The Signal
Lai, Chen and Hung Verify Structured Hamiltonian Encodings at 256 Qubits
A verifier certified 50 of 51 structured Hamiltonian block-encoding benchmarks within 30 minutes, including Ising and MaxCut instances with 256 system qubits. The paper also composes a verified block encoding and QSP wrapper into a circuit
Quantum.world Team · Written with AI · fact-checked ·
Fifty block encodings pass; one needs more time
A verifier proved 50 of 51 structured Hamiltonian block-encoding examples within a 30-minute limit, including Ising and MaxCut instances with 256 system qubits. The remaining case—a 256-qubit Heisenberg example containing 765 Pauli terms—did not finish before the cutoff, though the researchers report that it completed when they extended the run to 4,976.42 seconds.
Min Lai, Yu-Fang Chen and Shih-Han Hung posted their paper, “Automated Verification of Hamiltonian Simulation by Quantum Signal Processing,” to arXiv on October 29. The work targets a specific gap in quantum software checking: proving that a circuit approximates a requested time-evolution operator, rather than merely comparing it with another circuit that is presumed correct.
“The verifier proves 50 of the 51 instances within the timeout.”
Min Lai, Yu-Fang Chen and Shih-Han Hung, authors of the arXiv paper
Symbolic branches replace an exponentially large matrix
The method divides verification into two checks. First, it symbolically tracks the circuit’s action across ancilla branches, representing each branch as a weighted sum of Pauli strings. From the branch where all ancillas are zero, the verifier checks whether the circuit encodes the specified Hamiltonian and derives its normalization factor.
It then treats that verified block encoding as an abstract operation while analyzing the surrounding quantum signal processing (QSP) circuit. The checker extracts the polynomial transformation implemented by the circuit and certifies its approximation error against the desired time evolution. That structure lets it avoid building the full system matrix, whose dimension grows exponentially with the number of qubits.
The block-encoding benchmarks cover three 2-local Hamiltonian families: periodic Ising, cycle MaxCut and open-chain Heisenberg. The largest successful cases contain 512 Pauli terms for Ising and MaxCut; the verifier reports runtimes of 807.87 and 575.73 seconds, respectively, at 256 system qubits. All 51 block-encoding checks use a tolerance of 10-8.
The 1.52-million-gate figure is a composition
The paper’s headline scale figure combines two kinds of tests rather than reporting one expanded circuit checked gate by gate. The researchers pair a verified 256-qubit Ising block encoding with a separately verified QSP wrapper of degree 77, tested at effective simulation time 50. Their calculated combined circuit contains 1,518,078 nominal gates and has an error bound below 10-4 + (50/512) × 10-8.
The verifier proves that result compositionally: it checks the block encoding once, then reasons about its repeated use through an abstract interface. The gate estimate excludes extra decomposition costs for controlled oracle calls and multi-controlled projector operations. The QSP wrapper tests themselves use an opaque block encoding with one system qubit and one block-encoding ancilla, so the large-qubit composition is assembled from separately evaluated components.
Across the QSP tests, all 38 runs in a simulation-time sweep and all 12 runs in a tolerance sweep pass. At a fixed tolerance of 10-4, the polynomial degree rises from 3 to 77 as effective time increases from 0.1 to 50; the verifier’s runtime for those cases ranges from 0.40 to 3.07 seconds.
Injected faults expose the checker’s boundaries
The researchers also altered circuits to test whether the verifier would reject errors. It flagged all tested substitutions of a controlled-Z gate with a controlled-X, selector-normalization changes, reordered oracle calls and omitted Hadamard gates. Smaller phase changes were not always rejected: at the tested QSP tolerance of 10-4, all four perturbations of 10-6 passed, while the larger tested changes failed.
The paper reports a separate catch in a circuit generated by Cobble, a block-encoding software tool. For a two-qubit Heisenberg Hamiltonian with three equally weighted Pauli terms, the generated circuit encodes (H + I)/4 rather than the expected H/3. The authors trace the mismatch to a state-preparation routine that places amplitude on unused selector states when the number of equally weighted terms is not a power of two.
The largest Heisenberg case also shows the practical limit visible in the benchmark: increasing the number of terms expands the symbolic work, pushing one 256-qubit check beyond half an hour.