Skip to content
Quantum.world

Research

Agentic Pipelines for Mapping Topological Codes to Quantum Hardware

Formally verifies in Lean that a device Hamiltonian preserves a topological code space, turning hardware mapping from a simulation guess into a checkable proof.

Agentic Pipelines for Mapping Topological Codes to Quantum Hardware
arXiv
2611.03329
Published
2026-11-02
Authors
Arpit Arora, Adam Holmes, Krysta M. Svore, Taylor L. Patti

AI summary

Overview

Research area: Quantum error correction, analog Hamiltonian compilation, AI-agent-assisted scientific workflows, and formal verification with the Lean 4 proof assistant.

Technical level: Advanced. The paper assumes familiarity with stabilizer/topological codes, Pauli algebra, operator-norm perturbation theory, and driven superconducting circuit models.

Scope (one sentence): The paper presents an AI-agent-driven, Lean-verified pipeline that takes a topological code specification, constructs the corresponding ideal stabilizer Hamiltonian, and certifies whether the low-energy sector of a candidate superconducting device Hamiltonian reproduces the intended code space.

What This Paper Is About

Mapping a desired many-body Hamiltonian onto real quantum hardware is approximate: the device produces the interactions you want plus many you do not, and it is hard to prove that the intended physics survives. The authors build a workflow in which AI agents derive the mathematics, run circuit simulations, and formalize the results in Lean 4, so that a machine-checked proof certifies whether a device retains the correct number of isolated low-energy states and how closely those states match the target code space. They demonstrate the approach on an eight-qubit (L = 2) toric-code array realized with driven dc-SQUID couplers.

Key Contributions

  1. An end-to-end, Lean-verified pipeline for analog compilation of QEC codes. Starting from qubits and parity checks rather than a device layout, the workflow builds the ideal Hamiltonian, determines its code space and excitation gap, and certifies properties of the device's effective Hamiltonian. The authors describe this as, to their knowledge, the first Lean-verified pipeline connecting a code space to low-energy states of a device model.

  2. A commutation-aware residual bound (Theorem 2) adapted to QEC device Hamiltonians. The AI agent proposed, in response to a human request for a tighter bound, grouping anticommuting Pauli residual channels so their strengths add in quadrature rather than linearly. The authors state this adapts anticommuting-group methods from variational quantum algorithms to QEC device Hamiltonians for the first time, to their knowledge.

  3. A three-way comparison of residual bounds: the conventional coefficient-sum bound (Eq. 7), the anticommuting-group bound (Theorem 2 / Eq. 8), and a dense bound obtained by diagonalizing the full residual matrix (Eq. 9), evaluated against the finite encoded-sector criterion of Theorem 1.

  4. A diagnostic residual profile for hardware design. The pipeline identifies the strongest unwanted Pauli channels and the qubits they affect, giving specific targets for improving couplers and controls, and can be deployed before fabrication to reject or refine designs.

Main Findings

  • Certification criterion: Theorem 1 shows that if the residual satisfies 2r < g_code, where g_code is the target excitation gap, then the effective device Hamiltonian has exactly m eigenvalues below g_code/2 (m = rank of the ideal ground projector), separated from the rest of the spectrum by at least g_code − 2r. The low-energy projector obeys ‖P_low − U P₀ U†‖ ≤ r/(g_code − r). For r = g_code/4 this guarantees a separation of at least g_code/2 and projector distance of at most 1/3.

  • Ideal toric-code parameters: With J_X = J_Z = 2c, each violated check costs 2c and the smallest ideal excitation costs g_code = 4c. Topology fixes a four-dimensional ground space, and the code distance is d = L.

  • Pauli grouping is tighter than the conventional sum: The anticommuting-group bound lies between the coefficient-sum bound and the diagonalization result and is closer to the diagonalization value, avoiding overly conservative compensation requirements in hardware design. Physically, N anticommuting channels of comparable magnitude ε contribute ε√N rather than Nε.

  • Single-cell search: An agent ran an exhaustive search over 144 predefined circuit settings, with states of 2⁴ × 12 × 2 = 384 and a period T = 10.0 ns, using a 0.005 ns timestep for screening and 0.00125 ns for the selected computation. Design targets were a four-body interaction above 80 kHz, unwanted driven interactions below 20%, static interactions below 1%, and p_esc < 1%.

  • Selected single-cell result: The chosen circuit delivered h_tar = 80.916 kHz, with driven unwanted correlations of 11.4%, static unwanted correlations of 0.641%, and p_esc of 0.974%. The authors note the requested 80 kHz lies below the MHz-scale four-body couplings previously proposed for this architecture, and that the 1% static budget corresponds to 0.8 kHz.

  • Array-level result: For the periodic L = 2 torus, with four star and four plaquette couplers sharing eight edge qubits, Lean verifies four isolated low-energy states close to the ideal code space. The residual resolves into 8561 Pauli channels, aggregated by weight (one- through six-qubit interactions), and the ten individual channels with the largest |r_a| are reported as the dominant unwanted couplings.

  • Bound comparison in the array: Figure 2 reports that retaining the interaction algebra of Theorem 2 gives a bound closer to the dense witness of Eq. 9 than the coefficient-sum estimate does.

  • Explicit limits of the method: Theorem 1 is sufficient but not necessary; it can be violated by a term commuting with every stabilizer. The authors also caution that neither sandboxing nor proof checking validates the simulation programs or the physical assumptions, so human-directed auditing of equations, convergence, and error budgets is required.

Methodology in Plain English

A human-written "research contract" defines the scientific objective, permitted assumptions, and acceptance criteria. AI agents (OpenAI Codex, GPT-5-family models, for derivations, simulations, and Lean proofs; Anthropic Claude Opus 5 for external critical review) then carry out the work under a restricted execution sandbox.

  1. Build the ideal target. From the code's qubits and parity checks, the agent constructs the stabilizer Hamiltonian (XXXX star terms and ZZZZ plaquette terms), identifies its four-dimensional ground space, and computes the gap g_code = 4c.

  2. Model the device. The driven dc-SQUID cell is simulated with the full quantized flux and a multi-tone drive, retaining two states per qubit, 12 lower-mode Fock states, and two upper-mode Fock states. Each candidate setting yields the delivered four-body strength, leakage p_esc = 1 − σ_min(M)², and every parasitic Pauli term.

  3. Form the residual. The device Hamiltonian is written in the Pauli basis, the (axis-rotated) target is subtracted, and what remains is the residual R. The mismatch is bounded by r = r_cert + η, where η accounts for truncation, finite timesteps, and decimal-to-rational conversion.

  4. Bound and certify. Three bounds on ‖R‖ are computed: the sum of |r_a|, the anticommuting-group bound, and the dense eigenvalue witness. The agent translates coefficients into Lean syntax; Lean reconstructs the Hamiltonian, checks the proofs, and tests whether Theorem 1's condition 2r < g_code holds. Failed checks return errors to the agent, and unsatisfied bounds trigger further circuit search.

Why This Matters

Impact on research: The work turns "does this device realize this code?" from an informal numerical argument into a machine-checked claim, and it supplies a tighter, cheaper error budget than the standard coefficient-sum bound. It also offers a template for AI agents producing formally verified physics results under human-set scope.

Real-world applications:

  • Pre-fabrication screening of coupler and control designs for superconducting quantum processors.
  • Predictive engineering of analog quantum simulators targeting topological order and other strongly correlated phases.
  • Certified error budgets for Hamiltonian simulation and truncation-error analysis beyond QEC.
  • Reusable analog-compilation tooling that, per the authors, is independent of code family and hardware mechanism.

Industry relevance: All authors are at NVIDIA, and the workflow combines GPU-scale circuit simulation with LLM agents and formal verification, pointing to a commercial path for AI-assisted hardware design and verification in quantum computing.

Future Directions

  • Scaling beyond the eight-qubit L = 2 torus, where the dense diagonalization bound of Eq. 9 becomes impractical and the Pauli-based bounds must carry the certification.
  • Generalizing the anticommuting-group partitioning beyond finite anticommuting Pauli families, building on the restriction discussed in Ref. [68].
  • Feeding the residual channel profile back into coupler optimization to suppress the ten dominant channels identified in the array analysis.
  • Comparing Lean-certified predictions against measured device behavior, since the paper's validation is numerical rather than experimental.

Target Audience

Quantum error-correction and quantum-hardware researchers, superconducting-circuit designers, and groups applying LLMs or formal verification (Lean 4) to physics. The paper is most useful to readers already comfortable with stabilizer Hamiltonians and operator norms; the agentic-workflow and certification framing will also interest methodologists in AI-for-science and computer-assisted proof.

Authors’ abstract

Mapping a prescribed many-body Hamiltonian to a particular quantum hardware is generally an approximate process, resulting in both intended and unwanted quantum dynamics. We introduce an AI-agent-assisted workflow that uses Lean to establish whether an effective device Hamiltonian retains the intended low-energy subspace of the target many-body Hamiltonian. We demonstrate this approach with topological quantum error correction (QEC), providing, to our knowledge, the first…

The full abstract on arXiv
Read the original paper