QAOA Angle Offline Optimization

• QAOA angle offline optimization is a family of techniques that precompute or compress parameters, reducing the need for costly online variational searches.
• These methods leverage fixed-angle schedules, analytic derivations from adiabatic evolution, and symmetry-aware strategies to streamline parameter setting.
• The approaches improve performance by mitigating local minima and barren plateau issues, enhancing convergence and reducing total QPU time.

QAOA angle offline optimization denotes the family of methods that determine, compress, initialize, transfer, or constrain Quantum Approximate Optimization Algorithm parameters before deployment, or with substantially reduced online interaction with the target quantum device. In the literature, this includes optimization-free fixed-angle schedules for regular MaxCut graphs, closed-form schedules derived from Trotterized quantum adiabatic evolution, distribution-level angle computation for Grover-driven QAOA, symmetry-aware parameter tying in multi-angle QAOA, learned angle transfer across instance families, and query-efficient classical search policies that use offline predictions to restrict the online search region (Wurtz et al., 2021, Yoshioka et al., 2023, Headley et al., 2022, Shi et al., 2022, Huynh, 27 Apr 2026).

1. Formal setting and motivation

For depth $p$, standard QAOA prepares

$$\ket{\psi_p(\bm{\gamma},\bm{\beta})} = \prod_{j=1}^p e^{-i\beta_j H_x}e^{-i\gamma_j H_z}\ket{+}^{\otimes n},$$

and for MaxCut one commonly uses

$H_z=\frac{1}{2}\sum_{(j,k)\in E}(\mathbbm{1}-Z_jZ_k),$

with objective

$$F_p(\bm{\gamma},\bm{\beta})=\langle\psi_p(\bm{\gamma},\bm{\beta})|H_z|\psi_p(\bm{\gamma},\bm{\beta})\rangle.$$

The approximation ratio is written as $\alpha=F_p(\bm{\gamma}^*,\bm{\beta}^*)/C_{\max}$ in one formulation, and as $AR=\langle C\rangle/C^*$ in another MaxCut formulation (Lee et al., 2022, Wilkie et al., 2024).

Offline optimization is motivated by the cost and instability of the classical outer loop. The QAOA landscape is described as containing many local minima, barren plateaus can appear at larger qubit count or depth, and bad initialization can deteriorate the quality of the results, especially at great circuit depths. On hardware, the difficulty is amplified by sampling cost: many optimizers used for QAOA require $N\gtrsim 1000$ shots per point, while some neutral-atom platforms have measurement times as long as $30$ ms and total shot rates around $5$ Hz (Cheng et al., 2023, Lee et al., 2022, Polloreno et al., 2022).

The burden is even greater in multi-angle QAOA. Standard QAOA has $2p$ parameters, whereas ma-QAOA assigns one angle to each term in the cost and mixer Hamiltonians, giving $$\ket{\psi_p(\bm{\gamma},\bm{\beta})} = \prod_{j=1}^p e^{-i\beta_j H_x}e^{-i\gamma_j H_z}\ket{+}^{\otimes n},$$0 trainable parameters for MaxCut. This often improves performance relative to vanilla QAOA, but it makes offline classical optimization much more expensive because there are many more angles to search over (Shi et al., 2022).

2. Optimization-free schedules and analytic angle design

One line of work replaces instance-by-instance variational search by precomputed fixed schedules. For MaxCut on regular graphs, the fixed angle conjecture states that if QAOA angles are chosen to be optimal for the tree subgraph, then those same angles should give a guaranteed approximation ratio on any graph of the same regular degree. The reported 3-regular fixed angles for $$\ket{\psi_p(\bm{\gamma},\bm{\beta})} = \prod_{j=1}^p e^{-i\beta_j H_x}e^{-i\gamma_j H_z}\ket{+}^{\otimes n},$$1 yield guarantees from $$\ket{\psi_p(\bm{\gamma},\bm{\beta})} = \prod_{j=1}^p e^{-i\beta_j H_x}e^{-i\gamma_j H_z}\ket{+}^{\otimes n},$$2 at $$\ket{\psi_p(\bm{\gamma},\bm{\beta})} = \prod_{j=1}^p e^{-i\beta_j H_x}e^{-i\gamma_j H_z}\ket{+}^{\otimes n},$$3 to $$\ket{\psi_p(\bm{\gamma},\bm{\beta})} = \prod_{j=1}^p e^{-i\beta_j H_x}e^{-i\gamma_j H_z}\ket{+}^{\otimes n},$$4 at $$\ket{\psi_p(\bm{\gamma},\bm{\beta})} = \prod_{j=1}^p e^{-i\beta_j H_x}e^{-i\gamma_j H_z}\ket{+}^{\otimes n},$$5, and under the conjecture the $$\ket{\psi_p(\bm{\gamma},\bm{\beta})} = \prod_{j=1}^p e^{-i\beta_j H_x}e^{-i\gamma_j H_z}\ket{+}^{\otimes n},$$6 guarantee exceeds the Goemans–Williamson value $$\ket{\psi_p(\bm{\gamma},\bm{\beta})} = \prod_{j=1}^p e^{-i\beta_j H_x}e^{-i\gamma_j H_z}\ket{+}^{\otimes n},$$7. The operational consequence is an optimization-free QAOA protocol in which the circuit is evaluated once at the offline-computed angles (Wurtz et al., 2021).

A second line derives angles analytically from problem distributions or adiabatic schedules. For Grover-driven QAOA, if the probability density function of objective values is known, the expectation value can be expressed entirely through the characteristic function

$$\ket{\psi_p(\bm{\gamma},\bm{\beta})} = \prod_{j=1}^p e^{-i\beta_j H_x}e^{-i\gamma_j H_z}\ket{+}^{\otimes n},$$8

so that angles can be optimized classically from distributional data rather than from repeated quantum evaluations. At $$\ket{\psi_p(\bm{\gamma},\bm{\beta})} = \prod_{j=1}^p e^{-i\beta_j H_x}e^{-i\gamma_j H_z}\ket{+}^{\otimes n},$$9,

$H_z=\frac{1}{2}\sum_{(j,k)\in E}(\mathbbm{1}-Z_jZ_k),$0

and for a Gaussian random cost model the reported optimum is $H_z=\frac{1}{2}\sum_{(j,k)\in E}(\mathbbm{1}-Z_jZ_k),$1, $H_z=\frac{1}{2}\sum_{(j,k)\in E}(\mathbbm{1}-Z_jZ_k),$2 (Headley et al., 2022). In fermionic QAOA, the offline schedule is instead derived from midpoint Trotterization of

$H_z=\frac{1}{2}\sum_{(j,k)\in E}(\mathbbm{1}-Z_jZ_k),$3

leading to

$H_z=\frac{1}{2}\sum_{(j,k)\in E}(\mathbbm{1}-Z_jZ_k),$4

These angles can be used directly as fixed-angle FQAOA parameters or as a warm start for BFGS or conjugate-gradient refinement. On the portfolio instance $H_z=\frac{1}{2}\sum_{(j,k)\in E}(\mathbbm{1}-Z_jZ_k),$5, fixed-angle FQAOA at $H_z=\frac{1}{2}\sum_{(j,k)\in E}(\mathbbm{1}-Z_jZ_k),$6 outperforms previous $H_z=\frac{1}{2}\sum_{(j,k)\in E}(\mathbbm{1}-Z_jZ_k),$7-QAOA-I at $H_z=\frac{1}{2}\sum_{(j,k)\in E}(\mathbbm{1}-Z_jZ_k),$8 with about half the number of gate operations, and optimized FQAOA reaches $H_z=\frac{1}{2}\sum_{(j,k)\in E}(\mathbbm{1}-Z_jZ_k),$9 at $$F_p(\bm{\gamma},\bm{\beta})=\langle\psi_p(\bm{\gamma},\bm{\beta})|H_z|\psi_p(\bm{\gamma},\bm{\beta})\rangle.$$0 and $$F_p(\bm{\gamma},\bm{\beta})=\langle\psi_p(\bm{\gamma},\bm{\beta})|H_z|\psi_p(\bm{\gamma},\bm{\beta})\rangle.$$1 at $$F_p(\bm{\gamma},\bm{\beta})=\langle\psi_p(\bm{\gamma},\bm{\beta})|H_z|\psi_p(\bm{\gamma},\bm{\beta})\rangle.$$2 (Yoshioka et al., 2023).

A more recent outer-loop-free construction is Penta-O. The key statement is that, for fixed earlier-level parameters, the QAOA energy at level $$F_p(\bm{\gamma},\bm{\beta})=\langle\psi_p(\bm{\gamma},\bm{\beta})|H_z|\psi_p(\bm{\gamma},\bm{\beta})\rangle.$$3 can be written as

$$F_p(\bm{\gamma},\bm{\beta})=\langle\psi_p(\bm{\gamma},\bm{\beta})|H_z|\psi_p(\bm{\gamma},\bm{\beta})\rangle.$$4

This permits level-wise determination of the final mixer angle from a small number of sampled points. The paper reports total sampling overhead proportional to $$F_p(\bm{\gamma},\bm{\beta})=\langle\psi_p(\bm{\gamma},\bm{\beta})|H_z|\psi_p(\bm{\gamma},\bm{\beta})\rangle.$$5, quadratic time complexity $$F_p(\bm{\gamma},\bm{\beta})=\langle\psi_p(\bm{\gamma},\bm{\beta})|H_z|\psi_p(\bm{\gamma},\bm{\beta})\rangle.$$6, and a non-decreasing performance guarantee because $$F_p(\bm{\gamma},\bm{\beta})=\langle\psi_p(\bm{\gamma},\bm{\beta})|H_z|\psi_p(\bm{\gamma},\bm{\beta})\rangle.$$7 when $$F_p(\bm{\gamma},\bm{\beta})=\langle\psi_p(\bm{\gamma},\bm{\beta})|H_z|\psi_p(\bm{\gamma},\bm{\beta})\rangle.$$8, so some $$F_p(\bm{\gamma},\bm{\beta})=\langle\psi_p(\bm{\gamma},\bm{\beta})|H_z|\psi_p(\bm{\gamma},\bm{\beta})\rangle.$$9 always satisfies $\alpha=F_p(\bm{\gamma}^*,\bm{\beta}^*)/C_{\max}$0 (Jiang et al., 23 Jan 2025).

3. Initialization and warm-start strategies

A large fraction of offline work does not eliminate optimization, but makes it cheaper and more stable by predicting good initial parameters. One example is the depth-progressive bilinear strategy for MaxCut. The method exploits two empirical patterns: within a fixed depth, $\alpha=F_p(\bm{\gamma}^*,\bm{\beta}^*)/C_{\max}$1 tends to increase with $\alpha=F_p(\bm{\gamma}^*,\bm{\beta}^*)/C_{\max}$2 while $\alpha=F_p(\bm{\gamma}^*,\bm{\beta}^*)/C_{\max}$3 tends to decrease with $\alpha=F_p(\bm{\gamma}^*,\bm{\beta}^*)/C_{\max}$4; across depths, parameters optimal at depth $\alpha=F_p(\bm{\gamma}^*,\bm{\beta}^*)/C_{\max}$5 are generally not optimal at $\alpha=F_p(\bm{\gamma}^*,\bm{\beta}^*)/C_{\max}$6, but the shift is structured. Together with periodicity and symmetry, this permits bounded extrapolation of the next-depth parameters. For $\alpha=F_p(\bm{\gamma}^*,\bm{\beta}^*)/C_{\max}$7, the paper reports $\alpha=F_p(\bm{\gamma}^*,\bm{\beta}^*)/C_{\max}$8 to $\alpha=F_p(\bm{\gamma}^*,\bm{\beta}^*)/C_{\max}$9 fewer objective evaluations than the parameters-fixing strategy, while avoiding the premature saturation observed in layerwise training (Lee et al., 2022).

For ma-QAOA, initialization heuristics exploit regularities in optimized angles. One study reports that about $AR=\langle C\rangle/C^*$0 of optimized $AR=\langle C\rangle/C^*$1 and $AR=\langle C\rangle/C^*$2 angles on 4-vertex graphs and about $AR=\langle C\rangle/C^*$3 on 8-vertex graphs are multiples of $AR=\langle C\rangle/C^*$4 at $AR=\langle C\rangle/C^*$5. This motivates initializing every angle to a random multiple of $AR=\langle C\rangle/C^*$6 in $AR=\langle C\rangle/C^*$7 and then running one round of BFGS. The resulting average approximation ratios are $AR=\langle C\rangle/C^*$8, $AR=\langle C\rangle/C^*$9, and $N\gtrsim 1000$0 for $N\gtrsim 1000$1, essentially matching the corresponding single-seed BFGS baseline values $N\gtrsim 1000$2, $N\gtrsim 1000$3, and $N\gtrsim 1000$4. A related heuristic sets maximum-degree vertex angles to $N\gtrsim 1000$5, giving $N\gtrsim 1000$6, $N\gtrsim 1000$7, and $N\gtrsim 1000$8 after BFGS (Wilkie et al., 2024).

For $N\gtrsim 1000$9, initialization can also be transferred from standard QAOA to MA-QAOA. On the MaxCut datasets considered in one study, the best tested QAOA initializer is the constant rule

$30$0

and the best MA-QAOA initializer is “QAOA Relax,” which first solves QAOA with that constant initialization and then uses the optimized QAOA angles to seed MA-QAOA. With those best strategies, MA-QAOA reaches the average target quality in $30$1 layers instead of $30$2, and worst-case depth reductions range from $30$3 to $30$4 depending on the dataset (Gaidai et al., 2023).

Warm-starting can also be organized by ansatz expressivity rather than by depth. For constrained optimization with XY mixers, one strategy trains offline on a restricted circuit whose dynamical Lie algebra is polynomially sized, then transfers the learned angles to the full circuit where the DLA is exponential, initializes the newly added $30$5 angles to zero, and continues optimization. The reported effect is improved convergence to high quality optima and dramatic performance benefits in terms of solution sampling and approximation ratio for both shared-angle and multi-angle QAOA (Kordonowy et al., 23 May 2025).

4. Structure-aware reduction of the offline search space

Offline optimization can be reduced not only by better seeds but also by shrinking the parameter manifold itself. In ma-QAOA, each cost term and each mixer term has its own parameter, so the ansatz is much more flexible than standard QAOA. On 8-vertex graphs, one-layer ma-QAOA achieves average approximation ratio $30$6, compared with $30$7, $30$8, and $30$9 for 1-, 2-, and 3-layer standard QAOA, respectively. The same study also reports that many optimized angles are zero: for $5$0, $5$1 of vertex angles and $5$2 of edge angles are zero, so the corresponding gates can be omitted entirely (Herrman et al., 2021).

A more systematic compression uses graph automorphisms. If vertices and edges are grouped into orbits under the automorphism group, one can assign the same QAOA angle to all vertices in the same vertex orbit and the same angle to all edges in the same edge orbit. The max-sym-QAOA reduction computes a symmetry generator, identifies the maximum vertex and edge orbits, ties parameters within each orbit, and then optimizes only the reduced parameter set. On all $5$3 connected, non-isomorphic 8-node graphs with a nontrivial symmetry group, $5$4 can have their ma-QAOA parameter count reduced with no loss in objective value, with average parameter reduction ratio $5$5. Using only the largest symmetry is enough in $5$6 of the graphs. When reduction does decrease performance, the largest symmetry still reduces parameters by $5$7 at the cost of only a $5$8 decrease in the objective. A random grouping baseline with the same parameter count performs much worse, indicating that the benefit comes from symmetry rather than from reduction alone (Shi et al., 2022).

This suggests a general distinction between redundant angles and expressive angles. In the source literature, redundancy is explained by orbit structure, by the empirical clustering of optimized angles, and by the frequency of zero-valued optimal parameters. A plausible implication is that offline optimization is often more effective when it first identifies equivalence classes or inactive directions and only then refines the remaining free variables.

5. Data-driven transfer, learned search policies, and low-query optimization

Another major direction uses offline datasets of optimized angles to recommend parameters for new instances. One unsupervised approach first builds a database of approximate optimal angles using BFGS with $5$9 random restarts on training instances, then performs clustering either directly on the angle vectors or on instance encodings. On the MaxCut dataset, angle clustering with $2p$0 reaches median approximation ratio around $2p$1, and for QUBO around $2p$2. Across the main cross-validation setting, the reported reduction relative to the best optimized angles is less than $2p$3–$2p$4. In the Recursive-QAOA showcase up to depth $2p$5, where only $2p$6 candidate angle sets are tried per iteration, the median approximation ratio is about $2p$7 on $2p$8 Erdős–Rényi graphs, while the baseline BFGS procedures that generated the training database used median numbers of circuit calls $2p$9, $$\ket{\psi_p(\bm{\gamma},\bm{\beta})} = \prod_{j=1}^p e^{-i\beta_j H_x}e^{-i\gamma_j H_z}\ket{+}^{\otimes n},$$00, and $$\ket{\psi_p(\bm{\gamma},\bm{\beta})} = \prod_{j=1}^p e^{-i\beta_j H_x}e^{-i\gamma_j H_z}\ket{+}^{\otimes n},$$01 at depths $$\ket{\psi_p(\bm{\gamma},\bm{\beta})} = \prod_{j=1}^p e^{-i\beta_j H_x}e^{-i\gamma_j H_z}\ket{+}^{\otimes n},$$02 for MaxCut (Moussa et al., 2022).

Learned optimizers can instead model the objective landscape directly. DARBO treats QAOA angle selection as continuous black-box optimization and uses a Gaussian-process surrogate with a Matérn-5/2 kernel, an adaptive trust region, and an adaptive search region. On $$\ket{\psi_p(\bm{\gamma},\bm{\beta})} = \prod_{j=1}^p e^{-i\beta_j H_x}e^{-i\gamma_j H_z}\ket{+}^{\otimes n},$$03 weighted 3-regular graphs, the paper reports that Adam’s approximation gaps are about $$\ket{\psi_p(\bm{\gamma},\bm{\beta})} = \prod_{j=1}^p e^{-i\beta_j H_x}e^{-i\gamma_j H_z}\ket{+}^{\otimes n},$$04 larger than DARBO and COBYLA’s are about $$\ket{\psi_p(\bm{\gamma},\bm{\beta})} = \prod_{j=1}^p e^{-i\beta_j H_x}e^{-i\gamma_j H_z}\ket{+}^{\otimes n},$$05 larger in exact simulation; under finite-shot simulation at $$\ket{\psi_p(\bm{\gamma},\bm{\beta})} = \prod_{j=1}^p e^{-i\beta_j H_x}e^{-i\gamma_j H_z}\ket{+}^{\otimes n},$$06 to $$\ket{\psi_p(\bm{\gamma},\bm{\beta})} = \prod_{j=1}^p e^{-i\beta_j H_x}e^{-i\gamma_j H_z}\ket{+}^{\otimes n},$$07 shots, Adam gaps are $$\ket{\psi_p(\bm{\gamma},\bm{\beta})} = \prod_{j=1}^p e^{-i\beta_j H_x}e^{-i\gamma_j H_z}\ket{+}^{\otimes n},$$08 larger, COBYLA gaps $$\ket{\psi_p(\bm{\gamma},\bm{\beta})} = \prod_{j=1}^p e^{-i\beta_j H_x}e^{-i\gamma_j H_z}\ket{+}^{\otimes n},$$09 larger, and SPSA gaps $$\ket{\psi_p(\bm{\gamma},\bm{\beta})} = \prod_{j=1}^p e^{-i\beta_j H_x}e^{-i\gamma_j H_z}\ket{+}^{\otimes n},$$10 larger (Cheng et al., 2023). A different line studies the extreme low-shot regime directly: for $$\ket{\psi_p(\bm{\gamma},\bm{\beta})} = \prod_{j=1}^p e^{-i\beta_j H_x}e^{-i\gamma_j H_z}\ket{+}^{\otimes n},$$11 MaxCut on a 16-qubit instance, both dual annealing and natural evolution strategies successfully optimize with only $$\ket{\psi_p(\bm{\gamma},\bm{\beta})} = \prod_{j=1}^p e^{-i\beta_j H_x}e^{-i\gamma_j H_z}\ket{+}^{\otimes n},$$12 shot per evaluated point, using $$\ket{\psi_p(\bm{\gamma},\bm{\beta})} = \prod_{j=1}^p e^{-i\beta_j H_x}e^{-i\gamma_j H_z}\ket{+}^{\otimes n},$$13 and $$\ket{\psi_p(\bm{\gamma},\bm{\beta})} = \prod_{j=1}^p e^{-i\beta_j H_x}e^{-i\gamma_j H_z}\ket{+}^{\otimes n},$$14 circuit evaluations respectively, which at $$\ket{\psi_p(\bm{\gamma},\bm{\beta})} = \prod_{j=1}^p e^{-i\beta_j H_x}e^{-i\gamma_j H_z}\ket{+}^{\otimes n},$$15 Hz corresponds to about a minute (Polloreno et al., 2022).

The most explicit offline search policy in the recent literature is a graph-conditioned trust-region method. A GIN-based graph neural network predicts a Gaussian

$$\ket{\psi_p(\bm{\gamma},\bm{\beta})} = \prod_{j=1}^p e^{-i\beta_j H_x}e^{-i\gamma_j H_z}\ket{+}^{\otimes n},$$16

whose mean initializes local search, whose covariance defines an ellipsoidal trust region, and whose uncertainty sets an instance-dependent evaluation budget. For MaxCut at $$\ket{\psi_p(\bm{\gamma},\bm{\beta})} = \prod_{j=1}^p e^{-i\beta_j H_x}e^{-i\gamma_j H_z}\ket{+}^{\otimes n},$$17 on graphs with $$\ket{\psi_p(\bm{\gamma},\bm{\beta})} = \prod_{j=1}^p e^{-i\beta_j H_x}e^{-i\gamma_j H_z}\ket{+}^{\otimes n},$$18–$$\ket{\psi_p(\bm{\gamma},\bm{\beta})} = \prod_{j=1}^p e^{-i\beta_j H_x}e^{-i\gamma_j H_z}\ket{+}^{\otimes n},$$19, the method reduces the mean number of circuit evaluations from $$\ket{\psi_p(\bm{\gamma},\bm{\beta})} = \prod_{j=1}^p e^{-i\beta_j H_x}e^{-i\gamma_j H_z}\ket{+}^{\otimes n},$$20 and $$\ket{\psi_p(\bm{\gamma},\bm{\beta})} = \prod_{j=1}^p e^{-i\beta_j H_x}e^{-i\gamma_j H_z}\ket{+}^{\otimes n},$$21 to $$\ket{\psi_p(\bm{\gamma},\bm{\beta})} = \prod_{j=1}^p e^{-i\beta_j H_x}e^{-i\gamma_j H_z}\ket{+}^{\otimes n},$$22 relative to random restarts and the strongest learned point-prediction baseline, while maintaining sampled approximation ratios within $$\ket{\psi_p(\bm{\gamma},\bm{\beta})} = \prod_{j=1}^p e^{-i\beta_j H_x}e^{-i\gamma_j H_z}\ket{+}^{\otimes n},$$23 percentage points of concentration-based heuristics. The predictive uncertainty is reported as calibrated, with $$\ket{\psi_p(\bm{\gamma},\bm{\beta})} = \prod_{j=1}^p e^{-i\beta_j H_x}e^{-i\gamma_j H_z}\ket{+}^{\otimes n},$$24 and Spearman correlation $$\ket{\psi_p(\bm{\gamma},\bm{\beta})} = \prod_{j=1}^p e^{-i\beta_j H_x}e^{-i\gamma_j H_z}\ket{+}^{\otimes n},$$25 (Huynh, 27 Apr 2026).

Fixed-point transfer provides a stronger form of offline reuse. In one fpQAOA construction, parameters are trained on $$\ket{\psi_p(\bm{\gamma},\bm{\beta})} = \prod_{j=1}^p e^{-i\beta_j H_x}e^{-i\gamma_j H_z}\ket{+}^{\otimes n},$$26 QUBO instances and transferred to larger sizes using the two-parameter sine–cosine schedule

$$\ket{\psi_p(\bm{\gamma},\bm{\beta})} = \prod_{j=1}^p e^{-i\beta_j H_x}e^{-i\gamma_j H_z}\ket{+}^{\otimes n},$$27

with $$\ket{\psi_p(\bm{\gamma},\bm{\beta})} = \prod_{j=1}^p e^{-i\beta_j H_x}e^{-i\gamma_j H_z}\ket{+}^{\otimes n},$$28, Frobenius normalization of the QUBO matrix, and training objective $$\ket{\psi_p(\bm{\gamma},\bm{\beta})} = \prod_{j=1}^p e^{-i\beta_j H_x}e^{-i\gamma_j H_z}\ket{+}^{\otimes n},$$29 for target approximation ratio $$\ket{\psi_p(\bm{\gamma},\bm{\beta})} = \prod_{j=1}^p e^{-i\beta_j H_x}e^{-i\gamma_j H_z}\ket{+}^{\otimes n},$$30. The empirical finding is that the median number of shots required to reach the target approximation ratio decreases with $$\ket{\psi_p(\bm{\gamma},\bm{\beta})} = \prod_{j=1}^p e^{-i\beta_j H_x}e^{-i\gamma_j H_z}\ket{+}^{\otimes n},$$31; omitting any one of the three modifications—target-AR training, $$\ket{\psi_p(\bm{\gamma},\bm{\beta})} = \prod_{j=1}^p e^{-i\beta_j H_x}e^{-i\gamma_j H_z}\ket{+}^{\otimes n},$$32, or Frobenius normalization—restores exponential growth of shot complexity (Chernyavskiy et al., 23 Sep 2025).

A practical infrastructure for such workflows is provided by JuliQAOA, which implements offline angle finding as a classical outer loop around exact simulation. Its main method uses basinhopping, depth-progressive seeding from $$\ket{\psi_p(\bm{\gamma},\bm{\beta})} = \prod_{j=1}^p e^{-i\beta_j H_x}e^{-i\gamma_j H_z}\ket{+}^{\otimes n},$$33 upward, user-provided initial angles when desired, checkpointed runs, and automatic differentiation through Enzyme.jl. The paper reports that example data for multiple problem types and mixer choices at $$\ket{\psi_p(\bm{\gamma},\bm{\beta})} = \prod_{j=1}^p e^{-i\beta_j H_x}e^{-i\gamma_j H_z}\ket{+}^{\otimes n},$$34 and up to $$\ket{\psi_p(\bm{\gamma},\bm{\beta})} = \prod_{j=1}^p e^{-i\beta_j H_x}e^{-i\gamma_j H_z}\ket{+}^{\otimes n},$$35 was generated on an Apple M2 Max laptop in $$\ket{\psi_p(\bm{\gamma},\bm{\beta})} = \prod_{j=1}^p e^{-i\beta_j H_x}e^{-i\gamma_j H_z}\ket{+}^{\otimes n},$$36 hour (Golden et al., 2023).

6. Scope, limits, and complexity boundaries

Offline angle optimization is not a single universal principle, but a collection of methods with sharply different assumptions. Fixed angles for regular MaxCut rely on regularity, constant edge weights, and the large-loop conjecture for worst-case guarantees (Wurtz et al., 2021). Distribution-based Grover-QAOA requires a known or accurately approximated objective-value distribution and the Grover driver rather than the standard transverse-field mixer (Headley et al., 2022). FQAOA’s adiabatic schedule is tied to a constraint-preserving fermionic driver and an initial state that is the ground state of that driver within the constrained sector (Yoshioka et al., 2023).

There are also explicit complexity boundaries. For general MaxCut graphs and any fixed round $$\ket{\psi_p(\bm{\gamma},\bm{\beta})} = \prod_{j=1}^p e^{-i\beta_j H_x}e^{-i\gamma_j H_z}\ket{+}^{\otimes n},$$37, exactly evaluating the QAOA expectation at prescribed angles is NP-hard, and approximating it within additive error $$\ket{\psi_p(\bm{\gamma},\bm{\beta})} = \prod_{j=1}^p e^{-i\beta_j H_x}e^{-i\gamma_j H_z}\ket{+}^{\otimes n},$$38 is already NP-hard. On the other hand, a dynamic-programming algorithm based on tree decomposition gives exact evaluation in time exponential only in the $$\ket{\psi_p(\bm{\gamma},\bm{\beta})} = \prod_{j=1}^p e^{-i\beta_j H_x}e^{-i\gamma_j H_z}\ket{+}^{\otimes n},$$39-local treewidth; if the $$\ket{\psi_p(\bm{\gamma},\bm{\beta})} = \prod_{j=1}^p e^{-i\beta_j H_x}e^{-i\gamma_j H_z}\ket{+}^{\otimes n},$$40-local treewidth grows at most logarithmically with $$\ket{\psi_p(\bm{\gamma},\bm{\beta})} = \prod_{j=1}^p e^{-i\beta_j H_x}e^{-i\gamma_j H_z}\ket{+}^{\otimes n},$$41, exact evaluation is polynomial-time in $$\ket{\psi_p(\bm{\gamma},\bm{\beta})} = \prod_{j=1}^p e^{-i\beta_j H_x}e^{-i\gamma_j H_z}\ket{+}^{\otimes n},$$42 for fixed $$\ket{\psi_p(\bm{\gamma},\bm{\beta})} = \prod_{j=1}^p e^{-i\beta_j H_x}e^{-i\gamma_j H_z}\ket{+}^{\otimes n},$$43. This evaluator does not itself optimize angles, but it provides the classical oracle needed by offline grid search, Bayesian optimization, or parameter sweeps on structured graph families (Wang et al., 25 Nov 2025).

A recurring practical tradeoff is that reduced depth does not necessarily imply reduced total QPU time. One study on MA-QAOA for $$\ket{\psi_p(\bm{\gamma},\bm{\beta})} = \prod_{j=1}^p e^{-i\beta_j H_x}e^{-i\gamma_j H_z}\ket{+}^{\otimes n},$$44 finds depth reductions up to $$\ket{\psi_p(\bm{\gamma},\bm{\beta})} = \prod_{j=1}^p e^{-i\beta_j H_x}e^{-i\gamma_j H_z}\ket{+}^{\otimes n},$$45, but also concludes that MA-QAOA is not optimal for minimization of the total QPU time because the larger classical search can dominate in low-$$\ket{\psi_p(\bm{\gamma},\bm{\beta})} = \prod_{j=1}^p e^{-i\beta_j H_x}e^{-i\gamma_j H_z}\ket{+}^{\otimes n},$$46 regimes (Gaidai et al., 2023). Similarly, the multiscale QAOA literature improves low-depth performance mainly by reusing coarse-grained states and effective Hamiltonians across scales; it does not provide an explicit offline recursion for $$\ket{\psi_p(\bm{\gamma},\bm{\beta})} = \prod_{j=1}^p e^{-i\beta_j H_x}e^{-i\gamma_j H_z}\ket{+}^{\otimes n},$$47 themselves (Zou, 2023).

Recent work also suggests broader transfer principles, but with narrower empirical support. One study argues that optimized QAOA angles collapse onto universal quantum annealing trajectories in integrated Hamiltonian coordinates and that the coldest effective temperature scales as $$\ket{\psi_p(\bm{\gamma},\bm{\beta})} = \prod_{j=1}^p e^{-i\beta_j H_x}e^{-i\gamma_j H_z}\ket{+}^{\otimes n},$$48; the evidence is strongest for random fully connected QUBO/Ising instances up to $$\ket{\psi_p(\bm{\gamma},\bm{\beta})} = \prod_{j=1}^p e^{-i\beta_j H_x}e^{-i\gamma_j H_z}\ket{+}^{\otimes n},$$49 (Díez-Valle et al., 3 Jun 2025). This suggests offline schedule templates rather than exact universal angles.

Taken together, the literature supports a technical taxonomy. Offline angle optimization may mean exact fixed-angle reuse, analytic schedule derivation, initialization and warm-start design, symmetry-aware parameter tying, learned transfer across instances, or learned search policies that reduce query complexity. The common objective is to replace expensive online variational search by structure that is already available: symmetry, locality, adiabatic bias, distributional information, pretrained data, or graph-conditioned uncertainty.