The HALO Engine: $\mathcal{O}(1)$-Step Compilation and Localized String Rupture for Lattice Gauge Theories on Quantum Hardware

AI summary

Author: Abhiroop Gohar (Department of Physics, Indian Institute of Technology Indore) — arXiv:2608.19243v2 [quant-ph]

Overview

Research area: Quantum simulation of lattice gauge theories (LGTs) on near-term superconducting quantum hardware, with a focus on compiler/architecture design and error mitigation.

Technical level: Advanced. The paper assumes familiarity with Kogut-Susskind staggered fermions, Gauss's law operators, Jordan-Wigner mapping, the Quantum Link Model, Trotterization, and Zero-Noise Extrapolation, though the physical motivation is stated clearly enough for a physics-trained reader outside the subfield.

Scope (one sentence): The paper introduces a hardware-aware compiler ("HALO") that bounds the circuit depth of each Trotter step for lattice gauge theory simulation, then uses it — combined with error mitigation on IBM transmon processors — to observe localized string rupture and map a dynamical phase boundary on a 16-qubit mesoscopic lattice, and sketches a constant-depth 2D extension.

What This Paper Is About

Digital quantum simulations of lattice gauge theories normally use Trotterization, where the circuit depth needed for each time step grows with lattice size (as O(N) or O(N²)), so errors and decoherence overwhelm the signal before anything physical can be measured. This paper asks whether a different compilation strategy — one that maps gauge links directly onto hardware connectivity instead of using long Jordan-Wigner strings — can hold the per-step circuit depth fixed and thereby reach larger, "mesoscopic" lattices on today's noisy devices. The goal is to use that capability to actually observe non-perturbative physics: the rupture (breaking) of a stretched meson string and the confinement phase boundary of the finite lattice.

Key Contributions

1. The HALO compiler architecture. A Hardware-Aware Lattice Optimization scheme that natively maps composite gauge links onto the device topology, executing a global Trotter step in a bounded, "immutable" O(1) algorithmic depth per step and compressing a 16-qubit global step from 648 CNOTs (unoptimized Jordan-Wigner baseline) down to 56 CNOTs — characterized as up to a 91.36% reduction in entangling gate overhead.

2. A mesoscopic QLM simulation with composite gauge links. Elevating the Quantum Link Model truncation of the Schwinger model beyond prior N = 4-qubit digital demonstrations by using a composite multi-qubit (spin-1) gauge link representation, which supports non-trivial electric field dynamics rather than the constant electric-field matrix of a trivial spin-1/2 mapping. The encoding (N = 3n_links + 1) maps a subsystem of 5 composite gauge links and 6 matter sites onto 16 physical qubits.

3. Hardware observation of localized string rupture. A quantum quench of a heavily stretched meson string (coordinate length L = 15) on a 16-qubit superconducting transmon processor, combined with Zero-Noise Extrapolation, yielding a measured dynamical crossover at t ≈ 0.790 lattice units with an 18.3 ± 2.2% rupture probability, plus a g-sweep mapping an effective confinement boundary at g_c = 1.0.

4. A 2D unit-cell blueprint. A scalable, constant-depth checkerboard unit-cell architecture for four-local magnetic plaquette interactions that avoids global SWAP cascading, presented as an algorithmic pathway toward two-dimensional QCD on future grid-connected or fault-tolerant hardware.

Main Findings

• Baseline depth is the bottleneck. Standard Jordan-Wigner compilation requires 648 CNOT gates (Opt-Level 1) or 482 CNOTs (Opt-Level 3) for a 16-qubit single Trotter step; an explicit composite gauge encoding drops the Opt-Level 3 figure to 83 CNOTs, and the HALO parallel cascade reaches 56 CNOTs. The full single-step table spans 4, 7, 10, 13, and 16 qubits.

• The depth claim is specifically about routing per step, not total cost. The paper itself states that while per-step hardware routing depth stays O(1), the total number of Trotter steps needed to preserve global accuracy must scale as O(N²t²), so fundamental algorithmic complexity bounds are not bypassed. It also reports 4N − 8 two-qubit gates for HALO versus the O(N²) Jordan-Wigner count.

• Trotter error is small at the benchmark point. At t = 0.5 with 50 repetitions, the exact (infinite-depth) survival probability is 69.0309%, while the noiseless depth-2 Trotter circuit gives 68.9646% — a localized algorithmic error of 0.0663%.

• Variational ground-state preparation is accurate in the ideal case but fails on hardware. For an 8-qubit lattice, VQE (COBYLA, 48-dimensional parameter landscape) converged to −9.070575 against the exact analytical energy of −9.236068 (obtained with SciPy Lanczos sparse diagonalization) at iteration 1,900, a 1.792% ansatz representation error. On 20 qubits (exact E₀ = −23.82), execution on ibm_marrakesh produced a lowest energy of only −3.62 (peak circuit fidelity ≈ 15.2%), and at 96 qubits the unmitigated energy was +2.94.

• A hard coherence wall is measured. On ibm_marrakesh (10 qubits), the unmitigated signal falls from ≈14.4% at Depth 1 to the 0.09% white-noise floor by Depth 3. On ibm_fez (16 qubits, t = 0.8), it decays from 8.0% at Depth 1 to a noise floor reported in the figure caption as 0.0015% at depths ≥ 4.

• ZNE recovers the physical value. Using unitary folding at λ = 1, 3, 5 for the 10-qubit HALO simulation at t = 0.5 on ibm_marrakesh, raw readouts of 23.44%, 2.59%, and 0.29% extrapolated to a recovered probability of 70.6% (70.63% in the figure caption) against the 69.0% continuous physics truth. Repeat executions yielded recovered values ranging from 70% to 81%.

• String rupture crossover observed. Across 500 temporal frames, the survival probability of the original heavy meson intersects the rupture probability at t ≈ 0.790 with 18.3 ± 2.2%, where the uncertainty reflects shot noise from 2048 QPU measurements. The authors are explicit that, because of the finite Trotter depth and the Lieb-Robinson velocity, this is the localized onset of pair creation rather than a globally correlated rupture across the L = 15 string.

• Deep confinement shows oscillations, not decay. At g = 5, the electric field energy dominates the hopping term, and the string does not rupture; instead it exhibits persistent coherent oscillations.

• An effective phase boundary at g_c = 1.0. Sweeping g at fixed t = 0.8, the paper locates a boundary at exactly g_c = 1.0 separating a kinetic-dispersion regime (g < 1.0) from deep confinement. The authors note this is a known artifact of the finite-dimensional spin-1/2 QLM truncation (a transition from a ℤ₂ ordered phase to a disordered phase), since the continuum 1+1-dimensional Schwinger model is confining for all g.

• The 20-qubit VQE ansatz is algorithmically viable in simulation. A moderate-depth HALO-VQE (r = 4) with the L-BFGS-B quasi-Newton optimizer on an ideal statevector simulator converged to E ≈ −21.30 against the exact −23.82, capturing over 89% of the exact ground-state correlation energy and avoiding barren plateaus; the remaining 10.6% gap is attributed to the depth-4 circuit's intrinsic limit on long-range entanglement, not optimizer failure. Natively executing a minimal-depth (r = 2) variant on the 127-qubit ibm_marrakesh instead produced stochastic fluctuations between +0.64 and +3.32 over 20 optimization iterations.

• The 2D blueprint is algorithmic, not yet physical. The checkerboard unit-cell requires native grid connectivity; on current heavy-hex architectures such as IBM Heron, minor SWAP overheads would remain.

Methodology in Plain English

The starting point is the (1+1)-dimensional Schwinger model — quantum electrodynamics with one space and one time dimension — discretized on a lattice with staggered fermions, where matter sits on sites and the electric field lives on the links between them. Because infinite-dimensional gauge fields cannot be stored in qubits, the authors use the Quantum Link Model, which truncates the links to finite-dimensional operators. A plain spin-1/2 link would make the electric field energy a constant, so they instead use a composite, multi-qubit link that can hold several discrete energy states, restoring a dynamical string tension.

The central engineering move is to stop integrating the gauge fields out. Standard practice (Jordan-Wigner plus "integrate out the gauge field") produces an all-to-all Coulomb interaction whose long strings of Z operators force a transpiler to insert cascades of CNOTs and SWAPs — depth that grows with lattice size. HALO instead keeps explicit gauge-link qubits in place, so each kinetic and electric term acts only on a small, local neighborhood. That locality lets the compiler schedule the required gates in parallel, so the depth of one time step stops growing with qubit count. The paper benchmarks this in two ways: CNOT counts for single Trotter steps at several lattice sizes, and a StatevectorEstimator comparison of a depth-2 Trotter circuit against exact evolution.

For the physics run, the hardware experiment is set up not as adiabatic preparation of the interacting vacuum but as a deliberate quench: a quark at the first matter site, an antiquark at the last, empty vacuum sites between, and all links initialized to uniform positive flux. This creates a maximally unstable stretched string, which then evolves. The observable is not a global variance but a localized "link-neutralization" quantity tracked at the two center links (sites 7 and 8), chosen because it is a noise-resilient signature of the crossing. To combat decoherence, the authors use Zero-Noise Extrapolation: they deliberately stretch the circuit with unitary folding (odd integer scale factors), measure how the signal degrades, fit an exponential Lindblad decay, and extrapolate back to zero noise. They also run hardware saturation sweeps to document exactly where raw signals disappear into the noise floor, which is the justification for constant-depth compilation in the first place.

Two supporting protocols appear in the appendices: Nelder-Mead simplex calibration of the bare Lagrangian parameters (m_bare, g_bare) against the physical pion rest mass of 0.135 GeV using a 10-qubit, r = 4 HALO-VQE with L-BFGS-B (maxiter = 300), and a derivation of the CNOT scaling formulas. The authors emphasize that the gauge-invariant ansatz satisfies [Ĝ_n, Ĥ_k] = 0 block-by-block, so the optimization never leaks into unphysical gauge sectors.

Why This Matters

Impact on research. The paper reframes the near-term obstacle for lattice gauge theory simulation as a compilation problem rather than purely a hardware problem. Its most useful contribution may be the explicit bookkeeping: it separates what compilation can fix (per-step routing depth, entanglement overhead) from what it cannot (Trotter step count scaling as O(N²t²), and the depolarizing collapse measured at 20 and 96 qubits). The quantitative coherence-wall sweeps and the ZNE recovery are reusable benchmarks for other groups attempting mesoscopic gauge simulations. It also extends the digital frontier from the prior N = 4-qubit demonstration while noting the trade-off with analog Rydberg platforms, which have reached 51 qubits but lack strict local gauge enforcement.

Potential applications:

• High-energy and nuclear physics: digital simulation of real-time, out-of-equilibrium gauge dynamics where classical Monte Carlo hits the sign problem — string breaking, pair creation, and finite-density matter.
• Compiler and toolchain development: the localized-encoding idea and the O(1)-depth routing pattern are transferable to other lattice Hamiltonians with local interactions, not just gauge theories.
• Hardware benchmarking and error-mitigation methodology: the saturation sweeps and unitary-folding ZNE protocol give a template for establishing where a device's noise floor is before attempting a physics result.
• Path toward 2D QCD: the checkerboard unit-cell blueprint specifies concrete hardware requirements (native grid connectivity) that can inform device design for fault-tolerant lattice simulations.

Industry relevance. The workload runs on IBM Quantum cloud services (ibm_marrakesh, ibm_fez), and the results directly inform how quantum-hardware providers might co-design connectivity and compilation for lattice-physics customers. The finding that a physics-informed, gauge-restricted ansatz avoids barren plateaus on an ideal simulator while collapsing under depolarizing noise is a clear signal that fault-tolerant error correction, not better ansätze alone, is the gate to larger problems. The compiler source, folding scripts, transpiled circuits, and raw QPU JSONs are released at https://github.com/stark-069/HALO-Engine.

Future Directions

1. Fault-tolerant scaling of the 20-qubit-and-beyond regime. Unmitigated energies at 20 qubits (−3.62 against an exact −23.82, within a 15.2% peak fidelity) and at 96 qubits (+2.94) mark a hard boundary; the paper explicitly defers extracting physical observables at these sizes to future fault-tolerant hardware.

2. Physically realizing the 2D unit-cell. The blueprint's O(1) execution assumes native grid connectivity. Mapping it onto heavy-hex devices such as IBM Heron still incurs SWAP overhead, so a grid-connected processor — or a fault-tolerant implementation — is the open requirement before 2D plaquette simulations become practical.

3. Capturing genuinely global string rupture rather than its localized onset. The measured crossover is limited by Lieb-Robinson causality at shallow Trotter depth; larger circuit depth (or better error correction) is needed to observe correlations spanning the full L = 15 string, and to move beyond the finite-size ℤ₂ transition at g_c = 1.0 toward continuum confining behavior.

4. Closing the representation gap in variational state preparation. The depth-4 HALO-VQE leaves a 10.6% correlation-energy gap that deeper circuits would close but coherence limits currently forbid — an explicit trade-off the paper leaves open, along with extending the Nelder-Mead parameter calibration beyond the single pion-mass target.

Target Audience

Quantum-computing researchers working on near-term algorithm and compiler design; lattice gauge theorists interested in what digital quantum hardware can currently deliver; quantum-error-mitigation practitioners who want a worked example of ZNE and unitary folding applied to a physics observable; and hardware architects at superconducting-device providers deciding what connectivity and coherence budgets matter for lattice-physics workloads. Readers without a background in gauge theory or Trotterized circuit compilation will find the technical sections demanding, though the introduction's framing of the depth-scaling and coherence-wall problems is accessible.

Authors’ abstract

Simulating real-time dynamics in lattice gauge theories (LGTs) is severely constrained by the circuit depth overhead of standard fermion-to-qubit mappings, which scale linearly or quadratically with system size. To overcome this depth-scaling bottleneck, we introduce the Hardware-Aware Lattice Optimization (HALO) compiler, an architecture executing global time-evolution in an immutable $\mathcal{O}(1)$ circuit depth per Trotter step. By natively mapping composite gauge links to hardware topologies, HALO achieves a $91.36\%$ reduction in entangling gate overhead compared to unoptimized Jordan-Wigner baselines, compressing a 16-qubit global step to 56 CNOTs and bypassing extensive $\mathcal{O}(N)$ scaling limits. We validate this compiler on IBM superconducting transmon processors by simulating the mesoscopic Quantum Link Model (QLM) truncation of the Schwinger model. Coupling $\mathcal{O}(1)$ compilation with Zero-Noise Extrapolation (ZNE), we track localized string rupture, extracting the dynamical crossover of pair creation at $t \approx 0.790$ with an $18.3 \pm 2.2\%$ rupture probability. Furthermore, we map the dynamical phase diagram, identifying the confinement phase boundary at $g_c = 1.0$. Finally, we introduce a scalable 2D unit-cell blueprint, paving a direct pathway toward the fault-tolerant simulation of two-dimensional Quantum Chromodynamics (QCD).