WorksPhysics and Astronomy
Finding Low-Energy States of the Bernasconi Model via Simulated Transverse-Field Quantum Dynamics
Abstract
The Low-Autocorrelation Binary Sequence (LABS) problem, the ground-state problem of the Bernasconi model, has a rugged landscape with deep, isolated local minima. We study real-time transverse-field dynamics initialized in these minima as a means of reaching lower-energy configurations. Using GPU statevector simulation for N = 20-30 spins, we compare the escape probability per verified quantum sample with the success rate per energy evaluation of steepest-descent kicks, random restarts and tabu search. Using the same field and sampling window across all system sizes, we find that, relative to the best classical baseline, the sampler achieves a factor of 11.9-35.2 for the deepest traps and 13.4-27.2 for the next energy tier. These ratios describe a constant-factor improvement in sample efficiency over the sizes studied. At the adopted field, broad Hamming-space spreading coexists with a strong preference for low-energy states. This regime follows a sharp increase in displacement as the field is raised. To connect this behaviour to the landscape, a short-time expansion relates the onset of escape to target distance and intermediate energies. Bounds on isotropic spin flips and on the free transverse mixer quantify the role of the Bernasconi potential in the observed transfer. At N = 28, the ground-state probability at t = 30 exceeds a canonical cost prediction fitted to the same mean energy by a factor of about 1.9. The cost of generating each quantum sample is assessed separately through a hardware estimate, which places a direct circuit implementation two to three orders of magnitude beyond present processor capabilities.
Introduction
We study whether coherent spin dynamics can improve the escape from a deep classical minimum. The Bernasconi model provides a concrete setting in which to connect this question to optimisation of binary sequences and to earlier work on quantum search.
Binary sequences with small aperiodic autocorrelation are handy in radar ranging, pulse compression and spread-spectrum communication. Their design leads to the Low-Autocorrelation Binary Sequence (LABS) problem, whose quality measure is the merit factor [1, 2, 3]. Constructive families and correlation bounds have clarified the attainable merit factors [4, 5]. Even with this progress, finding the optimal sequence at a given length remains a demanding discrete search problem.
Bernasconi formulated this search as the ground-state problem of a long-range Ising model [6]. Its deterministic interactions produce a rugged energy landscape with many local minima, making it a practical setting for studying glassy behaviour without quenched disorder [7, 8, 9]. The distribution and structure of these minima help explain the difficulty of local search [10]. To assess search methods on this landscape, exhaustive and branch-and-bound calculations provide exact ground states and reliable benchmarks at finite sequence lengths, although their computational cost grows exponentially with length [11, 12].
Classical methods address this difficulty through structured local search, restarts and combinations of global and local moves. For example, relaxation-based search, memetic algorithms and self-avoiding walks have produced strong LABS solvers [13, 14, 15, 16]. The inclusion of LABS in the Quantum Optimization Benchmarking Library also makes it a common setting for comparing quantum and classical optimization [17]. Within this broader search problem, escape from a local minimum is a basic step shared by many heuristics. Iterated local search and basin hopping perturb the current configuration and then descend again, while multistart methods explore fresh basins [18, 19, 20]. The difficulty of this step becomes clear from the scarcity of other beneficial proposals. For the deepest traps studied here, the fraction of lower-energy configurations falls from at to at spins. This scarcity makes the distribution of escape proposals central to the performance of local search.
Thermal and quantum fluctuations offer different ways to explore such landscapes. Simulated annealing uses thermal acceptance of uphill moves [21], while quantum annealing introduces a kinetic term that couples classical configurations [22, 23]. Adiabatic algorithms, quantum annealing studies and finite-range tunnelling benchmarks establish the role of this coupling in optimization dynamics [24, 25, 26]. The corresponding algorithms and implementations are discussed in Refs. [27, 28]. A recurring physical issue in these studies is localization in configuration space, including its effect on adiabatic optimization [29].
With the Hamiltonian held fixed, the same coupling defines a continuous-time quantum walk on the graph of spin configurations. This viewpoint brings the escape problem into the framework of coherent transport and search [30, 31, 32], with the hypercube supplying the natural graph for independent spin flips [33]. In addition to that, quantum-walk studies of spin-glass optimization connect the competition between the cost Hamiltonian and the driver to the probability of finding low-energy states [34, 35].
Our protocol is closely related to population transfer, in which evolution starts from a low-energy classical configuration and redistributes amplitude among configurations of comparable energy [36, 37, 38]. Work on the quantum random energy model relates this process to localization, mobility edges and extended states [39, 40, 41]. Numerical studies further examine how the field and evolution time determine the transfer [42]. The choice of a convenient initial classical state also connects this approach to quantum local search and reverse annealing [43, 44]. Quantum-generated proposals have been studied in Markov chain Monte Carlo [45], and classical warm starts have been incorporated into variational quantum optimization [46].
For LABS itself, the quantum approximate optimization algorithm (QAOA) [47, 48] has been compared with strong classical solvers, with evidence of a scaling advantage reported for a fixed-parameter QAOA-based approach [49]. Along a different route, a recent variational study uses Pauli correlation encoding to represent larger LABS instances with few qubits [50]. These developments motivate a complementary question about a single optimization move. Given a deep classical minimum, how effectively can transverse-field dynamics produce a configuration of lower energy?
We address this question directly on the deterministic Bernasconi landscape. The initial states exhaust the symmetry-distinct minima in the two lowest excited energy tiers. Full-statevector simulations for compare the resulting sample yield with classical escape methods through their successes per energy evaluation. To understand the resulting escape probabilities, we develop a short-time expansion, geometric bounds on isotropic kicks and the free mixer, constraints from energy conservation, and a comparison with a canonical cost distribution fixed by the measured mean energy. These results connect the numerical observations to the distance, number and energies of the accessible configurations.
Model and escape problem
Our first step is to make the escape problem precise. We introduce the classical energy and the transverse-field dynamics, and then identify the initial traps and the lower-energy configurations that count as successful escape. This gives a common setting for the quantum and classical searches.
Energy and spin dynamics
We represent a binary sequence by a configuration of spins, , with . Its overlap with a shifted copy is measured by the aperiodic autocorrelation. For a lag , this is
To favour sequences with small correlations at all nonzero lags, we combine these overlaps into the classical Bernasconi energy,
The merit factor measures sequence quality, so maximizing is equivalent to minimizing [2, 3]. Expanding Eq. (2) gives the long-range spin interactions [8, 9, 12],
The energy is invariant under global spin inversion , sequence reversal and alternating sign reversal . These operations generate an order-8 symmetry group [12]. We therefore choose one canonical representative from each orbit when identifying distinct initial traps.
Having specified the energy landscape, we introduce dynamics between its configurations by replacing each spin with a Pauli- operator . The resulting diagonal cost Hamiltonian satisfies , where is the computational-basis state representing the string . During evolution, this term assigns an energy-dependent phase to each configuration, while a transverse field of strength produces the transitions between them. The full Hamiltonian is therefore
where is the Pauli- operator acting on spin , and the factor ensures that both terms scale extensively. The transverse operator couples states differing by a single spin flip, allowing probability amplitude to move across the Boolean hypercube.
Let denote the physical energy scale multiplying . We express Hamiltonians in units of and measure evolution time in units of , where is the reduced Planck constant. All quoted fields and times are numerical values in these units, and we set throughout.
Traps and lower-energy states
The escape problem begins at a configuration that resists a local energy-lowering move. We call a local minimum, or trap, when every single spin flip increases or preserves its energy,
where is the Hamming distance.
To quantify the difficulty of escaping a trap of energy , we define the minimal barrier distance as the minimum Hamming distance to any configuration of strictly lower energy,
Distance alone gives only part of the target geometry. We must also count how many lower-energy configurations lie at each distance. We denote the set of such configurations by . For an integer distance , let be the number of configurations in this set at distance from , and let be their total number. Using vertical bars to denote the number of elements in a set,
Uniform sampling of all configurations therefore has escape probability . These counts include every bitstring, including symmetry partners, whereas the seed counts in Table 1 refer to symmetry-distinct orbits.
| Tier | Orbits | range | SD hit fraction | |||
| 20 | 26 | Level 1 | 34 | 4 | 6–8 | 1.03% |
| Level 2 | 38 | 7 | 5–7 | 2.13% | ||
| 22 | 39 | Level 1 | 47 | 29 | 4–9 | 3.10% |
| Level 2 | 55 | 101 | 2–6 | 9.79% | ||
| 24 | 36 | Level 1 | 44 | 1 | 8 | 0.05% |
| Level 2 | 48 | 1 | 9 | 0.07% | ||
| 26 | 45 | Level 1 | 53 | 6 | 6–9 | 0.30% |
| Level 2 | 61 | 20 | 4–9 | 0.69% | ||
| 28 | 50 | Level 1 | 58 | 1 | 10 | 0.05% |
| Level 2 | 62 | 2 | 11 | 0.03% | ||
| 30 | 59 | Level 1 | 67 | 4 | 10–13 | 0.05% |
| Level 2 | 75 | 19 | 4–11 | 0.19% |
Table 1. Classification of exhaustive local minima tiers evaluated across system sizes . For each , the exact global optimum [12] is shown. Level 1 comprises all symmetry-distinct local minima at the lowest energy strictly above . Level 2 comprises all orbits at the next energy level above . SD hit fraction is the fraction of random-start steepest descents that end in any orbit of that tier.
A large places the nearest successful configurations many spin flips away. Within the Hamming shell at distance , which contains strings, the target fraction is , so escape must overcome both distance and scarcity.
To include the deepest traps systematically, we enumerate all one-flip local minima from the full table of energies and group them by energy after symmetry reduction. Let denote the global minimum energy and the lowest local-minimum energy strictly above it. We call the symmetry-distinct minima at Level 1, while Level 2 consists of those at the next local-minimum energy above . This exhaustive selection matters because random-start descent rarely reaches some of these low-lying basins. For example, it reaches the tiers at for and for in only 0.05% and 0.30% of runs, respectively.
The benchmark covers , from to configurations, where the lower-energy target set becomes increasingly sparse. For comparison, a 2,000-shot budget already covers more than 3% of the configurations at , where blind random sampling finds a ground state with 62% probability.
Table 1 gives the known global optimum [12], the two trap energies, their orbit counts and the range of for each size. The steepest-descent (SD) hit fractions show how frequently random-start descent reaches these tiers. With this classification in place, we evaluate every dynamical observable and escape ratio below on the orbits in this table.
Simulation and comparison
We now turn the escape problem into a common sampling test for quantum dynamics and classical search. To do so, we specify the evolution and its observables,
and compare the classical success rates through their energy evaluation cost. We then estimate the circuit resources needed to generate the quantum samples.
Time evolution and measured quantities
To follow coherent evolution under (4), we simulate the full statevector using NVIDIA cuStateVec [51] accelerator by custom CUDA kernels. We initialize the state in the computational-basis configuration of the trap, .
Write the diagonal part as and the transverse part as . For a time step , define the symmetric second-order product formula [52, 53, 54, 55],
which approximates with a local error of .
The implementation applies a full diagonal step followed by a transverse step. Denote this update by . Writing for a half diagonal step gives . For an integer number of steps, this relation becomes . The initial basis state acquires only a common phase under , while the final changes only the phases of the evolved state’s basis amplitudes. Consequently, the computational-basis probabilities agree exactly with those from (8) at every recorded time. The escape probabilities, Hamming displacements and energy distributions therefore inherit the symmetric formula’s accuracy.
The full diagonal step applies in parallel across all states using a custom kernel with hardware population counts (__popcll). The transverse mixer factorizes into single-qubit rotations, with denoting the identity operator,
applied sequentially via custatevecApplyMatrix. Simulations employ a Trotter step across a total horizon (1,200 Trotter steps).
Once the state has evolved, we use its amplitudes to follow both escape and spatial spreading. Summing their squared magnitudes over the target set defined in § 2 gives the escape probability . We evaluate its average over a fixed time interval . At the step , this interval contains 571 recorded times , with . The reported average assigns equal weight to these times,
The interval begins after the initial delay in escape and provides a common sampling protocol for every seed and system size. Equivalently, is the success probability when the measurement time is chosen uniformly from the 571 recorded times in this interval.
The window average measures the yield of the move, while the approach to the first peak reveals how quickly that yield develops. Let denote the time of the first local maximum of , using the overall peak when the trajectory rises monotonically. The first-peak growth indicator is
The figures label this quantity the initial rise rate. It measures growth up to a peak whose time depends on . § 4.1 relates this finite-time indicator to the short-time onset of escape.
To relate this growth to motion through configuration space, we also measure the mean Hamming displacement from the initial trap,
Together, the displacement and escape probability distinguish spatial spreading from the accumulation of weight on lower-energy configurations.
Classical search and success rates
We compare quantum escape with the perturbation and local-search moves used in classical heuristics. Kick followed by descent is a standard component of iterated local search and basin hopping [18, 19]. Tabu search [56] complements these moves by keeping a short memory of the recent search, and underlies strong LABS algorithms [14, 15]. All baseline implementations use cached autocorrelation updates with cost per proposed spin flip. Counting cost-function evaluations supplies a common measure of sample yield, while the quantum evolution resources are assessed in § 3.3. Explicit resource accounting is also central to broader optimization benchmarks [17].
In kick plus steepest descent (Kick-SD), we flip randomly chosen spins of and descend to a local minimum. For each , we run 50,000 trials. The success rate per evaluation is denoted by , with one evaluation charged for the sampled string and for each descent step. The best kick rate is . Restart plus steepest descent (Restart-SD) explores fresh basins by beginning from a uniformly random string. It uses 50,000 trials to estimate the rate under the same accounting [20].
To allow the local search to move beyond repeated visits to the same configurations, we also use tabu search. It begins either directly at (Tabu-direct), after a -spin kick (Tabu-kick, optimized over ), or at a uniformly random string (Tabu-restart). At each move, the tenure is chosen uniformly from . An aspiration criterion accepts a tabu move when it strictly improves the best energy found. Each trial scans all single-flip neighbours per move, charging one evaluation for the initial string and for each scan, and stops upon reaching or after 100 moves. We run 50,000 trials per mode and per tested kick distance and denote the best rate by .
Each classical rate is the number of successful trials divided by the total number of energy evaluations across all trials. The trial count is fixed, while the total evaluation count depends on the search path and stopping time. This accounting compares the methods through their yield per energy evaluation.
We use the largest of these classical rates, denoted by , to compare with the quantum yield. The quantum-to-classical escape ratio is therefore
Thus measures the yield of lower-energy configurations per energy evaluation, with one evaluation assigned to verification of each quantum sample. State preparation and evolution determine the additional cost of generating that sample. Tabu search is the stronger classical baseline on most traps, exceeding the best steepest-descent variant by a factor of 1.2–2, with larger gains on shallow traps that Tabu-direct escapes from the initial state.
Classical rates are estimated from the successes accumulated in the 50,000 trials. For a run with zero successes, let be its total number of evaluations. We use $1/N_{\mathrm{eval}}$ as the one-hit resolution of the rate, and mark the corresponding as a lower-limit estimate under this convention. For tiers with several orbits, we report medians and 68% bootstrap intervals using $B=2000 resamples of the symmetry-distinct orbits [57]. A tier containing a single orbit is reported as a single value, with orbit count .
The target counts also allow a direct geometric comparison with the perturbation itself. Consider a kick that first chooses a distance with probability and then chooses uniformly among the configurations at that distance from . Here and . We call this an isotropic kick and denote its probability of reaching by . Using the counts in Eq. (7),
Global spin inversion preserves the Bernasconi energy, so every target at distance has a target partner at distance . Thus , and targets occur only for , with defined in Eq. (6). Throughout this interval, . Denoting the resulting upper bound by , we obtain
Because the bound applies to every distribution of kick distances, it also covers a radius chosen separately for each seed and arbitrary mixtures of radii. It concerns the probability of reaching a target in the kick itself. Subsequent descent and tabu moves use the energies of visited configurations, so their performance is compared through Eq. (14). Together, these comparisons allow us to ask how much of the quantum yield comes from reaching distant shells and how much comes from favouring useful configurations within them.
Circuit cost
The sample yield becomes practically beneficial on a processor once the required evolution can be implemented. For the Bernasconi Hamiltonian, the main circuit cost comes from its two- and four-body phases. Expanding in Eq. (3) yields four-spin products before simplification. Products with reduce to two-body terms (), and coincident index sets merge, leaving a mixture of two- and four-body Pauli- operators. At , the Hamiltonian decomposes into exactly 210 two-body and 1,925 four-body commuting terms (and 182 two-body and 1,547 four-body terms at ). Synthesizing these terms on an idealized all-to-all topology requires 2 CNOTs per two-body and 6 CNOTs per four-body term, totaling two-qubit gates per Trotter step naively ( at ).
Routing on a processor with constrained connectivity increases the gate count, as shown for in Table 2. To assess how time discretization affects the escape probability and establish the largest step tolerable for hardware synthesis, Table 3 compares the window-averaged probability for coarse second-order Trotter steps with the reference at , using the deepest traps at , 20 and 24 (). A step of retains 92–98% of the reference probability (a loss of only 2–8%). At , the loss grows to 31–84%, and steps of retain at most 12% of the escape probability. The rapid deterioration past matches the threshold found for non-commuting Hamiltonians by commutator-based error analysis [55].
| Architecture | Coupling Topology | 2Q Gates / Step |
| Naive All-to-All | Idealized (unconstrained) | |
| Square lattice (e.g. IBM Nighthawk) | degree-4 planar grid | |
| IBM Heron | Heavy-hex planar |
Table 2. Two-qubit gate requirements for a single second-order Trotter step at across physical hardware topologies compared to naive all-to-all synthesis.
| () | 0.5 | 0.75 | 1.0 | |
| 16 (28) | 0.98 | 0.69 | 0.12 | 0.03 |
| 20 (34) | 0.92 | 0.26 | 0.01 | 0.01 |
| 24 (44) | 0.96 | 0.16 | 0.00 | 0.00 |
Table 3. Ratio of the window-averaged at Trotter step to its value at (second-order formula, ).
Using for the resource estimate, evolution to requires 40–120 steps. The resulting volume is of order two-qubit gates, approximately – on a square lattice and – on heavy-hex at .
Escape dynamics
With the sampling protocol fixed, we can ask which features of the dynamics account for the enhanced escape. We first establish the gain over classical search and relate it to the distance and energies of the accessible states. We then examine how the transverse field and the duration of evolution control the transfer into these states.
Escape from deep minima
We begin with the central comparison in Figure 1 and Table 4. They show the escape ratio for , using Level 1 traps at and Level 2 traps at . The energy tiers, orbit counts and barrier distances are those in Table 1. The same field and sampling window are used at every size, including and 30, so the comparison follows one common escape protocol.
| Energy Tier | ||||||
| Level 1 (Deepest Traps) | ||||||
| Trap Energy | 34 | 47 | 44 | 53 | 58 | 67 |
| Distinct Orbits () | 4 | 29 | 1 | 6 | 1 | 4 |
| Barrier Distance Range | 6–8 | 4–9 | 8 | 6–9 | 10 | 10–13 |
| Sample Median | 30.22 | 15.84 | 20.20 | 26.75 | 35.24 | 11.87 |
| 68% Bootstrap CI | [24.28, 66.17] | [14.87, 16.93] | [21.13, 32.79] | [7.93, 16.27] | ||
| Level 2 (Excited Traps) | ||||||
| Trap Energy | 38 | 55 | 48 | 61 | 62 | 75 |
| Distinct Orbits () | 7 | 101 | 1 | 20 | 2 | 19 |
| Barrier Distance Range | 5–7 | 2–6 | 9 | 4–9 | 11 | 4–11 |
| Sample Median | 27.22 | 16.10 | 25.53 | 23.87 | 13.36 | 23.18 |
| 68% Bootstrap CI | [26.26, 27.76] | [15.41, 16.83] | [19.23, 27.25] | [13.24, 13.48] | [21.21, 26.04] |
Table 4. Quantum-to-classical escape ratio at , window , where is the best of steepest-descent kicks, restarts and tabu search (Eq. (14)). Medians over symmetry-distinct orbits with 68% bootstrap intervals. Entries marked use the one-hit resolution convention for zero-success classical runs. Levels follow Table 1. Orbits with shallow barriers are marginal (a two-flip kick escapes them with probability).
Figure 1. Quantum-to-classical escape ratio versus system size for Level 1 (deepest traps) and Level 2 (second-hardest traps) at , against the best of steepest-descent kicks, restarts and tabu search. Points denote medians over symmetry-distinct orbits with 68% bootstrap intervals (). Upward arrows mark one-hit resolution estimates for zero-success classical runs. The dotted line denotes parity ().
The median ratio exceeds unity at every size in both tiers. More specifically, Level 1 medians lie between 11.9 and 35.2 and Level 2 medians between 13.4 and 27.2 over –30 (Table 4). The ratio varies with the orbits present at each size and stays within a constant-factor range over –30. The same field and sampling window give an enhanced yield also at and 30.
To understand this improvement, we turn to the distributions generated by the escape moves. An isotropic kick assigns equal probability to configurations at the chosen distance, after which descent concentrates the samples in nearby minima. Quantum evolution spreads amplitude over large Hamming distances (, Sec. 4.2) while enhancing the weight of low-energy configurations relative to blind sampling (Figure 4). The coexistence of broad spreading and enhanced low-energy weight leads us to examine the target geometry and the phases accumulated along the paths between configurations.
Figure 4. Energy distribution of the quantum state initialized in the deepest trap (, ). (a) Full distribution at compared to the blind density of states , where counts configurations of energy (grey dashed). While the mean energy – remains far below the random-string expectation , the wavepacket broadens into low-lying states. (b) Detail of the low-energy region . The ground-state level (8 energy units below the trap) accumulates probability by , an enrichment factor of over blind sampling ().
Figure 2 makes this connection through the initial rise rate and the window-averaged probability across the evaluated traps, whose minimal barrier distances span . Comparing these two observables shows how the distance to a target and the number of available targets enter at different stages of escape.
Figure 2. Dependence of quantum escape dynamics on minimal barrier distance . (a) Initial rise rate versus , showing the rapid decrease in the first-peak growth indicator as the distance increases. (b) Window-averaged escape probability versus , showing the range of later-time yields across target distances and energy tiers. Data points are colored by system size and shaped by coupling strength .
For a fixed barrier distance (for instance, ), window escape probabilities vary by nearly two orders of magnitude (ranging from up to ) depending on the trap’s energy tier . The deepest classical traps (Level 1, , with barriers –13) systematically exhibit lower window escape probabilities (– across ), while excited local minima (Level 2, with –11) achieve higher escape probabilities (up to ). This separation reflects the volume of the target set defined before Eq. (7). For excited traps, encompasses extensive lower-lying basins, whereas for Level 1 traps it is restricted to rare states near the global optimum ( at ).
To understand the early part of this evolution, we first use the distance in Eq. (6) to determine how escape begins. For a continuous-time quantum walk, the shortest paths determine the leading power of the transition probability [59]. Here the hypercube structure also fixes its coefficient. Consider a target at distance , and let be a nonnegative integer power of the Hamiltonian. Each off-diagonal factor in Eq. (4) flips one spin, so reaching requires at least such factors. At that order, the required flips can occur in different sequences, giving
The factorial cancels the Taylor coefficient of the evolution operator. Since the Hamiltonian is real in the computational basis, transition probabilities are even functions of time. Summing over the nearest targets in Eq. (10) gives the exact onset law
Distance fixes the power of time, while the number of nearest targets fixes its coefficient. This provides a direct explanation of the initial suppression in Figure 6.
Figure 6. Continuous-time escape probability for the hardest local minimum at each system size at . Horizontal dashed lines indicate blind search probability .
The leading onset depends only on the shortest paths and their number. To see where the landscape energies enter, we next examine the first correction. For a fixed target at distance , let be the configuration after flips along a shortest sequence, with and . Denote the full sequence by . Choose these sequences with equal weight. Define the energy offset at using the normalization of Eq. (4), and denote their sum along the sequence by ,
To keep track of the phase accumulated during a sequence, let be the fraction of the total time spent at . Integrating over the ordered flip times gives a uniform distribution over with . With these fractions, the energy averaged along the path is . We characterize its variation over both the paths and the time fractions by the variance . Writing for the uniform average over paths gives
To see how this variance enters the transition probability, write for the average over paths and time fractions . After removing the common phase , the contribution with flips becomes . Its squared modulus supplies the correction . In addition, paths with two further flips contribute the correction already present for the free transverse mixer. Combining the two contributions yields
The coefficient proportional to follows by expanding the free mixer probability . Thus the geometry determines the leading onset, and the spread of energies along the paths determines its first landscape-dependent correction. The nonnegative variance makes this first landscape correction suppressive. The enhanced escape measured in the sampling window therefore develops beyond the earliest shortest-path contribution, as further spin flips and relative phases enter the evolution. This later evolution connects the window probability in Eq. (11) and the first-peak growth indicator to the field dependence observed in Figure 2.
We can now compare this early-time structure with the transfer observed inside the sampling window. Broad Hamming-space spreading (, § 4.2) occurs together with enhanced weight on lower-energy configurations. This is the combination expected in energy-selective population transfer [37, 38]. The transverse field mixes configurations across Hamming shells, while the phases generated by give the transfer its energy dependence. Window probabilities reach order in the deepest tier and in the second tier, even among traps with –, where the number of possible kicks reaches order .
Figure 4 shows this directly for the deepest trap. At the weight on the ground-state energy is , about times the blind fraction of , while the mean energy stays far below the random-string mean . The state therefore combines broad Hamming displacement with a strong preference for low energies. The enhanced ground-state weight occurs while the mean cost energy rises from the trap value 58 to about 156. Escape is therefore accompanied by a redistribution of probability towards both lower and higher classical energies.
To quantify the preference for low energies within the distance shells, we use Eq. (16). For the hardest orbits in Table 5, the target counts are at and at , with and 13, respectively. The corresponding isotropic-kick bounds are
and . The window probabilities exceed these bounds by approximately 293 and 203, respectively. Since the bound holds for every choice of kick radius and for their mixtures, the enhancement requires a preference for successful configurations within the Hamming shells. The same bound applies to a time-averaged distribution, so the comparison applies directly to the sampling window used in the simulations.
This change in the mean cost energy has a neat interpretation through energy conservation. The relation between the cost and driver energies [35] connects the mean energy in Figure 4 to the transverse spin polarization. Write for expectation in the state evolved from . Initially , so continuous-time evolution under Eq. (4) conserves the total energy . Consequently,
where the inequality follows because the sum of Pauli- operators has eigenvalues between and . Define the variance of the full Hamiltonian by . Acting once on , the driver produces orthogonal one-flip states, each with amplitude . The conserved variance is therefore
At fixed and , every initial basis trap has the same standard deviation of the total energy, even though their escape probabilities differ. Their different yields probe the geometry of the lower-energy targets and the energies encountered on the way to them.
The low mean cost energy suggests a further question. How much of the enhanced ground-state weight follows from this mean alone, and how much depends on the way probability moves through the landscape? To make this comparison, write for the probability of measuring classical energy at time , and for its mean. Let count all configurations of energy , including their symmetry partners. Among distributions over these configurations with a prescribed mean, maximizing the Shannon entropy gives a canonical cost distribution [65]. Denote its dimensionless inverse temperature by and its partition function by . With the energy normalization of Eq. (4), they enter as
Thus the density of states fixes the relative abundance of each energy level, while the Boltzmann factor fixes the preference for low energies. For each snapshot, define the effective inverse temperature by matching its mean,
The canonical mean decreases monotonically with , since its derivative is minus the canonical energy variance divided by . Consequently, this condition determines a unique inverse temperature, after which every energy-level probability is fixed.
For the run in Figure 4, the saved density of states contains 803 distinct energies and sums to exactly configurations. We use the recorded first moments 154.0269 at and 155.8906 at in Eq. (25).1
The corresponding inverse temperatures are 1.4351 and 1.3968. There are eight ground-state configurations at , so Eq. (24) predicts ground-state probabilities of and , respectively. The measured values are and . To compare the distributions at every energy, define the enrichment as the measured probability relative to this canonical prediction,
Figure 5 shows both distributions and this ratio. At the ground-state energy, and . In addition, the probability at the initial energy remains 0.0826 at , about 881 times its canonical value. Substantial excess weight also persists in the high-energy tail. Between these two snapshots, the ground-state probability grows by a factor of 12.5 while the mean cost energy changes by only about 1.2%. The transfer therefore carries information beyond the mean, simultaneously retaining weight at the initial trap energy and increasing the population of the lower ground-state level.
Figure 5. Canonical cost comparison for the deepest trap (, ), using the archived snapshots in Figure 4. (a) Recorded probabilities at and (solid lines with markers) and canonical distributions with the same recorded first moments (dashed lines in matching colours). The grey dotted curve is . Each canonical curve is normalized using the complete density of states, including energies beyond the displayed range. (b) Enrichment from Eq. (26) in the low-energy region. The horizontal line denotes agreement with the canonical prediction. The ground-state enrichment rises from 0.129 to 1.86, while the initial trap energy retains a much larger excess. Lines join discrete energy levels to guide the eye.
The distinction between cost energy and total energy is useful when interpreting this test. For a canonical reference based on the full quantum Hamiltonian, let be its inverse temperature and let project onto the configurations of classical energy . Taking the trace over the full spin Hilbert space gives its prediction for a cost measurement,
Here is fixed by the conserved mean total energy , and the transverse field remains inside the Gibbs operator. Equilibration within conserved symmetry sectors also retains their initial weights [66]. Equation (26) quantifies the additional energy selection beyond a mean-matched canonical cost distribution. Field and time enter that reference through the measured mean, while records the further structure generated by the dynamics.
Energy conservation also constrains how broadly the classical energy can be sampled. A uniform distribution over bitstrings has mean energy , as follows by averaging Eq. (3). For a seed below this mean, define as the lower bound on the field magnitude required to reach the uniform mean. Equation (22) gives
Table 5 lists for the six hardest orbits. Its values lie between 0.3802 and 0.4089. For the , seed, . At the adopted field , Eq. (22) gives , below the uniform mean 378. The continuous-time dynamics thus retains a mean energy below the uniform-string value even as the state spreads in Hamming distance. At , the field exceeds for all six seeds, so this energy constraint allows the mean to reach the uniform value. The actual distribution is then determined by the transfer dynamics. These conservation relations refer to the continuous-time Hamiltonian used as the reference for the Trotter evolution in § 3.
| Window | Best Classical Method | Ratio | ||||||
| 20 | 34 | 0.3900 | 8 | Tabu Direct | 34.9× | |||
| 22 | 47 | 0.3802 | 9 | Kick-SD () | 15.6× | |||
| 24 | 44 | 0.4028 | 8 | Tabu Restart | 20.2× | |||
| 26 | 53 | 0.4024 | 9 | Tabu Kick () | 21.1× | |||
| 28 | 58 | 0.4082 | 10 | Tabu Kick () | 35.2× | |||
| 30 | 67 | 0.4089 | 13 | Tabu Kick () | 8.1× |
Table 5. Hardest single orbit at each system size (), defined as the Level 1 orbit with the largest (first evaluated orbit on ties). is the fraction of all configurations below . The ratio uses the best classical method listed. is the necessary field bound for reaching the uniform mean cost energy in Eq. (28). These are single-orbit values. Table 4 and the abstract quote medians over all orbits in a tier.
Effect of the transverse field
The preceding analysis shows how the landscape shapes the transfer. We now ask how this transfer changes when the transverse field becomes stronger. The field controls the balance between the diagonal Bernasconi potential and the spin flips generated by . Figure 3 follows this balance through the window-averaged escape probability and mean Hamming displacement , for and .
Figure 3. Effect of transverse field on quantum escape and Hamming-space spreading for system sizes . (a) Window-averaged escape probability versus , showing the field dependence of the low-energy yield within the chosen window. (b) Mean Hamming displacement versus , showing increasing displacement towards . Error bars indicate standard error of the mean (SEM).
The transverse field alone gives the continuous-time walk on the hypercube [32, 33], providing a useful reference for this spreading. With the diagonal term removed, Eq. (9) makes each spin flip independently with probability upon measurement. Denote the resulting escape probability by and the mean Hamming displacement by . Then
For the free mixer, the field enters only through , so changing rescales the time along the same evolution. Once the Bernasconi term is present, changing also changes the balance between spin flips and the phases accumulated from the fixed energy landscape.
To bound the escape produced by this free spreading, consider a flip probability and a target at distance . The probability is largest at . Over the allowed target distances , this maximum is largest at the endpoints. Summing the bound over the targets gives
The bound holds for every time and field, and hence for every sampling window. For the hardest and seeds it gives and , respectively. The observed window probabilities are about and times larger. The diagonal Bernasconi term is therefore essential to the observed transfer into lower-energy configurations, while
the mixer alone already produces a large Hamming displacement.
With this reference in place, the field dependence in Figure 3 traces a progression from partial spreading to broad exploration. At , the mean Hamming displacement is approximately – for , or – of . For example, at , whose hypercube midpoint is 15. At the displacement is 7.15, or of . The potential retains substantial weight near the initial trap. This slower spreading can nevertheless produce a larger late-time escape probability for excited traps, as discussed in § 4.3.
Increasing the field to produces a pronounced expansion. For , the mean displacement rises to –, or – of , reaching at . The wavepacket now explores configurations well beyond the initial basin while retaining elevated escape probabilities, with the averages in Figure 3 ranging from about to .
Increasing the field further to brings the mean displacement to approximately . Its values are , , , and for , , , and , respectively. However, the escape probability decreases at the same time, reaching about at . Stronger driving thus gives a larger displacement but a smaller yield of lower-energy configurations. This decrease occurs above the energy-conservation bound listed in Table 5, where a wider range of mean cost energies becomes accessible. The comparison with Eq. (30) shows why both observables are useful. Efficient escape requires the diagonal potential to shape the distribution as the transverse field spreads it. A superposition with equal weight on a configuration and its globally inverted partner has mean Hamming displacement from the initial trap. Together with this displacement, the energy distribution identifies which parts of the landscape carry the probability.
The rapid change from partial spreading at to displacement above 85% of at marks a finite-time crossover to broad Hamming-space spreading at the sizes studied. The adopted field, , combines this broad spreading with a strong preference for low-energy configurations within the chosen window. A similar competition between spreading and energy selectivity appears in population transfer in random-energy and impurity-band models [37, 38, 42], whose extended-state regimes have also been studied analytically [41]. Here the deterministic Bernasconi potential retains the energetic preference while the field supplies enough mixing to traverse wide Hamming barriers. At weaker fields the transfer builds more slowly, and at stronger fields the additional spreading reduces the low-energy yield within the same window.
Growth of escape over time
The field scan compares averages over one interval. To see how those averages develop, we now follow the full escape trajectories and then vary the time at which they are compared. Figure 6 shows for the hardest local minimum at each , using .
As expected from the short-time expansion, the trajectories begin with a delay of order –. All these seeds have , so reaching the target requires several spin flips and across the sizes shown. This suppression follows from the power of time in Eq. (18). During –, the escape probability grows by two to four orders of magnitude as the state develops weight in lower-energy basins, reaching order for the larger-probability trajectories.
For the deepest Level 1 traps at the late-time behaviour differs by instance. At (, ) the probability has essentially plateaued by the end of the window ( against a window average of ), whereas at () it is still rising at ( for the orbit, its window average). The late rise is more pronounced for Level 2 traps and for , as the comparisons in Table 7 will show. For these rising trajectories, the probability measured at exceeds the fixed-window average.
| Trap | Field | Factor | ||||
| 24 | 44 | 8 | 0.30 | 0.93× | ||
| 24 | 48 | 9 | 0.30 | 4.30× | ||
| 28 | 58 | 10 | 0.35 | 1.10× | ||
| 28 | 62 | 11 | 0.30 | 2.69× | ||
| 30 | 67 | 11 | 0.35 | 2.54× | ||
| 30 | 67 | 13 | 0.35 | 2.34× | ||
| 30 | 67 | 13 | 0.30 | 1.90× |
Table 7. Comparison of window-averaged escape probability versus late-time escape probability for deep traps across .
To appreciate the size of this transfer, we compare these trajectories with the fraction of configurations that qualify as targets. For the hardest orbit at each size, Table 5 lists the blind success probability , using the target count in Eq. (7). It falls by more than two orders of magnitude, from at to at . The quantum window probabilities range from to , showing a strong concentration of weight on the sparse lower-energy set.
Classical local search also uses the landscape to improve its proposals, so it gives a more demanding comparison than the blind fraction. Their best rate for the hardest orbit is at and at , where the barrier distances in the table grow from 8 to 13. Relative to this best baseline, quantum sampling gives – for the hardest individual orbits within the fixed window . Table 4 gives the corresponding medians over all orbits in each tier.
The same trajectories also let us revisit the choice of field. Denote by the transverse field giving the largest escape probability at the observation time. Although produces faster early transfer, the initially slower trajectory can continue to grow until it matches or exceeds the stronger-field result. Thus the preferred field depends on how long the move is allowed to develop.
Table 6 connects the preferred field to the time available for the move. At , the state spreads rapidly beyond the initial basin and develops a larger early escape probability. At , the initial growth is slower, but the probability continues to rise beyond and can overtake the stronger-field trajectory. The same competition between exploration and energetic preference therefore appears both in the field scan and in the ordering of the late-time probabilities. The dependence on observation time is also natural in quantum-walk optimization and population transfer [34, 35, 42]. Under unitary evolution, probability can flow both into and out of the lower-energy set. The crossing of trajectories at different fields therefore makes the measurement time part of the escape move, alongside the field strength.
| Time | Dynamic Regime | ||||
| (, ) | 5.0 | 0.43 | faster rise | ||
| 10.0 | 0.43 | faster rise | |||
| 38.5 | 0.82 | catching up | |||
| 50.0 | 1.96 | superior | |||
| 60.0 | 1.37 | superior | |||
| (, ) | 10.0 | 0.21 | faster rise | ||
| 38.5 | 0.70 | catching up | |||
| 50.0 | 1.55 | superior | |||
| 60.0 | 1.50 | superior | |||
| (, ) | 10.0 | 0.30 | faster rise | ||
| 38.5 | 0.89 | catching up | |||
| 50.0 | 1.31 | superior | |||
| 60.0 | 1.37 | superior |
Table 6. Escape probability comparing couplings and for evaluated traps across . Early on (), leads. Later (), maintains amplitude in low-energy basins, overtaking . In the probability columns, the argument denotes the field strength at the listed time.
This time dependence also explains the role of the measurement window. Fixing the window gives a common protocol for every instance, while following the trajectories out to reveals how its choice affects the measured yield. Several Level 2 orbits and runs continue to rise, as does the orbit at under . Other trajectories, including the deepest and traps at , have largely plateaued within the observation interval. Table 7 compares the fixed-window average with the observed endpoint probability for representative traps. For the rising trajectories in Table 7, exceeds the window average by factors up to 4.30. At , and , for example, the trajectory has and , a factor of 2.54. For the hardest orbit at the same field, the fixed-window ratio is and the endpoint probability is 2.34 times the window average. These observations identify a further gain in sample yield available at the measured later time. They complement the medians at every size and tier under the same field and sampling window.
The same effect becomes more pronounced at the weaker field. At , and , gives within , or relative to . By , the escape probability reaches , a factor of 1.90 higher and slightly above the classical rate. The additional evolution time is therefore a physical resource that can change the ranking of the escape modes.
Conclusions
The numerical comparison and analytical results together identify the physical ingredients of successful escape. We bring these observations together here and use them to outline the next steps for understanding and implementing the dynamics.
Real-time transverse-field dynamics provides an effective escape move from the deep local minima of the Bernasconi model. The GPU statevector calculations cover –30, up to approximately basis states, and begin from the complete set of symmetry-distinct traps in the two lowest excited tiers. With the same field and sampling window across all sizes, the median yield of lower-energy configurations exceeds the best classical perturbation rate by factors of 11.9–35.2 for Level 1 and 13.4–27.2 for Level 2. The improvement remains a constant factor across this range of sizes, including the larger instances at and 30.
The analytical results explain how this gain develops from the structure of the dynamics. The minimum target distance fixes the power of the short-time onset, the number of nearest targets fixes its coefficient, and the intermediate energies enter through the first correction. By the sampling times, this initially constrained motion has developed into broad Hamming-space spreading together with a strong preference for low-energy configurations. For the hardest and seeds, the window probabilities exceed the bound for every isotropic kick distribution by factors of approximately 293 and 203. The free-mixer bounds give an additional direct measure of the role played by the Bernasconi potential. Energy conservation explains how the mean cost energy remains below the uniform-string value during this broad spreading. The canonical cost comparison further separates the effect of this low mean energy from the selection generated by the dynamics. At , the ground-state weight rises from below its mean-matched canonical value at to an enhancement of about 1.9 at , while substantial weight persists at the initial trap energy.
The field scan brings the geometry and energy dependence together. At , the mean displacement remains at 43%–84% of , falling to at most 54% for . At it exceeds 85% of , while the low-energy yield remains enhanced. At , the displacement approaches and the escape probability falls. The operating field thus lies just beyond the rapid increase in spreading, where spatial exploration and energetic preference work together. The later overtaking by shows that the optimum also depends on the duration of the move. Even as the blind target fraction falls to at , the measured window probabilities remain of order – for the hardest traps.
Turning from the escape mechanism to its implementation, the ratio characterizes sample efficiency, while state preparation and evolution determine the cost of generating each sample. Full-statevector simulation requires resources exponential in and reaches about on a single GPU in this study. The circuit estimate identifies the two- and four-body phases as the main hardware cost. These resource requirements give a concrete target for improved simulation and compilation while the present unitary dynamics establishes the escape mechanism.
These observations also suggest a natural next analytical step, namely to connect the Bernasconi spectrum near to the transfer rate and its field dependence, building on the population-transfer treatments of Refs. [37, 38]. The short-time expansion and the bounds in Eqs. (16) and (30) already constrain this development using the existing trajectories. In addition, time-dependent fields offer a way to combine rapid initial spreading with the larger late-time yield seen at weaker coupling. Matrix-product-state evolution and variational time-evolution methods provide possible routes to larger instances, with their cost set by the entanglement generated during the escape [60, 61, 62]. Finally, sample-based diagonalization and Krylov methods help clarify the role of the sampled states [63, 64]. For the diagonal cost Hamiltonian used here, its restriction to a span of sampled computational-basis states has those sampled energies as its eigenvalues. Its minimum is therefore the best energy already sampled, which also remained unchanged when descent was applied to the lowest-energy samples in the present runs. The quantum contribution lies in the distribution that makes these low-energy configurations available to the classical search.
Acknowledgements
The authors thank the Sharanga HPC cluster for providing the computational facility.
RB thanks support from an intramural grant CDRF C2/24/282 from BITS Pilani, and from ANRF India through ANRF/ARG/2025/009241/PS. RB acknowledges stimulating discussions with Rajatj Majumdar at IBM India and Indrakshi Raychowdhury on quantum algorithms, which motivated the present work.
ChatGPT Astra (OpenAI) was used to assist in framing the logical flow of the manuscript, and Gemini 3.8 Flash (Google) was used to assist in writing and maintaining the simulation and analysis code. The authors take responsibility for the final results and conclusions.
The code and numerical data are hosted on GitHub at github.com/ssk1001/bernasconi-krylov, with access provided upon request to the authors.
References
- [1]M. Golay, “Sieves for low autocorrelation binary sequences,” IEEE Transactions on Information Theory, vol. 23, pp. 43–51, 1977. doi:10.1109/tit.1977.1055653DOI
- [2]M. Golay, “The merit factor of long low autocorrelation binary sequences (Corresp.),” IEEE Transactions on Information Theory, vol. 28, pp. 543–549, 1982. doi:10.1109/tit.1982.1056505DOI
- [3]J. Jedwab, “A Survey of the Merit Factor Problem for Binary Sequences,” in Sequences and Their Applications—SETA 2004, Lecture Notes in Computer Science, vol. 3486, Springer, pp. 30–55, 2005. doi:10.1007/11423461_2DOI
- [4]J. Jedwab, D. J. Katz, and K.-U. Schmidt, “Advances in the merit factor problem for binary sequences,” Journal of Combinatorial Theory, Series A, vol. 120, pp. 882–906, 2013. doi:10.1016/j.jcta.2013.01.010DOI
- [5]K.-U. Schmidt, “Sequences with small correlation,” Designs, Codes and Cryptography, vol. 78, pp. 237–267, 2016. doi:10.1007/s10623-015-0154-7DOI
- [6]J. Bernasconi, “Low autocorrelation binary sequences: statistical mechanics and configuration space analysis,” Journal of Physique, vol. 48, pp. 559–567, 1987. doi:10.1051/jphys:01987004804055900
- [7]J. P. Bouchaud and M. Mézard, “Self induced quenched disorder: a model for the glass transition,” Journal of Statistical Physics I, vol. 4, pp. 1109–1114, 1994. doi:10.1051/jp1:1994240DOI
- [8]E. Marinari, G. Parisi, and F. Ritort, “Replica field theory for deterministic models. II. A non-random spin glass with glassy behaviour,” Journal of Physics A: Mathematical and General, vol. 27, pp. 7647–7668, 1994. doi:10.1088/0305-4470/27/23/011DOI
- [9]S. Mertens and C. Bessenrodt, “On the ground states of the Bernasconi model,” Journal of Physics A: Mathematical and General, vol. 31, pp. 3731–3749, 1998. doi:10.1088/0305-4470/31/16/004DOI
- [10]F. F. Ferreira, J. F. Fontanari, and P. F. Stadler, “Landscape statistics of the low-autocorrelation binary string problem,” Journal of Physics A: Mathematical and General, vol. 33, pp. 8635–8647, 2000. doi:10.1088/0305-4470/33/48/304DOI
- [11]S. Mertens, “Exhaustive search for low-autocorrelation binary sequences,” Journal of Physics A: Mathematical and General, vol. 29, pp. L473–L481, 1996. doi:10.1088/0305-4470/29/18/005DOI
- [12]T. Packebusch and S. Mertens, “Low autocorrelation binary sequences,” Journal of Physics A: Mathematical and Theoretical, vol. 49, p. 165001, 2016. doi:10.1088/1751-8113/49/16/165001DOI
- [13]S. Prestwich, “Exploiting relaxation in local search for LABS,” Annals of Operations Research, vol. 156, pp. 129–141, 2007. doi:10.1007/s10479-007-0226-9DOI
- [14]J. E. Gallardo, C. Cotta, and A. J. Fernández, “Finding low autocorrelation binary sequences with memetic algorithms,” Applied Soft Computing, vol. 9, pp. 1252–1262, 2009. doi:10.1016/j.asoc.2009.03.005DOI
- [15]B. Bošković, F. Brglez, and J. Brest, “Low-autocorrelation binary sequences: On improved merit factors and runtime predictions to achieve them,” Applied Soft Computing, vol. 56, pp. 262–285, 2017. doi:10.1016/j.asoc.2017.02.024DOI
- [16]B. Bošković, J. Herzog, and J. Brest, “Parallel self-avoiding walks for a low-autocorrelation binary sequences problem,” Journal of Computational Science, vol. 77, p. 102260, 2024. doi:10.1016/j.jocs.2024.102260
- [17]T. Koch, D. E. Bernal Neira, Y. Chen et al., “The Quantum Optimization Benchmarking Library,” Nature Computational Science, vol. 6, pp. 653–671, 2026. doi:10.1038/s43588-026-00991-1DOI
- [18]H. R. Lourenço, O. C. Martin, and T. Stützle, “Iterated Local Search,” in Handbook of Metaheuristics, vol. 57, Kluwer Academic Publishers, pp. 320–353, 2003. doi:10.1007/0-306-48056-5_11
- [19]D. J. Wales and J. P. K. Doye, “Global Optimization by Basin-Hopping and the Lowest Energy Structures of Lennard-Jones Clusters Containing up to 110 Atoms,” The Journal of Physical Chemistry A, vol. 101, pp. 5111–5116, 1997. doi:10.1021/jp970984nDOI
- [20]R. Martí, “Multi-Start Methods,” in Handbook of Metaheuristics, vol. 57, Kluwer, pp. 355–368, 2003. doi:10.1007/0-306-48056-5_12
- [21]S. Kirkpatrick, C. D. Gelatt, and M. P. Vecchi, “Optimization by Simulated Annealing,” Science, vol. 220, pp. 671–680, 1983. doi:10.1126/science.220.4598.671DOI
- [22]A. B. Finnila, M. A. Gomez, C. Sebenik, C. Stenson, and J. D. Doll, “Quantum annealing: A new method for minimizing multidimensional functions,” Chemical Physics Letters, vol. 219, pp. 343–348, 1994. doi:10.1016/0009-2614(94)00117-0DOI
- [23]T. Kadowaki and H. Nishimori, “Quantum annealing in the transverse Ising model,” Physical Review E, vol. 58, pp. 5355–5363, 1998. doi:10.1103/physreve.58.5355DOI
- [24]E. Farhi, J. Goldstone, S. Gutmann, J. Lapan, A. Lundgren, and D. Preda, “A Quantum Adiabatic Evolution Algorithm Applied to Random Instances of an NP-Complete Problem,” Science, vol. 292, pp. 472–475, 2001. doi:10.1126/science.1057726DOI
- [25]G. E. Santoro, R. Martoňák, E. Tosatti, and R. Car, “Theory of Quantum Annealing of an Ising Spin Glass,” Science, vol. 295, pp. 2427–2430, 2002. doi:10.1126/science.1068774DOI
- [26]V. S. Denchev, S. Boixo, S. V. Isakov, N. Ding, R. Bab-bush, V. Smelyanskiy, J. Martinis, and H. Neven, “What is the Computational Value of Finite-Range Tunneling?,” Physical Review X, vol. 6, p. 031015, 2016. doi:10.1103/physrevx.6.031015DOI
- [27]T. Albash and D. A. Lidar, “Adiabatic quantum computation,” Reviews of Modern Physics, vol. 90, p. 015002, 2018. doi:10.1103/revmodphys.90.015002DOI
- [28]P. Hauke, H. G. Katzgraber, W. Lechner, H. Nishimori, and W. D. Oliver, “Perspectives of quantum annealing: methods and implementations,” Reports on Progress in Physics, vol. 83, p. 054401, 2020. doi:10.1088/1361-6633/ab85b8DOI
- [29]B. Altshuler, H. Krovi, and J. Roland, “Anderson localization makes adiabatic quantum optimization fail,” Proceedings of the National Academy of Sciences, vol. 107, pp. 12446–12450, 2010. doi:10.1073/pnas.1002116107DOI
- [30]E. Farhi and S. Gutmann, “Quantum computation and decision trees,” Physical Review A, vol. 58, pp. 915–928, 1998. doi:10.1103/physreva.58.915DOI
- [31]A. M. Childs and J. Goldstone, “Spatial search by quantum walk,” Physical Review A, vol. 70, p. 022314, 2004. doi:10.1103/physreva.70.022314DOI
- [32]O. Mülken and A. Blumen, “Continuous-time quantum walks: Models for coherent transport on complex networks,” Physics Reports, vol. 502, pp. 37–87, 2011. doi:10.1016/j.physrep.2011.01.002
- [33]C. Moore and A. Russell, “Quantum Walks on the Hypercube,” in Randomization and Approximation Techniques in Computer Science, Lecture Notes in Computer Science, vol. 2483, Springer, pp. 164–178, 2002. doi:10.1007/3-540-45726-7_14DOI
- [34]A. Callison, N. Chancellor, F. Mintert, and V. Kendon, “Finding spin glass ground states using quantum walks,” New Journal of Physics, vol. 21, p. 123022, 2019. doi: 10.1088/1367-2630/ab5ca2DOI
- [35]A. Callison, M. Fenstentein, J. Chen, L. Nita, V. Kendon, and N. Chancellor, “Energetic Perspective on Rapid Quenches in Quantum Annealing,” PRX Quantum, vol. 2, p. 010338, 2021. doi: 10.1103/prxquantum.2.010338DOI
- [36]K. Kechedzhi, V. Smelyanskiy, J. R. McClean, V. Denchev, M. Mohseni, S. Isakov, S. Boixo, B. Altshuler, and H. Neven, “Efficient population transfer via non-ergodic extended states in quantum spin glass,” arXiv:1807.04792, 2018.arxiv.org/abs/1807.04792
- [37]V. N. Smelyanskiy, K. Kechedzhi, S. Boixo, S. V. Isakov, H. Neven, and B. Altshuler, “Nonergodic Delocalized States for Efficient Population Transfer within a Narrow Band of the Energy Landscape,” Physical Review X, vol. 10, p. 011017, 2020. doi:10.1103/physrevx.10.011017DOI
- [38]C. L. Baldwin and C. R. Laumann, “Quantum algorithm for energy matching in hard optimization problems,” Physical Review B, vol. 97, p. 224201, 2018. doi:10.1103/physrevb.97.224201DOI
- [39]C. R. Laumann, A. Pal, and A. Scardicchio, “Many-Body Mobility Edge in a Mean-Field Quantum Spin Glass,” Physical Review Letters, vol. 113, p. 200405, 2014. doi:10.1103/physrevlett.113.200405DOI
- [40]C. L. Baldwin, C. R. Laumann, A. Pal, and A. Scardicchio, “The many-body localized phase of the quantum random energy model,” Physical Review B, vol. 93, p. 024202, 2016. doi:10.1103/physrevb.93.024202DOI
- [41]L. Faoro, M. V. Feigel’man, and L. Ioffe, “Non-ergodic extended phase of the Quantum Random Energy model,” Annals of Physics, vol. 409, p. 167916, 2019. doi:10.1016/j.aop.2019.167916DOI
- [42]T. Parolini and G. Mossi, “Multifractal Dynamics of the QREM,” arXiv:2007.00315, 2020.arxiv.org/abs/2007.00315
- [43]N. Chancellor, “Modernizing quantum annealing using local searches,” New Journal of Physics, vol. 19, p. 023024, 2017. doi:10.1088/1367-2630/aa59c4DOI
- [44]M. Ohkuwa, H. Nishimori, and D. A. Lidar, “Reverse annealing for the fully connected p-spin model,” Physical Review A, vol. 98, p. 022314, 2018. doi: 10.1103/physreva.98.022314DOI
- [45]D. Layden, G. Mazzola, R. V. Mishmash, M. Motta, P. Wojcian, J.-S. Kim, and S. Sheldon, “Quantum-enhanced Markov chain Monte Carlo,” Nature, vol. 619, pp. 282–287, 2023. doi:10.1038/s41586-023-06095-4
- [46]D. J. Egger, J. Mareček, and S. Woerner, “Warm-starting quantum optimization,” Quantum, vol. 5, p. 479, 2021. doi:10.22331/q-2021-06-17-479DOI
- [47]E. Farhi, J. Goldstone, and S. Gutmann, “A Quantum Approximate Optimization Algorithm,” arXiv:1411.4028, 2014.arxiv.org/abs/1411.4028
- [48]L. Zhou, S.-T. Wang, S. Choi, H. Pichler, and M. D. Lukin, “Quantum Approximate Optimization Algorithm: Performance, Mechanism, and Implementation on Near-Term Devices,” Physical Review X, vol. 10, p. 021067, 2020. doi:10.1103/physrevx.10.021067DOI
- [49]R. Shaydulin, C. Li, S. Chakrabarti et al., “Evidence of scaling advantage for the quantum approximate optimization algorithm on a classically intractable problem,” Science Advances, vol. 10, p. eadm6761, 2024. doi:10.1126/sciadv.adm6761DOI
- [50]M. Sciorilli, G. Camilo, T. O. Maciel, A. Canabarro, L. Borges, and L. Aolita, “A competitive NISQ and qubit-efficient solver for the LABS problem,” arXiv:2506.17391, 2025.arxiv.org/abs/2506.17391
- [51]NVIDIA, “cuStateVec: A High-Performance Library for State Vector Quantum Simulators,” cuQuantum SDK documentation, accessed November 10, 2026.
- [52]H. F. Trotter, “On the product of semi-groups of operators,” Proceedings of the American Mathematical Society, vol. 10, pp. 545–551, 1959. doi:10.1090/S0002-9939-1959-0108732-6DOI
- [53]G. Strang, “On Construction and Comparison of Difference Schemes,” SIAM Journal on Numerical Analysis, vol. 5, pp. 506–517, 1968. doi:10.1137/0705041DOI
- [54]M. Suzuki, “Generalized Trotter’s formula and systematic approximants of exponential operators and inner derivations with applications to many-body problems,” Communications in Mathematical Physics, vol. 51, pp. 183–190, 1976. doi:10.1007/bf01609348DOI
- [55]A. M. Childs, Y. Su, M. C. Tran, N. Wiebe, and S. Zhu, “Theory of Trotter Error with Commutator Scaling,” Physical Review X, vol. 11, p. 011020, 2021. doi:10.1103/physrevx.11.011020DOI
- [56]F. Glover and M. Laguna, Tabu Search, Kluwer Academic Publishers, 1997. doi:10.1007/978-1-4615-6089-0DOI
- [57]B. Efron and R. J. Tibshirani, An Introduction to the Bootstrap, Chapman & Hall/CRC, 1994. doi:10.1201/9780429246593DOI
- [59]B. E. Szigeti, G. Homa, Z. Zimborás, and N. Barankai, “Short-time behavior of continuous-time quantum walks on graphs,” Physical Review A, vol. 100, p. 062320, 2019. doi:10.1103/physreva.100.062320DOI
- [60]U. Schollwöck, “The density-matrix renormalization group in the age of matrix product states,” Annals of Physics, vol. 326, pp. 96–192, 2011. doi:10.1016/j.aop.2010.09.012DOI
- [61]J. Haegeman, J. I. Cirac, T. J. Osborne, I. Pižorn, H. Verschelde, and F. Verstraete, “Time-Dependent Variational Principle for Quantum Lattices,” Physical Review Letters, vol. 107, p. 070601, 2011. doi:10.1103/physrevlett.107.070601DOI
- [62]S. Paeckel, T. Köhler, A. Swobodа, S. R. Manmana, U. Schlöckow, and C. Hubig, “Time-evolution methods for matrix-product states,” Annals of Physics, vol. 411, p. 167998, 2019. doi:10.1016/j.aop.2019.167998DOI
- [63]J. Robledo-Moreno, M. Motta, H. Haas et al., “Chemistry beyond the scale of exact diagonalization on a quantum-centric supercomputer,” Science Advances, vol. 11, p. eadu9991, 2025. doi:10.1126/sciadv.adu9991DOI
- [64]J. Yu, J. Robledo Moreno, J. T. Iosue et al., “Quantum-Centric Algorithm for Sample-Based Krylov Diagonalization,” arXiv:2501.09702, 2025.arxiv.org/abs/2501.09702
- [65]E. T. Jaynes, “Information Theory and Statistical Mechanics,” Phys. Rev. 106, 620–630 (1957). doi:10.1103/PhysRev.106.620.DOI
- [66]L. D’Alessio, Y. Kafri, A. Polkovnikov, and M. Rigol, “From quantum chaos and eigenstate thermalization to statistical mechanics and thermodynamics,” Adv. Phys. 65, 239–362 (2016). doi:10.1080/00018732.2016.1198134.DOI





