Leveraging Landau-Zener-Stückelberg interference for accelerating diabatic quantum annealing

Abstract: Diabatic quantum annealing with variationally optimized schedules can exhibit exponential speedups over conventional adiabatic quantum annealing, as was demonstrated numerically for a frustrated Ising ring model by C{ô}t{é} et al. Here we identify Landau-Zener-St{ü}ckelberg interference as the underlying mechanism for this speedup, and based on this insight we propose a variational schedule ansatz with far fewer parameters. This simplified ansatz allows us to show analytically that the classical optimization of the schedule parameters can be done in polynomial time and discuss conditions when we expect this type of mechanism to provide speedups over adiabatic annealing. Furthermore, we provide an analytical argument that coherence is an essential resource for this mechanism, which we verify numerically. We perform extensive numerical tests of the proposed ansatz and observe substantial improvements over adiabatic annealing and competitive performance in particularly challenging problem instances, including a well-studied MAXCUT instance commonly used for benchmarking. Our work shows that explicitly leveraging physical mechanisms can lead to more effective designs of variational annealing algorithms.

• The paper identifies Landau-Zener-Stückelberg interference as the mechanism behind diabatic annealing speedups, using repeated spectral sweeps and waiting plateaus to transfer population before exponentially small gaps and restore the ground state afterward.
• The paper replaces schedules with roughly 100 parameters by a seven-parameter ansatz, achieving sub-quadratic scaling on a frustrated Ising ring and reducing the fitted scaling exponent to approximately 1.2–1.4 with cubic-spline interpolation.
• The paper proves that coherent LZS gates can implement arbitrary two-level transformations and supports polynomial-time schedule optimization, while showing that decoherence can destroy performance, with optimization failing near coupling strengths of g ≳ 0.1 in simulations.

Overview and motivation

This paper addresses the question of why variationally optimized diabatic annealing schedules can outperform adiabatic quantum annealing (AQA) on problems with exponentially closing spectral gaps, and uses that understanding to design a drastically simpler schedule ansatz. The starting point is the observation by Côté et al. that optimized non-monotonic schedules avoid the exponential slowdown of AQA on a frustrated Ising ring, where the interpolating Hamiltonian exhibits a perturbative anti-crossing with an exponentially small gap despite the ground state being trivial to find (2606.09706). The authors identify Landau-Zener-Stückelberg (LZS) interference as the underlying mechanism: repeated sweeps through a "control interval" of the spectrum, separated by waiting plateaus that tune relative phases, allow coherent population transfer into the first excited state before the minimal gap, so that ramping through the anti-crossing merely inverts populations back to the ground state.

The practical payoff is an ansatz with only seven parameters — two plateau locations $s_1, s_2$ and five time intervals $t_{1\ldots5}$ — versus roughly 100 parameters in the concatenated-ramp schedules of Côté et al. This parameter reduction enables three results that would otherwise be inaccessible: a proof that LZS-gates are universal on the two-level subspace, a proof that classical optimization of the schedule is polynomial-time, and a clean analytical account of how decoherence destroys the mechanism.

Theoretical framework

Universality of LZS gates

Within the assumption that dynamics are confined to the two lowest instantaneous eigenstates, each segment of the schedule acts as a gate $U(\phi,\theta)$ consisting of a fixed-mixing-angle rotation (the ramp, with transition probability $r = \sin^2\lambda$) flanked by tunable $Z$-rotations (the waiting plateaus). The key structural result is Proposition 1: for any fixed $0 < r < 1$, concatenating at most $$N_0 = 2^{\lceil \log_2 \frac{\pi}{2\lambda_{\text{eff}}} \rceil}$$ such gates realizes any element of SU(2), where $$\lambda_{\text{eff}} = \min\{\lambda, \pi/2 - \lambda\}$$. For small transition rates this scales as $\mathcal{O}(r^{-1/2})$, and when $r = 1/2$ only two sweeps suffice. The proof proceeds by showing that pairs of gates generate effective mixing angles whose accessible range doubles under iteration until it covers $t_{1\ldots5}$0.

Two consequences follow directly. First, if a "control interval" exists along the anneal — a region where the gap to the second excited state remains polynomially large while transitions within the low-energy doublet can be stimulated — then any state that evolves into the target ground state can be prepared there, regardless of exponentially small gaps elsewhere. Second, exploiting coherent interference yields a quadratic advantage over naive repetition: preparing the excited state requires $t_{1\ldots5}$1 sweeps rather than $t_{1\ldots5}$2 repetitions, analogous to Grover search.

The paper is explicit that these conditions are strict. The control interval requires $t_{1\ldots5}$3, guaranteed either by a polynomially scaling gap to higher levels or by selection rules; using a result of Callison et al., the authors bound the achievable transition rate by the eigenstate overlap across the interval, $t_{1\ldots5}$4, so intervals where eigenstates barely change cannot serve as control points. They concede these conditions may hold mainly in artificially constructed or rare instances, and make no claim about computational hardness of instances satisfying them; the conditions are sufficient but possibly not necessary.

Coherence as a resource

Decomposing the population-update rule reveals that its first two terms reproduce a classical Markov chain with transition matrix $t_{1\ldots5}$5, which equilibrates at $t_{1\ldots5}$6. Only the phase-dependent interference term permits pumping population away from equilibrium. Under dephasing, amplitude damping, or depolarizing channels inserted between gates, the reachable set of excited-state populations shrinks from full ellipses covering $t_{1\ldots5}$7 to restricted regions, and invertibility of the gate is lost. The prediction is concrete: decoherence imposes an upper bound on achievable population inversion, verified numerically in an open two-level system coupled to a thermal bath via the adiabatic master equation. For coupling strengths $t_{1\ldots5}$8 the optimizer fails entirely; for $t_{1\ldots5}$9 and $U(\phi,\theta)$0 near-complete inversion is achieved, with performance degrading monotonically in $U(\phi,\theta)$1. This establishes the speedup mechanism as inherently coherent, which limits applicability to devices with sufficiently long coherence times.

Polynomial-time optimization

Because the parameter count is fixed at seven, the loss landscape admits an affinely bounded gradient with polynomial constants whenever $U(\phi,\theta)$2 and a solution exists with $U(\phi,\theta)$3. An adaptive grid search then finds parameters to accuracy $U(\phi,\theta)$4 in polynomially many annealer calls (Proposition 2). This addresses a standard objection to variational speedups — that exponential effort is outsourced to the classical optimizer. The caveat is stated plainly: the guarantee is conditional on an efficient schedule existing, and the proven bound comes from brute-force grid search with a high-order polynomial runtime that is not practically competitive; no claim of optimizer optimality is made.

Numerical results

Frustrated Ising ring

Simulations use the Nambu formalism (free fermions, $U(\phi,\theta)$5-dimensional evolution), enabling system sizes up to those comparable with prior work. With linear interpolation, the seven-parameter ansatz achieves sub-quadratic power-law scaling of optimal anneal time $U(\phi,\theta)$6, improving on the quadratic scaling reported by Côté et al., while reproducing their qualitative dynamics: the optimized schedule populates the first excited state at the paramagnetic-to-ferromagnetic transition near $U(\phi,\theta)$7 and restores ground-state population across the perturbative anti-crossing at $U(\phi,\theta)$8, with $U(\phi,\theta)$9 throughout. The authors attribute the exponent improvement partly to optimizer budget artifacts in the earlier work rather than to a fundamental difference in mechanism.

Replacing linear interpolation with cubic splines between the same six control points reduces the fitted exponent further, from $r = \sin^2\lambda$0 to $r = \sin^2\lambda$1 ($r = \sin^2\lambda$2) or $r = \sin^2\lambda$3 ($r = \sin^2\lambda$4). No conclusive theoretical explanation is offered; the authors speculate that smooth turnarounds permit effectively local-adiabatic crossing of the phase transition, and flag this as an open question. For context, Wang et al. achieve linear scaling with a digitized approach but with a parameter count growing with system size and dynamics that evacuate the low-energy subspace, suggesting a different mechanism; Arezzo et al. establish quadratic circuit-depth bounds for QAOA on the same model.

Toy model and MAXCUT instance 1

On a toy model with a single local minimum whose depth determines whether a localized-localized avoided crossing appears, the ansatz cleanly separates the two cases: on instances without the anti-crossing, monotonic schedules beat linear ramps by a constant factor; on instances with the anti-crossing, non-monotonic schedules reach near-unit fidelity while the linear ramp fails even at maximal tested times. This confirms that localized-localized crossings — a known failure mode of AQA traceable to Anderson localization arguments — can be systematically overcome when a favorable control interval exists.

On MAXCUT instance 1 ($r = \sin^2\lambda$5), a benchmark instance with an exceptionally small gap at $r = \sin^2\lambda$6, three regimes emerge as $r = \sin^2\lambda$7 increases: (i) monotonic schedules outperforming linear ramps by a constant factor; (ii) non-monotonic LZS schedules reaching residual energies below the first excited state energy, nearly two orders of magnitude faster in $r = \sin^2\lambda$8 than regime (iii); and (iii) a distinct mechanism in which the schedule halts at the minimal gap, where the Hamiltonian acts as an $r = \sin^2\lambda$9-rotation requiring time $Z$0 — still quadratically better than an adiabatic ramp, but far slower than LZS interference. Performance is competitive with the QAOA-based results of Pecci et al. using only seven parameters. However, on 20 random 3-regular MAXCUT instances the ansatz yields only mild improvements over linear ramps, which the authors attribute to the absence of perturbative anti-crossings in unstructured random instances rather than to a fundamental limitation of the method.

Limitations and open questions

Several limitations are acknowledged explicitly. The universality and efficiency guarantees rest on the two-level confinement assumption, valid only when leakage is suppressed; whether the sufficient conditions for speedup are also necessary is unresolved, as is whether efficient trainability of this ansatz implies tractable classical simulation — the same concern raised for barren-plateau-free variational circuits. The cubic-spline improvement lacks a theoretical explanation. The mechanism's dependence on coherence restricts hardware applicability, and the weak performance on random MAXCUT instances leaves open which problem classes of practical relevance possess the required spectral structure. Finally, since Wang et al. demonstrate linear-time solvability on the frustrated ring, the true optimal exponent achievable by LZS-style schedules remains undetermined.

Conclusion

The paper converts an empirically observed diabatic-annealing speedup into a mechanistic explanation grounded in LZS interference, and leverages that explanation to compress the schedule ansatz from ~100 to 7 system-size-independent parameters. This compression simultaneously enables a universality theorem for sweep-wait-sweep gates, a conditional polynomial-time optimization guarantee, and a quantitative theory of decoherence-induced failure, all validated numerically on the frustrated Ising ring, structured toy models, and a pathological MAXCUT instance. The central lesson — that encoding known physical mechanisms into ansatz design can yield both better performance and provable properties — stands alongside the candid finding that the advantage appears confined to instances with appropriate spectral structure and demands coherent devices.