Quantum-classical crossover in fault-tolerant quantum dynamics simulation

Abstract: While quantum computers promise to solve classically intractable problems, identifying the point at which fault-tolerant quantum computation outperforms the best classical algorithms for practical applications remains an outstanding challenge. Here we establish a concrete quantum-classical crossover for quantum many-body dynamics under realistic hardware conditions. We introduce a scalable fault-tolerant framework that combines coherent observable estimation with a space-time-efficient implementation of non-Clifford rotations, suppressing the residual logical errors that limit existing partially fault-tolerant approaches. A benchmark against state-of-the-art tensor-network and variational Monte Carlo algorithms reveals a concrete crossover for mixed-field Ising dynamics at modest system sizes. For a physical error rate of $p=10<sup>{-3}$, fault-tolerant simulation requires approximately 2 hours and $3.7 \times 10<sup>5$ physical qubits for a 100-site 1D system, whereas tensor network approaches would require about 100 years. For 2D models, where rapid entanglement growth limits the classical evolution time, we project quantum runtimes within minutes. A physical error rate of $p=10<sup>{-4}$ leads to at least an order of magnitude reduction in qubit count ($3.1 \times 10<sup>4$ physical qubits) and runtime (minutes for 1D and seconds for 2D). The reduction in quantum runtime arises from our improved rotation-state injection and co-design of quantum error correction and observable-estimation protocols, which jointly suppress logical-error accumulation and reduce sampling overhead. Our results establish a scalable route towards practical quantum advantage and identify quantitative engineering targets for future fault-tolerant architectures.

• The paper identifies quantitative quantum–classical crossover points for mixed-field Ising dynamics, with quantum methods becoming competitive around 18–22 sites in 1D and 26 sites in 2D.
• The authors combine Gaussian-sampled Chebyshev amplitude estimation, fourth-order Trotterization, surface-code error correction, and rotation-state injection to balance sampling costs, circuit depth, and logical errors.
• The results estimate 31,200–376,000 physical qubits and runtimes from minutes to hours for 100-site simulations, while showing that lower physical error rates reduce resources by roughly an order of magnitude.

Overview and motivation

This paper establishes a concrete, quantitative crossover point at which fault-tolerant quantum simulation of many-body dynamics outperforms the best available classical algorithms. The authors address a question that has repeatedly undermined claims of quantum advantage: noisy-device demonstrations of many-body dynamics have been matched or surpassed by improved classical methods such as tensor networks (TN) and time-dependent variational Monte Carlo (tVMC). The work argues that only by combining realistic fault-tolerant resource estimates with rigorously bounded classical baselines can one locate the practical boundary of advantage.

The target task is estimation of an observable expectation value $\bra{\psi(t)}O\ket{\psi(t)}$ for real-time evolution under the non-integrable mixed-field Ising model, evolved to $t=n/2$ in 1D and $t=\sqrt{n}$ in 2D — timescales chosen so that classical error control remains possible, enabling a fair comparison. The framework is full-stack: it co-designs the observable-estimation algorithm, Trotterised circuit construction, surface-code quantum error correction (QEC), and residual logical-error mitigation, reporting costs directly in physical qubits and wall-clock time.

Algorithmic layer: Gaussian-sampled Chebyshev amplitude estimation

Direct sampling of observables incurs the standard quantum limit $\mathcal{O}(\varepsilon^{-2})$, while fully coherent Heisenberg-limited schemes require deep circuits whose residual logical errors accumulate catastrophically in partially fault-tolerant architectures. The paper introduces Gaussian-sampled Chebyshev amplitude estimation (GCAE), built on the low-depth amplitude-estimation circuit structure of Huang and Koczor (Huang et al., 5 Mar 2026). GCAE samples the number of coherent queries $m$ to the evolution operator from a truncated discrete Gaussian distribution and fits measurement outcomes to Chebyshev polynomials $T_m(b)$ via least squares.

A key technical contribution is proving that GCAE admits a uniform metric lower bound on its least-squares loss over the entire parameter range $b\in[-1,1]$, unlike the related GLSAE algorithm, which degenerates near endpoints where the angular distance becomes quartic rather than quadratic. The proven trade-off is $M \in \widetilde{O}(\varepsilon^{-1+\beta})$ maximum queries per circuit against $N \in \widetilde{O}(\varepsilon^{-2\beta})$ samples, with $\beta\in[0,1]$ interpolating continuously between Heisenberg and shot-noise scaling. This allows the algorithm to balance coherent depth against sampling cost, which is essential because deeper circuits amplify accumulated logical errors and exponentially inflate the error-mitigation overhead ($t=n/2$0).

The Trotterisation uses fourth-order product formulas, with segment counts determined not by worst-case operator-norm commutator bounds but by entanglement-informed bounds [Nature Physics 21, 1338 (2025)] and empirical estimates for the specific initial state $t=n/2$1. For entangled states the Trotter error approaches average-case scaling $t=n/2$2, an $t=n/2$3 improvement over worst-case analysis; for general Pauli-decomposed observables the paper further derives a bound via operator scrambling suggesting an additional quadratic speed-up in the Frobenius norm $t=n/2$4. This substantially reduces required circuit depth per query.

Fault-tolerant layer: rotation-state injection versus magic-state distillation

Two non-Clifford implementations are analysed. Conventional Clifford+T synthesis via magic-state distillation (MSD) is found to impose prohibitive space-time overhead: at $t=n/2$5 and $t=n/2$6, MSD requires roughly $t=n/2$7 physical qubits, rendering it impractical below one million qubits. The paper also compares direct Clifford+T scheduling against sequential Pauli-based compilation (SPBC) under fixed factory budgets, finding direct execution faster in 23 of 30 tested instances; SPBC's limited advantage traces to reduced magic-state consumption from standard-form simplification rather than to its PPM scheduling structure.

The preferred approach is a Clifford+$t=n/2$8 architecture using repeat-until-success (RUS) injection of small-angle rotation states prepared on rotated surface codes with 1-fault-tolerant (1-FT) primitives (Zeng et al., 25 Nov 2025). The central result is that the injected-rotation logical error scales as $t=n/2$9 — quadratic suppression compared to the first-order $t=\sqrt{n}$0 scaling of prior STAR-type schemes. This is what enables fully fault-tolerant simulation beyond $t=\sqrt{n}$1, a scale inaccessible to existing partially fault-tolerant designs. The RUS layer cost scales as $t=\sqrt{n}$2 with $t=\sqrt{n}$3, giving logarithmic parallel-layer overhead; angle wrapping (S-gate or T-gate based) keeps intermediate angles bounded during failure retries.

Quantum-classical crossover results

The headline numbers are stark. At physical error rate $t=\sqrt{n}$4, simulating a 100-site 1D chain to $t=\sqrt{n}$5 at accuracy $t=\sqrt{n}$6 requires approximately 2 hours (6804 s) and $t=\sqrt{n}$7 physical qubits. At $t=\sqrt{n}$8 this drops to about 5 minutes and $t=\sqrt{n}$9 physical qubits — an order-of-magnitude reduction in both resources, underscoring the paper's argument that improving gate fidelity matters more than increasing qubit count.

On the classical side, MPS simulations with systematically controlled bond dimension yield minimum bond dimensions $\mathcal{O}(\varepsilon^{-2})$0–$\mathcal{O}(\varepsilon^{-2})$1 and runtimes growing as $\mathcal{O}(\varepsilon^{-2})$2–$\mathcal{O}(\varepsilon^{-2})$3 depending on target precision. Extrapolation projects roughly $\mathcal{O}(\varepsilon^{-2})$4 years for $\mathcal{O}(\varepsilon^{-2})$5 at $\mathcal{O}(\varepsilon^{-2})$6 — far worse than the abstract's "about 100 years" figure, which corresponds to a looser accounting. Full state-vector GPU simulations grow as $\mathcal{O}(\varepsilon^{-2})$7, with peak memory already reaching 223 GB at $\mathcal{O}(\varepsilon^{-2})$8. The 1D tVMC baseline with a Jastrow ansatz cannot control its error below 0.07 even at short times $\mathcal{O}(\varepsilon^{-2})$9 regardless of body order. Crossovers occur at $m$0–22 (MPS), $m$1–27 (state-vector), and $m$2 overall in 1D.

In 2D, classical error control fails outright beyond $m$3: PEPS energy-density errors saturate above $m$4 for $m$5, never reaching $m$6. At $m$7, the best classical runs take 5.1–20 hours while the quantum estimate takes 57 seconds — two to three orders of magnitude faster and more accurate. The quantum runtime stays below 9 minutes up to $m$8 in 2D. The crossover lies at $m$9 versus state-vector methods, and approximate TNs are already slower and less accurate at $T_m(b)$0.

Limitations and open questions

Several caveats bear directly on these conclusions. First, the comparison is restricted to few-body Pauli observables — deliberately classically easy targets — yet even here classical error control fails; extension to general observables relies on block-encoding constructions not benchmarked here. Second, the Trotter step counts depend partly on empirical extrapolation for the specific initial state rather than rigorous guarantees, and the claimed entanglement-informed speed-ups assume sufficiently high entanglement entropy across relevant subsystem partitions. Third, the resource estimates assume a 500 ns QEC cycle time and specific code distances ($T_m(b)$1–18); the state-supply overheads $T_m(b)$2 are acknowledged as conservative, and adaptive scheduling could reduce them further. Fourth, magic-state cultivation is not included in the MSD optimisation, and qLDPC-based rotation injection is left as future work. Fifth, the 2D classical baseline cannot reach $T_m(b)$3 at all, so the 2D crossover is defined against a weaker classical accuracy standard than the 1D comparison. Finally, whether the GCAE sample-depth trade-off holds under realistic noise correlations within the QEC cycle model remains unverified empirically.

Conclusion

By co-designing coherent observable estimation, entanglement-informed Trotterisation, and quadratic-suppression rotation-state injection, this work converts the abstract promise of fault-tolerant quantum simulation into concrete engineering targets: roughly $T_m(b)$4–$T_m(b)$5 physical qubits and seconds-to-hours runtimes suffice for a definitive crossover over state-of-the-art tensor-network, tVMC, and state-vector methods at modest system sizes. The identification of the crossover in 1D — historically the strongest regime for classical methods — makes the case for practical quantum advantage unusually direct, and the quantitative dependence on physical error rate provides a clear metric for hardware development priorities.