphys}}\rangle,\qquad\Delta\Phi=\sum_n\phi_{k,n}(g_n-1). \tag*{(21)}

The off-diagonal terms of the corresponding density matrix carry e±iΔΦe^{\pm i\Delta\Phi}; since the angles vary between circuit realizations, ⟨e±iΔΦ⟩≈0\langle e^{\pm i\Delta\Phi}\rangle\approx0 and the ensemble average becomes

⟨ρ⟩≈∣ψphys⟩⟨ψphys∣+ϵ2∣ψunphys⟩⟨ψunphys∣.(22)\langle\rho\rangle\approx|\psi_{\mathrm{phys}}\rangle\langle\psi_{\mathrm{phys}}|+\epsilon^2|\psi_{\mathrm{unphys}}\rangle\langle\psi_{\mathrm{unphys}}|. \tag*{(22)}

Applying independent random transformations at every step continuously randomizes the phases of newly generated gauge-violating amplitudes, so leakage generated at different times no longer interferes constructively and the coherent accumulation is suppressed. Related schemes based on dynamical decoupling have also been proposed [59].

Implementation. Both schemes are free in our construction. Since the generators GiG_i are encoded into single matter qubits at the midpoint of each Trotter step, both reduce to single-qubit Rz(θprot)R_z(\theta_{\mathrm{prot}}) rotations on the matter qubits at that point, Figure A6, with θprot=−2α dt\theta_{\mathrm{prot}}=-2\alpha\,dt for the energy penalty and θprot=−2ϕk,n\theta_{\mathrm{prot}}=-2\phi_{k,n} for the random gauge transformation.

Classical simulability and entanglement

Quantum Monte Carlo is excluded for real-time dynamics due to the sign problem [60], and the state-of-the-art classical approach is based on tensor networks [36, 61], whose efficiency is tied to the entanglement of the evolving state. The optimal TN representation varies across dynamical regimes and excitation geometries, so that geometry, bond dimension or contraction strategy must be adapted case by case. In two dimensions this is sharper still [62]: matrix product states [63] require mapping the lattice onto an effective chain, with efficiency depending strongly on the correlation structure [64], while projected entangled pair states preserve the geometry naturally [65] but remain expensive to contract and evolve [66]. A state-of-the-art reference for the dynamics would use Belief-Propagation for contracting PEPS tensor networks expectation values, although we leave this to further work as our main current limitation is the low bond-dimension of the quantum state itself.

References

  1. [1]C. W. Bauer, Z. Davoudi, A. B. Balantekin, T. Bhat- tacharya, M. Carena, W. A. de Jong, P. Draper, A. El-Khadra, N. Gemelke, M. Hanada, D. Kharzeev, H. Lamm, Y.-Y. Li, J. Liu, M. Lukin, Y. Meurice, C. Monroe, B. Nachman, G. Pagano, J. Preskill, E. Ri- naldi, A. Roggero, D. I. Santiago, M. J. Savage, I. Sid- diqi, G. Siopsis, D. Van Zanten, N. Wiebe, Y. Yamauchi, K. Yeter-Aydemir, and S. Zorzetti, Quantum simulation for high-energy physics, PRX Quantum 4, 027001 (2023).DOI
  2. [2]A. Di Meglio, K. Jansen, I. Tavernelli, C. Alexandrou, S. Arunachalam, C. W. Bauer, K. Borras, S. Carrazza, A. Crippa, V. Croft, R. de Putter, A. Delgado, V. Dun- jko, D. J. Egger, E. Fernández-Combarro, E. Fuchs, L. Funcke, D. González-Cuadra, M. Grossi, J. C. Hal- imeh, Z. Holmes, S. Kühn, D. Lacroix, R. Lewis, D. Luc- chesi, M. L. Martinez, F. Meloni, A. Mezzacapo, S. Mon- tangero, L. Nagano, V. R. Pascuzzi, V. Radescu, E. R. Ortega, A. Roggero, J. Schuhmacher, J. Seixas, P. Silvi, P. Spentzouris, F. Tacchino, K. Temme, K. Terashi, J. Tura, C. Tüyüsüz, S. Vallecorsa, U.-J. Wiese, S. Yoo, and J. Zhang, Quantum Computing for High-Energy Physics: State of the Art and Challenges, PRX Quantum 5, 037001 (2024).
  3. [3]J. C. Halimeh, N. Mueller, J. Knolle, Z. Papić, and Z. Davoudi, Quantum simulation of out-of-equilibrium dynamics in gauge theories (2025), arXiv:2509.03586 [quant-ph].arxiv.org/abs/2509.03586
  4. [4]E. A. Martinez, C. A. Muschik, P. Schindler, D. Nigg, A. Erhard, M. Heyl, P. Hauke, M. Dalmonte, T. Monz, P. Zoller, and R. Blatt, Real-time dynamics of lattice gauge theories with a few-qubit quantum computer, Nature 534, 516 (2016).
  5. [5]R. C. Farrell, I. A. Chernyshev, S. J. M. Powell, N. A. Zemlevsky, M. Illa, and M. J. Savage, Preparations for quantum simulations of quantum chromodynamics in 1 + 1 dimensions. II. Single-baryon β-decay in real time, Physical Review D 107, 054513 (2023).DOI
  6. [6]A. De, A. Lerose, D. Luo, F. M. Surace, A. Schuckert, E. R. Bennewitz, B. Ware, W. Morong, K. S. Collins, Z. Davoudi, A. V. Gorshkov, O. Katz, and C. Monroe, String-breaking dynamics in a quantum simulator, Nature Physics , 1 (2026).DOI
  7. [7]B. Yang, H. Sun, R. Ott, H.-Y. Wang, T. V. Zache, J. C. Halimeh, H.-S. Yuan, P. Hauke, and J.-W. Pan, Observation of gauge invariance in a 71-site Bose–Hubbard quantum simulator, Nature 587, 392 (2020).
  8. [8]Z.-Y. Zhou, G.-X. Su, J. C. Halimeh, R. Ott, H. Sun, P. Hauke, B. Yang, Z.-S. Yuan, J. Berges, and J.-W. Pan, Thermalization dynamics of a gauge theory on a quantum simulator, Science 377, 311 (2022), arXiv:2107.13563 [cond-mat].DOI
  9. [9]G.-X. Su, H. Sun, A. Hudomal, J.-Y. Desaules, Z.-Y. Zhou, B. Yang, J. C. Halimeh, Z.-S. Yuan, Z. Papić, and J.-W. Pan, Observation of many-body scattering in a bose-hubbard quantum simulator, Phys. Rev. Res. 5, 023010 (2023).
  10. [10]Z. Wang, Z.-Y. Ge, Z. Xiang, X. Song, R.-Z. Huang, P. Song, X.-Y. Guo, L. Su, K. Xu, D. Zheng, and H. Fan, Observation of emergent Z₂ gauge invariance in a superconducting circuit, Physical Review Research 4, L022060 (2022).
  11. [11]A. N. Ciavarella, Quantum simulation of lattice QCD with improved Hamiltonians, Physical Review D 108, 094513 (2023).arxiv.org/abs/2307.05593
  12. [12]R. C. Farrell, M. Illa, A. N. Ciavarella, and M. J. Savage, Quantum simulations of hadron dynamics in the Schwinger model using 112 qubits, Physical Review D 109, 114510 (2024).arxiv.org/abs/2401.08044
  13. [13]N. A. Zemlevskiy, Scalable quantum simulations of scattering in scalar field theory on 120 qubits, Physical Review D 112, 034502 (2025).DOI
  14. [14]J. Schuhmacher, G.-X. Su, J. J. Osborne, A. Gandon, J. C. Halimeh, and I. Tavernelli, Observation of hadron scattering in a lattice gauge theory on a quantum computer (2025), arXiv:2505.20387 [quant-ph].arxiv.org/abs/2505.20387
  15. [15]C. Alexandrou, F. A. Athanodorou, K. Blekos, G. Polykratis, and S. Kühn, Realizing string breaking dynamics in a Z₂ lattice gauge theory on quantum hardware, Physical Review D 112, 114506 (2025).
  16. [16]J. Mildenberger, W. Mruckczkiewicz, J. C. Halimeh, Z. Jiang, and P. Hauke, Confinement in a Z2 lattice gauge theory on a quantum computer, Nature Physics 21, 312 (2025).
  17. [17]M. Lüscher, Symmetry-breaking aspects of the roughening transition in gauge theories, Nuclear Physics B 180, 317 (1981).DOI
  18. [18]M. Lüscher, G. Münster, and P. Weisz, How thick are chromo-electric flux tubes?, Nuclear Physics B 180, 1 (1981).DOI
  19. [19]M. Caselle, Effective string description of the confining flux tube at finite temperature (2021).arxiv.org/abs/2104.10486
  20. [20]F. D. Marcantonio, S. Pradhan, S. Vallecorsa, M. C. Baňuls, and E. R. Ortega, Roughening and dynamics of an electric flux string in a (2+1)D lattice gauge theory (2025), arXiv:2505.23853 [hep-lat].
  21. [21]K. Xu, U. Borla, S. Moroz, and J. C. Halimeh, String Breaking Dynamics and Glueball Formation in a 2 + 1D Lattice Gauge Theory (2025), arXiv:2507.01950 [hep-lat].arxiv.org/abs/2507.01950
  22. [22]Y. Tian, N. S. Srivatsa, K. Xu, J. J. Osborne, U. Borla, and J. C. Halimeh, Role of Plaquette Term in Genuine 2 + 1D String Dynamics on Quantum Simulators (2025), arXiv:2508.05736 [quant-ph].arxiv.org/abs/2508.05736
  23. [23]Google Quantum AI and Collaborators† et al., Observation of disorder-free localization using a (2+1)d lattice gauge theory on a quantum processor, Science 393, 71 (2026), https://www.science.org/doi/pdf/10.1126/science.adg9680.DOI
  24. [24]J. Cobos, J. Fraxanet, C. Benito, F. di Marcantonio, P. Rivero, K. Kapás, M. A. Werner, Ö. Legeza, A. Bermudez, and E. Rico, Real-Time Dynamics in a (2+1)-D Gauge Theory: The Stringy Nature on a Superconducting Quantum Simulator (2025).arxiv.org/abs/2507.08088
  25. [25]T. A. Cochran, B. Jobst, and E. t. Rosenberg, Visualizing dynamics of charges and strings in (2 + 1)D lattice gauge theories, Nature 642, 315 (2025).DOI
  26. [26]K. Xu, U. Borla, K. Hemery, R. Joshi, H. Dreyer, E. Rinaldi, and J. C. Halimeh, Observation of glueball excitations and string breaking in a 2 + 1D Z₂ lattice gauge theory on a trapped-ion quantum computer (2026), arXiv:2604.07435 [hep-lat].arxiv.org/abs/2604.07435
  27. [27]D. González-Cuadra, M. Hamdan, T. V. Zache, B. Braverman, M. Kornjača, A. Lukin, S. H. Cantú, F. Liu, S.-T. Wang, A. Keesling, M. D. Lukin, P. Zoller, and A. Bylinskii, Observation of string breaking on a (2 + 1)D Rydberg quantum simulator, Nature 642, 321 (2025).
  28. [28]R. Joshi, Y. Tian, K. Hemery, N. S. Srivatsa, J. J. Osborne, H. Dreyer, E. Rinaldi, and J. C. Halimeh, Observation of genuine $Z_2+1$D string dynamics in a U(1)$_1$ lattice gauge theory with a tunable plaquette term on a trapped-ion quantum computer (2026), 2604.07436.DOI
  29. [29]M. Carena, H. Lamm, Y.-Y. Li, and W. Liu, Quantum error thresholds for gauge-redundant digitizations of lattice field theories (2024), arXiv:2402.16780 [hep-lat].arxiv.org/abs/2402.16780
  30. [30]A. Gonzales, Quantum Distribution Error Mitigation via the Circulant Structure of Pauli Noise (2026), arXiv:2501.01953 [quant-ph].arxiv.org/abs/2501.01953
  31. [31]E. Fradkin and S. H. Shenker, Phase diagrams of lattice gauge theories with Higgs fields, Physical Review D 19, 3682 (1979).DOI
  32. [32]F. Wu, Y. Deng, and N. Prokof’ev, Phase diagram of the toric code model in a parallel magnetic field, Physical Review B 85, 195104 (2012).DOI
  33. [33]I. S. Tupitsyn, A. Kitaev, N. V. Prokof’ev, and P. C. E. Stamp, Topological multicritical point in the phase diagram of the toric code model and three-dimensional lattice gauge Higgs model, Physical Review B 82, 085114 (2010).DOI
  34. [34]J. J. Wallman and J. Emerson, Noise tailoring for scalable quantum computation via randomized compiling, Physical Review A 94, 052325 (2016).DOI
  35. [35]W.-T. Xu, T. Rakovszky, M. Knap, and F. Pollmann, Entanglement Properties of Gauge Theories from Higher-Form Symmetries, Physical Review X 15, 011001 (2025), arXiv:2311.16235 [cond-mat].DOI
  36. [36]F. Verstraete and J. I. Cirac, Renormalization algorithms for Quantum-Many Body Systems in two and higher dimensions (2004), arXiv:cond-mat/0407066.DOI
  37. [37]Y. Yuan, Y. Zhao, and D. E. Liu, Zeno-Enhanced Probabilistic Error Cancellation with Quantum Error Detection Codes (2026), arXiv:2605.12149 [quant-ph].arxiv.org/abs/2605.12149
  38. [38]M. M. Wauters, E. Ballini, A. Biella, and P. Hauke, Symmetry-protection Zeno phase transition in monitored lattice gauge theories (2025), arXiv:2405.18504 [quant-ph].arxiv.org/abs/2405.18504
  39. [39]J. C. Halimeh, H. Lang, J. Mildenberger, Z. Jiang, and P. Hauke, Gauge-Symmetry Protection Using Single-Body Terms, PRX Quantum 2, 040311 (2021), arXiv:2007.00668 [quant-ph].DOI
  40. [40]J. C. Halimeh and P. Hauke, Reliability of Lattice Gauge Theories, Physical Review Letters 125, 030503 (2020).DOI
  41. [41]J. C. Halimeh and P. Hauke, Stabilizing Gauge Theories in Quantum Simulators: A Brief Review (2022), arXiv:2204.13709 [cond-mat.quant-gas].arxiv.org/abs/2204.13709
  42. [42]H. Lamm, S. Lawrence, and Y. Yamauchi, Suppressing Coherent Gauge Drift in Quantum Simulations (2020), arXiv:2005.12688 [quant-ph].arxiv.org/abs/2005.12688
  43. [43]N. H. Nguyen, M. C. Tran, Y. Zhu, A. M. Green, C. H. Alderete, Z. Davoudi, and N. M. Linke, Digital Quan- tum Simulation of the Schwinger Model and Symmetry Protection with Trapped Ions, PRX Quantum 3, 020324 (2022).
  44. [44]E. Zohar, A. Farace, B. Reznik, and J. I. Cirac, Digital lattice gauge theories, Physical Review A 95, 023604 (2017), arXiv:1607.08121 [quant-ph].DOI
  45. [45]A. Y. Kitaev, Fault-tolerant quantum computation by anyons, Annals of Physics 303, 2 (2003), arXiv:quant-ph/9707021.
  46. [46]A. G. Fowler, M. Mariantoni, J. M. Martinis, and A. N. Cleland, Surface codes: Towards practical large-scale quantum computation, Physical Review A 86, 032324 (2012), arXiv:1208.0928 [quant-ph].DOI
  47. [47]A. M. Stephens, Fault-tolerant thresholds for quantum error correction with the surface code, Physical Review A 89, 022321 (2014), arXiv:1311.5003 [quant-ph].DOI
  48. [48]Google Quantum AI and Collaborators, Quantum error correction below the surface code threshold, Nature 638, 920 (2025).
  49. [49]K. t. Satzinger, Realizing topologically ordered states on a quantum processor, Science 374, 1237 (2021).DOI
  50. [50]S. Chandrasekharan and U.-J. Wiese, Quantum Link Models: A Discrete Approach to Gauge Theories, Nuclear Physics B 492, 455 (1997), arXiv:hep-lat/9609042.DOI
  51. [51]J. R. Stryker, Oracles for Gauss’s law on digital quantum computers, Physical Review A 99, 10.1103/PhysRevA.99.042301 (2019).DOI
  52. [52]A. Rajput, A. Roggero, and N. Wiebe, Quantum error correction with gauge symmetries, npj Quantum Information 9, 41 (2023).arxiv.org/abs/2112.05186
  53. [53]L. Spagnoli, A. Roggero, and N. Wiebe, Fault-tolerant simulation of Lattice Gauge Theories with gauge covariant codes, Quantum 10, 1968 (2026).DOI
  54. [54]L. Spagnoli, A. Roggero, and N. Wiebe, Qudit stabiliser codes for Z_n lattice gauge theories with matter (2026), arXiv:2602.20661 [quant-ph].arxiv.org/abs/2602.20661
  55. [55]E. J. Gustafson and H. Lamm, Robustness of Gauge Digitization to Quantum Noise (2023), arXiv:2301.10207 [hep-lat].arxiv.org/abs/2301.10207
  56. [56]X. Yao, Quantum Error Correction Codes for Truncated SU(2) Lattice Gauge Theories (2025), arXiv:2511.13721 [quant-ph].DOI
  57. [57]B. Pato and N. Klco, Trade-offs in Gauss’s law error correction for lattice gauge theory quantum simulations (2026), arXiv:2602.22121 [quant-ph].arxiv.org/abs/2602.22121
  58. [58]F. H. Lieb and D. W. Robinson, The finite group velocity of quantum spin systems, Communications in Mathematical Physics 28, 251 (1972).
  59. [59]V. Kasper, T. V. Zache, F. Jendrzejewski, M. Lewenstein, and E. Zohar, Non-Abelian gauge invariance from dynamical decoupling, Physical Review D 107, 014506 (2023).arxiv.org/abs/2012.08620
  60. [60]K. G. Wilson, Quarks and Strings on a Lattice, in New Phenomena in Subnuclear Physics: Part A, edited by A. Zichichi (Springer US, Boston, MA, 1977) pp. 69–142.DOI
  61. [61]R. Orus, Tensor networks for complex quantum systems, Nature Reviews Physics 1, 538 (2019), arXiv:1812.04011 [cond-mat.str-el].DOI
  62. [62]G. Magnifico, G. Cataldi, M. Rigobello, P. Majcen, D. Jaschke, P. Silvi, and S. Montangero, Tensor networks for lattice gauge theories beyond one dimension, Communications Physics 8, 322 (2025).DOI
  63. [63]S. Östlund and S. Rommer, Thermodynamic Limit of Density Matrix Renormalization, Physical Review Letters 75, 3537 (1995).pubmed.ncbi.nlm.nih.gov/10059611
  64. [64]A. Scardicchio, On the Optimal Layout of Two-Dimensional Lattices for Density Matrix Renormalization Group (2026), arXiv:2601.13762 [cond-mat.str-el].arxiv.org/abs/2601.13762
  65. [65]A. Kelman, U. Borla, P. Emonts, and E. Zohar, Projected Entangled Pair States for Lattice Gauge Theories with Dynamical Fermions (2024), arXiv:2412.16951 [hep-lat].arxiv.org/abs/2412.16951
  66. [66]N. Schuch, M. M. Wolf, F. Verstraete, and J. I. Cirac, The computational complexity of PEPS, Physical Review Letters 98, 140506 (2007), arXiv:quant-ph/0611050.DOI

Paper details

vlnc:W_pybgd4pq
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY-SA 4.0
Usage

views · 1

downloads · 0

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

  • Physics and Astronomy

ChatGPT, Claude

View statement
Looking for related papers…
Contents