Scalable Parameter Init for QAOA (SPIQ)

• SPIQ is a scalable methodology for initializing QAOA parameters by exploiting structure and symmetries to reduce search complexity.
• It leverages depth-progressive warm starts, annealing-inspired schedules, and learned predictors to enhance convergence and avoid poor local optima.
• Practical implementations show improved approximation quality and significant reductions in classical optimization overhead for combinatorial tasks.

Scalable Parameter Initialization for QAOA (SPIQ) denotes a class of strategies for choosing QAOA variational angles so that optimization remains effective as circuit depth, problem size, or architectural complexity increase. Across the literature, the central objective is consistent: reduce the classical burden of searching over $2p$ or more parameters, avoid convergence to poor local optima, and preserve or improve approximation quality at larger depth $p$. In this sense, SPIQ encompasses depth-progressive warm starts, annealing-derived schedules, symmetry-bounded extrapolation rules, topology-aware learned predictors, multilevel transfer mechanisms, and classical pre-search over relaxed ansätze (Lee et al., 2021, Lee et al., 2022, Bharadwaj et al., 15 Feb 2026).

1. Variational setting and initialization problem

For a combinatorial optimization problem specified by a cost Hamiltonian $H_C$ and a mixer Hamiltonian $H_B$, the $p$-layer QAOA state is

$$|\psi_p(\gamma,\beta)\rangle = \prod_{\ell=1}^{p} e^{-i \beta_\ell H_B} e^{-i \gamma_\ell H_C} |s\rangle,$$

where $\gamma = (\gamma_1,\dots,\gamma_p)$, $\beta = (\beta_1,\dots,\beta_p)$, and $|s\rangle = |+\rangle^{\otimes n}$ is the uniform superposition over computational basis states. For Max-Cut on a graph $G=(V,E)$,

$p$0

The variational objective is the energy of $p$1,

$p$2

with optimization problem

$p$3

and Max-Cut approximation ratio

$p$4

Equivalent notation using $p$5 is also standard in the literature (Lee et al., 2021, Lee et al., 2022).

Initialization is consequential because the QAOA optimization landscape becomes increasingly multimodal at larger $p$6, and random starts often converge to false local minima or false local maxima. This behavior was visualized explicitly for MaxCut on random graphs, where random initialization was shown to be prone to converging to local minima with sub-optimal performance, and the best result among many random starts improved only slowly as the number of starts grew (Sack et al., 2021). A recurring conclusion across SPIQ methods is therefore that high-quality seeds can be as important as the subsequent local optimizer.

2. Landscape structure, symmetries, and bounded search domains

A major line of SPIQ research exploits the fact that QAOA parameter landscapes are not arbitrary. For unweighted Max-Cut, the expectation has angle-reversal symmetry,

$p$7

together with periodicities that imply redundant optima. For general unweighted graphs, canonical domains are

$p$8

while for unweighted regular graphs,

$p$9

Odd-regular graphs additionally exhibit a symmetry that mixes $H_C$0 with alternating-index transformations of $H_C$1 (Lee et al., 2022).

Once these bounds remove redundancy, near-optimal parameters display two empirical regularities. First, within a fixed depth $H_C$2, $H_C$3 tends to increase and $H_C$4 tends to decrease with the layer index $H_C$5, in a pattern reminiscent of discretized adiabatic evolution: $H_C$6 Second, optimal angles are depth-nonstationary: as $H_C$7 increases, $H_C$8 decreases and $H_C$9 increases for fixed $H_B$0. This directly implies that “optimal at depth $H_B$1” is not optimal at depth $H_B$2, which is why layerwise freezing is generally suboptimal (Lee et al., 2022).

Landscape studies also provide a mechanistic account of why SPIQ works. Under parameters fixing, fixing earlier parameters at values associated with high $H_B$3 produces $H_B$4 slices with expanded high-value regions, and isolated maxima in the $H_B$5 slice empirically morph into “maximum lines” as $H_B$6 increases. This reduces trapping risk and facilitates convergence (Lee et al., 2021). A plausible implication is that SPIQ is not merely a heuristic speedup; it is a method for altering the effective local geometry seen by the optimizer.

3. Depth-progressive warm starts

Depth-progressive initialization is the most direct form of SPIQ. In the “parameters fixing strategy,” the best parameters found at depth $H_B$7 are used as the initialization for depth $H_B$8, and only the new layer’s parameters are sampled randomly: $H_B$9 with $p$0 and $p$1. Although termed “parameters fixing,” earlier-layer parameters are not frozen in practice; they are re-optimized and empirically change only slightly across depths (Lee et al., 2021).

In that study, the optimization used Nelder–Mead with max function evaluations $p$2, absolute tolerance $p$3, and $p$4 random initializations per depth. The total classical effort scales as $p$5 local runs. On 3-regular graphs with $p$6, mean $p$7 showed a clear increasing trend with $p$8, and mean $p$9 was achieved for $$|\psi_p(\gamma,\beta)\rangle = \prod_{\ell=1}^{p} e^{-i \beta_\ell H_B} e^{-i \gamma_\ell H_C} |s\rangle,$$0 across tested $$|\psi_p(\gamma,\beta)\rangle = \prod_{\ell=1}^{p} e^{-i \beta_\ell H_B} e^{-i \gamma_\ell H_C} |s\rangle,$$1. On Erdős–Rényi graphs $$|\psi_p(\gamma,\beta)\rangle = \prod_{\ell=1}^{p} e^{-i \beta_\ell H_B} e^{-i \gamma_\ell H_C} |s\rangle,$$2 with $$|\psi_p(\gamma,\beta)\rangle = \prod_{\ell=1}^{p} e^{-i \beta_\ell H_B} e^{-i \gamma_\ell H_C} |s\rangle,$$3, mean $$|\psi_p(\gamma,\beta)\rangle = \prod_{\ell=1}^{p} e^{-i \beta_\ell H_B} e^{-i \gamma_\ell H_C} |s\rangle,$$4 was achieved for $$|\psi_p(\gamma,\beta)\rangle = \prod_{\ell=1}^{p} e^{-i \beta_\ell H_B} e^{-i \gamma_\ell H_C} |s\rangle,$$5 across tested $$|\psi_p(\gamma,\beta)\rangle = \prod_{\ell=1}^{p} e^{-i \beta_\ell H_B} e^{-i \gamma_\ell H_C} |s\rangle,$$6. By contrast, the random-initialization baseline often stagnated or degraded for $$|\psi_p(\gamma,\beta)\rangle = \prod_{\ell=1}^{p} e^{-i \beta_\ell H_B} e^{-i \gamma_\ell H_C} |s\rangle,$$7, and its standard deviation often increased with depth (Lee et al., 2021).

A more aggressive progressive variant is the bilinear depth-prediction rule. Given optimized parameters at depths $$|\psi_p(\gamma,\beta)\rangle = \prod_{\ell=1}^{p} e^{-i \beta_\ell H_B} e^{-i \gamma_\ell H_C} |s\rangle,$$8 and $$|\psi_p(\gamma,\beta)\rangle = \prod_{\ell=1}^{p} e^{-i \beta_\ell H_B} e^{-i \gamma_\ell H_C} |s\rangle,$$9, it predicts depth-$\gamma = (\gamma_1,\dots,\gamma_p)$0 parameters by extrapolation. For $\gamma = (\gamma_1,\dots,\gamma_p)$1,

$\gamma = (\gamma_1,\dots,\gamma_p)$2

with special boundary rules for $\gamma = (\gamma_1,\dots,\gamma_p)$3 and $\gamma = (\gamma_1,\dots,\gamma_p)$4, followed by projection to the canonical parameter domain. This method requires only one bounded optimization per depth after seeding $\gamma = (\gamma_1,\dots,\gamma_p)$5. On Max-Cut over 30 non-isomorphic instances up to $\gamma = (\gamma_1,\dots,\gamma_p)$6, it tracked the optimal approximation ratio found by parameters fixing with minimal error, while reducing $\gamma = (\gamma_1,\dots,\gamma_p)$7 by $\gamma = (\gamma_1,\dots,\gamma_p)$8–$\gamma = (\gamma_1,\dots,\gamma_p)$9 for $\beta = (\beta_1,\dots,\beta_p)$0 because parameters fixing was measured as totals over $\beta = (\beta_1,\dots,\beta_p)$1 trials per depth and the bilinear method used one trial (Lee et al., 2022).

These two schemes articulate complementary SPIQ principles. Parameters fixing emphasizes multistart only on the new layer; bilinear prediction emphasizes single-shot extrapolation after symmetry reduction. Both rely on the same empirical fact: shallow-depth optima remain near-optimal when embedded into deeper circuits.

4. Schedule-based and analytically derived initialization

Another major SPIQ lineage derives QAOA parameters from adiabatic or annealing schedules. In the Trotterized quantum annealing (TQA) approach, one discretizes a linear schedule

$\beta = (\beta_1,\dots,\beta_p)$2

so that

$\beta = (\beta_1,\dots,\beta_p)$3

which yields

$\beta = (\beta_1,\dots,\beta_p)$4

For unweighted 3-regular graphs at $\beta = (\beta_1,\dots,\beta_p)$5, a broad interval $\beta = (\beta_1,\dots,\beta_p)$6 yielded post-optimization approximation ratios within $\beta = (\beta_1,\dots,\beta_p)$7 of optimal, and the optimal step size was approximately $\beta = (\beta_1,\dots,\beta_p)$8. A single TQA-initialized run matched the best outcome among $\beta = (\beta_1,\dots,\beta_p)$9 random initializations, and for Erdős–Rényi graphs it could even outperform the best random initialization (Sack et al., 2021).

A more radical analytic variant is linear-ramp QAOA (LR-QAOA), which fixes

$|s\rangle = |+\rangle^{\otimes n}$0

With Hamiltonian normalization and default endpoints $|s\rangle = |+\rangle^{\otimes n}$1, $|s\rangle = |+\rangle^{\otimes n}$2, the same schedule was used across nine combinatorial optimization problems. Simulations up to $|s\rangle = |+\rangle^{\otimes n}$3 qubits and $|s\rangle = |+\rangle^{\otimes n}$4 layers suggested

$|s\rangle = |+\rangle^{\otimes n}$5

For 42-qubit weighted MaxCut, the random baseline $|s\rangle = |+\rangle^{\otimes n}$6 increased to an average of $|s\rangle = |+\rangle^{\otimes n}$7 at $|s\rangle = |+\rangle^{\otimes n}$8, and an observed case reached $|s\rangle = |+\rangle^{\otimes n}$9 at $G=(V,E)$0. Hardware demonstrations extended to $G=(V,E)$1 qubits and $G=(V,E)$2, with approximately $G=(V,E)$3 CNOT gates (Montanez-Barrera et al., 2024).

Adiabatic-passage-based parameter setting for 3-SAT pushes the analytic approach further by normalizing $G=(V,E)$4 and $G=(V,E)$5 so that parameter magnitudes are decoupled from instance scale. This yields a problem-independent initialization

$G=(V,E)$6

followed by a continuity-aware reparameterization in $G=(V,E)$7. On satisfiable random 3-SAT near threshold, the reported average number of expectation evaluations remained in the range $G=(V,E)$8–$G=(V,E)$9 over $p$00, whereas FOURIER showed superlinear growth and TQA showed roughly linear growth (Wu et al., 2023).

Gap-informed schedule transfer reduces the variational search dimension to two global hyperparameters. Learning a spectral-gap profile $p$01 on small instances, it defines

$p$02

and discretizes the schedule to

$p$03

In experiments on random QUBO and 3-regular MaxCut, gap profiles learned at $p$04 transferred effectively to $p$05, and the classical optimization problem was compressed from $p$06 parameters to $p$07 (Nzongani et al., 16 Feb 2026).

In a different setting, ground-state preparation for the fully-connected $p$08-spin ferromagnet showed that schedule-derived initialization can be decisive when $p$09. There, a digitized QA warm start was necessary to obtain good performance at small $p$10, while for $p$11 the landscape simplified and all minima became degenerate, with $p$12 for even $p$13 and $p$14 for odd $p$15 (Wauters et al., 2020).

5. Learned, generative, and topology-aware SPIQ

A second broad SPIQ family learns mappings from instance structure or prior optimized runs to good parameter seeds. For MaxCut on Erdős–Rényi graphs, a simple three-layer fully connected network took the upper-triangular adjacency matrix as input and predicted all $p$16 angles in one forward pass. Training used $p$17 graphs per setup combination, with labels obtained by BFGS optimization seeded by TQA. On $p$18, $p$19 random-ER graphs, the neural initializer converged in about one iteration, whereas TQA required about $p$20–$p$21 iterations to reach the same approximation ratio; across benchmarks it yielded up to $p$22 fewer optimization iterations and was presented as effectively realizing an iterative-free QAOA scheme (Amosy et al., 2022).

The hybrid GRU-CNN bilinear method decomposes the task by depth. A GRU generates favorable parameters for $p$23, a CNN maps optimized depth-1 angles to a depth-2 prediction, and a bilinear recurrence extrapolates beyond $p$24. On Erdős–Rényi graphs, it forecast parameters up to depth $p$25 and reported an approximation ratio of $p$26 across $p$27 qubits, while on edge-probability $p$28 graphs it reported approximately $p$29 at $p$30, $p$31 at $p$32, and $p$33 at $p$34 for $p$35 (Xu et al., 2023).

Conditional diffusion treats parameter selection as generative modeling. A denoising diffusion probabilistic model conditioned on Graph2Vec embeddings was trained on $p$36 graphs with $p$37. After $p$38 optimization iterations, average approximation-ratio improvements over random initialization were reported as $p$39 for random graphs, $p$40 for regular graphs, and $p$41 for Watts–Strogatz graphs, with maxima $p$42, $p$43, and $p$44, respectively. Training on $p$45 and testing on $p$46, the same model yielded up to $p$47 and $p$48 average improvement on larger graphs (Meng et al., 2024).

Neural QAOA$p$49 extends SPIQ to divide-and-conquer settings where qubit limits force graph partitioning. Its generative evaluative network jointly learns partitions and initial parameters, using topology features such as degree, weighted degree, clustering coefficient, PageRank, and betweenness centrality, together with a differentiable quantum evaluator. On $p$50 QUBO, Ising, and MaxCut instances from $p$51 to $p$52 variables, it ranked first on $p$53 instances, achieved best overall average rank $p$54 on out-of-distribution distributions, and preserved strong ranks under zero-shot generalization (Zheng et al., 13 May 2026).

These learned methods differ in conditioning and training cost, but they share a common SPIQ premise: parameter initialization can itself be modeled as a transferable prediction problem rather than a fresh local search on every instance.

6. Relaxed ansätze, multilevel transfer, and system-level frameworks

Several recent SPIQ methods modify the search space instead of only predicting within the standard $p$55-angle parametrization. In scalable Clifford-based initialization, a relaxed multi-angle QAOA ansatz is restricted to discrete Clifford angle sets,

$p$56

so that candidate states remain stabilizer states and can be evaluated classically. A genetic algorithm searches these “Clifford points,” after which selected points seed continuous QAOA. Reported gains include up to $p$57 ground-state overlap on moderate instances, up to $p$58 absolute accuracy improvement over state-of-the-art initialization, and up to $p$59 reduction in initial-state diversity across QUBO, PUBO, and PCBO problems spanning tens to hundreds of qubits (Bharadwaj et al., 15 Feb 2026).

A related idea appears in multi-angle QAOA for $p$60. There, the best MA-QAOA initializer was “QAOA Relax”: first optimize standard QAOA at depth $p$61, then copy each scalar $p$62 uniformly across all cost-term and mixer-term angles at that layer, and finally release them for full MA-QAOA optimization. On the 9-node dataset, QAOA Relax reached $p$63 on average at $p$64 and achieved $p$65 in the worst case at $p$66. Across datasets, MA-QAOA reduced required depth by factors from $p$67 to $p$68 relative to standard QAOA, although it was not always optimal for minimizing total QPU time (Gaidai et al., 2023).

Multilevel QAOA realizes SPIQ within a hierarchy for a single large instance. A large QUBO is coarsened into progressively smaller graphs, low-depth QAOA parameters are computed on the coarsest level, and those parameters are transferred unchanged down the hierarchy while refining bitstrings with Quantum Relax-and-Round and a genetic algorithm. The similarity score between consecutive $p$69 landscapes satisfied $p$70 for most consecutive levels, with the coarsest transition sometimes $p$71, supporting the claim that relaxation-based coarsening preserves the landscape features needed for parameter transfer (Bach et al., 16 May 2025).

At the systems level, SPIQ has also been implemented as a service layer. Stone-in-Waiting is a cloud-based accelerator that integrates exact matching, parameter-based approximate graph matching, factor-based approximate graph matching, and formula generation, together with database updates and reverse parameter submission. On the hackathon benchmark, it improved the Baseline Algorithm’s total score by $p$72, from $p$73 to $p$74 (Zeng, 20 Mar 2026).

7. Limitations, trade-offs, and open problems

SPIQ methods improve optimization reliability, but their assumptions differ sharply. Parameters fixing requires $p$75 local optimizations and was evaluated only with Nelder–Mead, on 3-regular and Erdős–Rényi graphs, in noiseless Qiskit Aer simulation; robustness under NISQ noise, finite shots, and hardware constraints was not studied (Lee et al., 2021). Bilinear depth prediction removes multistart overhead, but it depends on high-quality solutions at $p$76 and $p$77, and mis-specified canonical bounds can lead to non-adiabatic starts that degrade prediction beyond moderate depth (Lee et al., 2022).

Schedule-based methods trade expressivity for compression. LR-QAOA eliminates outer-loop optimization but can require modest tuning of $p$78 as $p$79 grows and may underperform on edge-case instances with dense near-optimal degeneracy (Montanez-Barrera et al., 2024). Adiabatic-passage and gap-informed methods depend on meaningful spectral-range normalization or transferable gap profiles; this suggests that their success is strongest when the instance class has stable ensemble-level spectral structure rather than highly irregular, instance-specific bottlenecks (Wu et al., 2023, Nzongani et al., 16 Feb 2026).

Learned SPIQ methods introduce distribution-shift risks. The fully connected neural predictor was evaluated on Erdős–Rényi graphs and did not enforce permutation invariance (Amosy et al., 2022). Conditional diffusion was trained only at $p$80 and on specific graph families (Meng et al., 2024). Neural QAOA$p$81 relies on a surrogate evaluator; if that evaluator is poorly calibrated, generator optimization can degrade, although the reported Pearson’s $p$82 on unseen test points indicates strong alignment in the studied setting (Zheng et al., 13 May 2026).

Relaxed-ansatz and pre-search frameworks likewise have nontrivial scope conditions. Clifford-restricted search may miss high-quality non-Clifford basins and depends on subsequent continuous refinement (Bharadwaj et al., 15 Feb 2026). MA-QAOA reduces depth but may increase total QPU time because its parameter count is $p$83 rather than $p$84 (Gaidai et al., 2023). Multilevel transfer assumes that coarsening preserves QAOA-relevant structure, which was empirically supported for the tested instances but not elevated to a general theorem (Bach et al., 16 May 2025).

Open questions identified across the literature include analytic characterization of angle trajectories at high $p$85, extending symmetry-bounded domains to weighted or irregular cases, integrating hardware-aware surrogates and noise models, scaling learned initializers beyond narrow training distributions, and establishing theoretical guarantees for discrete pre-search spaces such as Clifford-restricted multi-angle QAOA (Lee et al., 2022, Zheng et al., 13 May 2026, Bharadwaj et al., 15 Feb 2026).

Taken together, these works show that SPIQ is not a single algorithm but a design space. Its unifying principle is scalable reduction of QAOA parameter-search difficulty: by exploiting depth continuity, canonical symmetries, annealing structure, learned topology-to-angle maps, relaxed classical proxies, or multilevel transfer, SPIQ attempts to replace unguided search in a rugged $p$86-dimensional landscape with a structured initialization procedure that remains effective as $p$87, $p$88, and problem heterogeneity increase.