Challenges
Datasets
Workspaces
Papers
Discussions
Leaderboard
Log inSign up
Challenges
Datasets
Workspaces
Papers
Discussions
Leaderboard
Blog
Job Board
Q3AS

© 2026 Aqora Quantum S.A.S.

TermsPrivacyLegal Notice

arXiv:2609.18209

The Expressive Power of Constrained QAOA: What You Might Have MISsed

Boris Tsvelikhovskiy, Bao Bach, Ilya Safro

quant-phSubmitted Sep 16, 2026• Announced Sep 17, 2026• 4 citations
arXivHTMLPDF

Abstract

Feasibility-preserving mixer Hamiltonians offer an attractive alternative to penalty encodings for constrained quantum optimization, yet how enforcing constraints alters the state space reachable by the Quantum Approximate Optimization Algorithm (QAOA) remains largely unexplored. We address this question within the context of the Maximum Independent Set (MIS) problem. Alongside the standard controlled bit-flip mixer, we analyze the alternative $\mathit{Flip\text{-}or\text{-}Stay}$ mixer, whose restriction to the feasible subspace $W_{\mathcal{F}}$ corresponds precisely to the shifted negative Laplacian of the independent-set reconfiguration graph. Although both mixers induce identical transitions between independent sets, the diagonal modification changes the spectral structure and can substantially enlarge the set of reachable quantum states. For every connected input graph with at least two vertices, we prove that independently parameterized local controls associated with either mixer generate the full unitary Lie algebra on the feasible subspace. Standard QAOA exhibits a more nuanced algebraic structure. We prove that the dynamical Lie algebra associated with MIS-QAOA employing the conventional mixer embeds into its Flip-or-Stay counterpart. Furthermore, we construct an infinite family of graphs for which the corresponding vacuum-state dynamical group orbits exhibit strictly distinct dimensions, establishing a qualitative expressivity separation between the two architectures. For either mixer, we show that standard QAOA initialized in the empty set can prepare a state supported entirely on maximum independent sets at finite depth. For the free architectures, we also derive an exact loss-variance formula in the unitary-design limit, expressed through the number of feasible independent sets and the variance of their cardinalities.

AI Summary

The paper asks how feasibility-preserving mixers affect the states reachable by QAOA for Maximum Independent Set (MIS). It compares the conventional controlled-flip (“Flip-or-Void”) mixer with a “Flip-or-Stay” variant that adds diagonal terms when flips are blocked. On the feasible space $W_{\mathcal F}$, these are respectively the adjacency operator $A_{\mathcal R}$ and shifted negative Laplacian $nI-L_{\mathcal R}$ of the independent-set reconfiguration graph. For every connected graph with $n\ge2$, either mixer with independent local controls generates $\mathfrak u(W_{\mathcal F})$. A global mixer plus local $Z$ controls suffices; standard two-parameter-per-layer QAOA is also fully controllable with positive, pairwise-distinct vertex weights. The proofs isolate transitions through commutators and generate matrix units. For unweighted standard QAOA, the conventional dynamical Lie algebra is contained in the modified one. An explicit family with $n=r(r-1)/2$ has vacuum-state orbit dimension at most $7$ for the conventional mixer versus $3r-2$ for Flip-or-Stay. Both architectures can exactly prepare a state supported on maximum independent sets from the vacuum at finite depth, using dynamical-orbit arguments and adiabatic convergence. For fully controlled circuits forming an exact unitary $2$-design, the loss variance is $4\operatorname{Var}(|S|)/(|\mathcal F|+1)$ for uniformly sampled independent sets $S$, bounded below by $\Omega(|\mathcal F|^{-2})$ and above by $O(\alpha(\Gamma)^2/|\mathcal F|)$. Reachability guarantees are existential, without depth bounds or parameter-finding algorithms. The mixer separation establishes no approximation-ratio advantage, and deep-design concentration results need not describe shallow-circuit trainability.

Some signs of AI-generated textAI use disclosedStrong assumptions
the authors used OpenAI ChatGPT (GPT-5.6 Sol) to assist in reviewing relevant literature, verifying the correctness of results, and refining the exposition.

Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.

Related papers

Constraint-preserving QAOA for multi-agent task allocation

Hyungseop Yu, Sang-Won Kang, Kwang‐Kyo Oh +1 more

We study a constraint-preserving Quantum Approximate Optimization Algorithm (QAOA) for the multi-agent task allocation problem under injective allocation constraints. The allocation cost is encoded into a diagonal cost Hamiltonian derived from the given cost matrix, while feasibility is enforced by a mixer Hamiltonian that restricts transitions along the edges of the feasible-state graph, thereby preventing probability leakage to infeasible allocations. We further provide a graph-theoretic sensitivity bound showing that the mixer unitary’s β-perturbation is bounded by the maximum degree of the feasible-state graph, clarifying how graph connectivity impacts mixing dynamics. Numerical simulations on a toy instance using Qiskit validate that the proposed approach concentrates sampling probability on low-cost feasible allocations.

IEEE Conference on Control Technology and Applications (CCTA)•Aug 1, 2026•0 citations

XY vs X mixer in Quantum Alternating Operator Ansatz for optimization problems with constraints

Zhihui Wang, Nicholas C. Rubin, Eleanor Gilbert Rieffel

Quantum Approximate Optimization Algorithm, further generalized as Quantum Alternating Operator Ansatz (QAOA), is a family of algorithms for combinatorial optimization problems. It is a leading candidate to run on emerging universal quantum computers to gain insight into quantum heuristics. In constrained optimization, penalties are often introduced so that the ground state of the cost Hamiltonian encodes the solution (a standard practice in quantum annealing). An alternative is to choose a mixing Hamiltonian such that the constraint corresponds to a constant of motion and the quantum evolution stays in the feasible subspace. Better performance of the algorithm is speculated due to a much smaller search space. We consider problems with a constant Hamming weight as the constraint. We also compare different methods of generating the generalized W-state, which serves as a natural initial state for the Hamming-weight constraint. Using graph-coloring as an example, we compare the performance of using XY model as a mixer that preserves the Hamming weight with the performance of adding a penalty term in the cost Hamiltonian.

Bulletin of the American Physical Society•Mar 8, 2018•0 citations

Beyond Feasibility: Finite-Depth Accessibility in Constrained QAOA

Rushikesh Ubale, Gregory T. Byrd, Yasar Mulani +1 more

Constraint-preserving mixers keep QAOA within a feasible subspace, but feasibility and global mixer connectivity do not determine which feasible configurations are available to a particular shallow circuit. We study constrained QAOA at the level of the circuit actually executed: a specified feasible initial state, finite depth, and finite schedule of mixer interactions. We define the finite-depth accessible set $R_p(M,x_0)$ and show that the computational-basis support of the ideal QAOA circuit is contained within it. This gives a parameter-independent objective ceiling $f_R^\star=\max_{x\in R_p}f(x)$ and a normalized reachable-quality diagnostic $Q_R$, separating accessible volume from the objective quality contained within it. We evaluate this framework using resource-matched block-local XY mixer schedules for constrained portfolio optimization and synthetic block-constrained quadratic problems. Across 207 favorable-tail CVaR QAOA configurations, $Q_R$ shows the strongest association with realized shallow-QAOA performance among the structural diagnostics considered, with Pearson and Spearman correlations of $0.7003$ and $0.8057$; the relationship remains positive under controlled and dependence-aware analyses. The distinction persists across changes in feasible initialization, portfolio size, and objective family. A compilation-only study further shows that, for the dense portfolio Hamiltonians considered here, phase-separator routing dominates the compiled two-qubit resource scale, while mixer schedules with comparable compiled two-qubit cost can exhibit markedly different finite-depth accessible quality. These results motivate objective-relevant finite-depth accessibility as a complementary diagnostic for resource-bounded constrained QAOA.

quant-pharXiv:2610.00736•Oct 2, 2026•1 citation

Reductions of QAOA Induced by Classical Symmetries: Theoretical Insights and Practical Implications

Boris Tsvelikhovskiy, Bao Bach, Jose Falla +1 more

The performance of the Quantum Approximate Optimization Algorithm (QAOA) is closely tied to the structure of the dynamical Lie algebra (DLA) generated by its Hamiltonians, which determines both its expressivity and trainability. In this work, we show that classical symmetries can be systematically exploited as a design principle for QAOA. Focusing on the MaxCut problem with global bit-flip symmetry, we analyze reduced QAOA instances obtained by fixing a single variable and study how this choice affects the associated DLAs. We show that the structure of the DLAs can change dramatically depending on which variable is held fixed. In particular, we construct explicit examples where the dimension collapses from exponential to quadratic, uncovering phenomena that do not appear in the original formulation. Numerical experiments on asymmetric graphs indicate that such reductions often produce DLAs of much smaller dimension, suggesting improved trainability. We also prove that any graph can be embedded into a slightly larger one (requiring only quadratic overhead) such that the standard reduced DLA coincides with the free reduced DLA, in most cases implying exponential dimension and irreducibility on the Hilbert space for reduced QAOA instances. These results establish symmetry-aware reduction as a principled tool for designing expressive and potentially trainable QAOA circuits.

quant-pharXiv:2602.16141•Feb 19, 2026•8 citations

Provable avoidance of barren plateaus for the Quantum Approximate Optimization Algorithm with Grover mixers

Boris Tsvelikhovskiy, Matthew Nuyten, Bojko Bakalov

We analyze the dynamical Lie algebras (DLAs) associated with the Grover-mixer variant of the Quantum Approximate Optimization Algorithm (GM-QAOA). When the initial state is the uniform superposition of computational basis states, we show that the corresponding DLA is isomorphic to $\mathfrak{su}(d) \oplus \mathfrak{u}(1)\oplus \mathfrak{u}(1)$, where $d$ denotes the number of distinct values of the objective function. We also establish an analogous result for other choices of initial states and Grover-type mixers. Furthermore, we prove that the DLA of GM-QAOA has the largest possible commutant among all QAOA variants initialized with the same state, corresponding physically to the maximal set of conserved quantities. We derive an explicit formula for the variance of the GM-QAOA loss function in terms of the objective function values, and we show that for a broad class of optimization problems, GM-QAOA with sufficiently many layers avoids barren plateaus.

arXiv:2509.10424•Sep 12, 2025•14 citations

Equivariant Quantum Approximate Optimization Algorithm

Boris Tsvelikhovskiy, Ilya Safro, Yuri Alexeev

Constructing effective mixer Hamiltonians is essential for enhancing the performance of the Quantum Approximate Optimization Algorithm (QAOA) in solving combinatorial optimization problems. In this work, we develop a systematic methodology for designing QAOA mixers that align with the symmetries of the classical objective function, with the goal of achieving values (mean, median, and minimum over multiple runs) that are closer to the true optimum. Our main idea is to design QAOA operators that are explicitly adapted to the action of symmetry groups on the Hilbert space. We focus on subgroups of the symmetric group$ S_{d}$, where$ d = 2^\ell$, to ensure compatibility with qudit-based quantum architectures. In particular, we construct QAOA mixers invariant under the full symmetric group$ S_{d}$as well as its cyclic subgroup$ \mathbb {Z}_{d} \subset S_{d}$. These constructions are natural in that they respect the decomposition of the Hilbert space into isotypic components under the symmetry group action. Notably, to the best of our knowledge, the QAOA algorithm based on the$ \mathbb {Z}_{d}$-invariant mixer provides the first example of a QAOA protocol whose dynamics (up to final measurement) are confined entirely within a nontrivial irreducible representation of a symmetry group of the objective function. Although our work does not investigate the benefits of exploiting such subspaces as computational resources, we think that the very realization of a variational algorithm whose evolution is restricted to a nontrivial symmetry-adapted subspace is of fundamental conceptual interest. We provide closed-form expressions for these mixers, together with explicit quantum circuit implementations. To empirically evaluate our approach, we compare QAOA variants employing the standard mixer$B = \sum X_{i}$with those using our proposed Hamiltonians$H_{M}$and$H_\chi$on edge coloring and graph partitioning problems. Across multiple graph instances, our symmetry-adapted mixers consistently yield objective values closer to the optimum, demonstrating statistically significant improvements over classical baselines.

arXiv:2405.07211•IEEE Transactions on Quantum Engineering•Jan 1, 2026•10 citations
Feasibility-preserving mixer Hamiltonians offer an attractive alternative to penalty encodings for constrained quantum optimization, yet how enforcing constraints alters the state space reachable by the Quantum Approximate Optimization Algorithm (QAOA) remains largely unexplored. We address this question within the context of the Maximum Independent Set (MIS) problem. Alongside the standard controlled bit-flip mixer, we analyze the alternative Flip-or-Stay\mathit{Flip\text{-}or\text{-}Stay}Flip-or-Stay mixer, whose restriction to the feasible subspace WFW_{\mathcal{F}}WF​ corresponds precisely to the shifted negative Laplacian of the independent-set reconfiguration graph. Although both mixers induce identical transitions between independent sets, the diagonal modification changes the spectral structure and can substantially enlarge the set of reachable quantum states.
For every connected input graph with at least two vertices, we prove that independently parameterized local controls associated with either mixer generate the full unitary Lie algebra on the feasible subspace. Standard QAOA exhibits a more nuanced algebraic structure. We prove that the dynamical Lie algebra associated with MIS-QAOA employing the conventional mixer embeds into its Flip-or-Stay counterpart. Furthermore, we construct an infinite family of graphs for which the corresponding vacuum-state dynamical group orbits exhibit strictly distinct dimensions, establishing a qualitative expressivity separation between the two architectures. For either mixer, we show that standard QAOA initialized in the empty set can prepare a state supported entirely on maximum independent sets at finite depth. For the free architectures, we also derive an exact loss-variance formula in the unitary-design limit, expressed through the number of feasible independent sets and the variance of their cardinalities.
The paper asks how feasibility-preserving mixers affect the states reachable by QAOA for Maximum Independent Set (MIS). It compares the conventional controlled-flip (“Flip-or-Void”) mixer with a “Flip-or-Stay” variant that adds diagonal terms when flips are blocked. On the feasible space WFW_{\mathcal F}WF​, these are respectively the adjacency operator ARA_{\mathcal R}AR​ and shifted negative Laplacian nI−LRnI-L_{\mathcal R}nI−LR​ of the independent-set reconfiguration graph.
For every connected graph with n≥2n\ge2n≥2, either mixer with independent local controls generates u(WF)\mathfrak u(W_{\mathcal F})u(WF​). A global mixer plus local ZZZ controls suffices; standard two-parameter-per-layer QAOA is also fully controllable with positive, pairwise-distinct vertex weights. The proofs isolate transitions through commutators and generate matrix units. For unweighted standard QAOA, the conventional dynamical Lie algebra is contained in the modified one. An explicit family with n=r(r−1)/2n=r(r-1)/2n=r(r−1)/2 has vacuum-state orbit dimension at most 777 for the conventional mixer versus 3r−23r-23r−2 for Flip-or-Stay. Both architectures can exactly prepare a state supported on maximum independent sets from the vacuum at finite depth, using dynamical-orbit arguments and adiabatic convergence.
For fully controlled circuits forming an exact unitary 222-design, the loss variance is 4Var⁡(∣S∣)/(∣F∣+1)4\operatorname{Var}(|S|)/(|\mathcal F|+1)4Var(∣S∣)/(∣F∣+1) for uniformly sampled independent sets SSS, bounded below by Ω(∣F∣−2)\Omega(|\mathcal F|^{-2})Ω(∣F∣−2) and above by O(α(Γ)2/∣F∣)O(\alpha(\Gamma)^2/|\mathcal F|)O(α(Γ)2/∣F∣). Reachability guarantees are existential, without depth bounds or parameter-finding algorithms. The mixer separation establishes no approximation-ratio advantage, and deep-design concentration results need not describe shallow-circuit trainability.
We study a constraint-preserving Quantum Approximate Optimization Algorithm (QAOA) for the multi-agent task allocation problem under injective allocation constraints. The allocation cost is encoded into a diagonal cost Hamiltonian derived from the given cost matrix, while feasibility is enforced by a mixer Hamiltonian that restricts transitions along the edges of the feasible-state graph, thereby preventing probability leakage to infeasible allocations. We further provide a graph-theoretic sensitivity bound showing that the mixer unitary’s β-perturbation is bounded by the maximum degree of the feasible-state graph, clarifying how graph connectivity impacts mixing dynamics. Numerical simulations on a toy instance using Qiskit validate that the proposed approach concentrates sampling probability on low-cost feasible allocations.
Quantum Approximate Optimization Algorithm, further generalized as Quantum Alternating Operator Ansatz (QAOA), is a family of algorithms for combinatorial optimization problems. It is a leading candidate to run on emerging universal quantum computers to gain insight into quantum heuristics. In constrained optimization, penalties are often introduced so that the ground state of the cost Hamiltonian encodes the solution (a standard practice in quantum annealing). An alternative is to choose a mixing Hamiltonian such that the constraint corresponds to a constant of motion and the quantum evolution stays in the feasible subspace. Better performance of the algorithm is speculated due to a much smaller search space. We consider problems with a constant Hamming weight as the constraint. We also compare different methods of generating the generalized W-state, which serves as a natural initial state for the Hamming-weight constraint. Using graph-coloring as an example, we compare the performance of using XY model as a mixer that preserves the Hamming weight with the performance of adding a penalty term in the cost Hamiltonian.
Constraint-preserving mixers keep QAOA within a feasible subspace, but feasibility and global mixer connectivity do not determine which feasible configurations are available to a particular shallow circuit. We study constrained QAOA at the level of the circuit actually executed: a specified feasible initial state, finite depth, and finite schedule of mixer interactions. We define the finite-depth accessible set Rp(M,x0)R_p(M,x_0)Rp​(M,x0​) and show that the computational-basis support of the ideal QAOA circuit is contained within it. This gives a parameter-independent objective ceiling fR⋆=max⁡x∈Rpf(x)f_R^\star=\max_{x\in R_p}f(x)fR⋆​=maxx∈Rp​​f(x) and a normalized reachable-quality diagnostic QRQ_RQR​, separating accessible volume from the objective quality contained within it. We evaluate this framework using resource-matched block-local XY mixer schedules for constrained portfolio optimization and synthetic block-constrained quadratic problems. Across 207 favorable-tail CVaR QAOA configurations, QRQ_RQR​ shows the strongest association with realized shallow-QAOA performance among the structural diagnostics considered, with Pearson and Spearman correlations of 0.70030.70030.7003 and 0.80570.80570.8057; the relationship remains positive under controlled and dependence-aware analyses. The distinction persists across changes in feasible initialization, portfolio size, and objective family. A compilation-only study further shows that, for the dense portfolio Hamiltonians considered here, phase-separator routing dominates the compiled two-qubit resource scale, while mixer schedules with comparable compiled two-qubit cost can exhibit markedly different finite-depth accessible quality. These results motivate objective-relevant finite-depth accessibility as a complementary diagnostic for resource-bounded constrained QAOA.
The performance of the Quantum Approximate Optimization Algorithm (QAOA) is closely tied to the structure of the dynamical Lie algebra (DLA) generated by its Hamiltonians, which determines both its expressivity and trainability. In this work, we show that classical symmetries can be systematically exploited as a design principle for QAOA. Focusing on the MaxCut problem with global bit-flip symmetry, we analyze reduced QAOA instances obtained by fixing a single variable and study how this choice affects the associated DLAs. We show that the structure of the DLAs can change dramatically depending on which variable is held fixed. In particular, we construct explicit examples where the dimension collapses from exponential to quadratic, uncovering phenomena that do not appear in the original formulation. Numerical experiments on asymmetric graphs indicate that such reductions often produce DLAs of much smaller dimension, suggesting improved trainability. We also prove that any graph can be embedded into a slightly larger one (requiring only quadratic overhead) such that the standard reduced DLA coincides with the free reduced DLA, in most cases implying exponential dimension and irreducibility on the Hilbert space for reduced QAOA instances. These results establish symmetry-aware reduction as a principled tool for designing expressive and potentially trainable QAOA circuits.
We analyze the dynamical Lie algebras (DLAs) associated with the Grover-mixer variant of the Quantum Approximate Optimization Algorithm (GM-QAOA). When the initial state is the uniform superposition of computational basis states, we show that the corresponding DLA is isomorphic to su(d)⊕u(1)⊕u(1)\mathfrak{su}(d) \oplus \mathfrak{u}(1)\oplus \mathfrak{u}(1)su(d)⊕u(1)⊕u(1), where ddd denotes the number of distinct values of the objective function. We also establish an analogous result for other choices of initial states and Grover-type mixers. Furthermore, we prove that the DLA of GM-QAOA has the largest possible commutant among all QAOA variants initialized with the same state, corresponding physically to the maximal set of conserved quantities. We derive an explicit formula for the variance of the GM-QAOA loss function in terms of the objective function values, and we show that for a broad class of optimization problems, GM-QAOA with sufficiently many layers avoids barren plateaus.
Constructing effective mixer Hamiltonians is essential for enhancing the performance of the Quantum Approximate Optimization Algorithm (QAOA) in solving combinatorial optimization problems. In this work, we develop a systematic methodology for designing QAOA mixers that align with the symmetries of the classical objective function, with the goal of achieving values (mean, median, and minimum over multiple runs) that are closer to the true optimum. Our main idea is to design QAOA operators that are explicitly adapted to the action of symmetry groups on the Hilbert space. We focus on subgroups of the symmetric group$ S_{d},where, where,where d = 2^\ell,toensurecompatibilitywithqudit−basedquantumarchitectures.Inparticular,weconstructQAOAmixersinvariantunderthefullsymmetricgroup, to ensure compatibility with qudit-based quantum architectures. In particular, we construct QAOA mixers invariant under the full symmetric group,toensurecompatibilitywithqudit−basedquantumarchitectures.Inparticular,weconstructQAOAmixersinvariantunderthefullsymmetricgroup S_{d}aswellasitscyclicsubgroupas well as its cyclic subgroupaswellasitscyclicsubgroup \mathbb {Z}_{d} \subset S_{d}.TheseconstructionsarenaturalinthattheyrespectthedecompositionoftheHilbertspaceintoisotypiccomponentsunderthesymmetrygroupaction.Notably,tothebestofourknowledge,theQAOAalgorithmbasedonthe. These constructions are natural in that they respect the decomposition of the Hilbert space into isotypic components under the symmetry group action. Notably, to the best of our knowledge, the QAOA algorithm based on the.TheseconstructionsarenaturalinthattheyrespectthedecompositionoftheHilbertspaceintoisotypiccomponentsunderthesymmetrygroupaction.Notably,tothebestofourknowledge,theQAOAalgorithmbasedonthe \mathbb {Z}_{d}−invariantmixerprovidesthefirstexampleofaQAOAprotocolwhosedynamics(uptofinalmeasurement)areconfinedentirelywithinanontrivialirreduciblerepresentationofasymmetrygroupoftheobjectivefunction.Althoughourworkdoesnotinvestigatethebenefitsofexploitingsuchsubspacesascomputationalresources,wethinkthattheveryrealizationofavariationalalgorithmwhoseevolutionisrestrictedtoanontrivialsymmetry−adaptedsubspaceisoffundamentalconceptualinterest.Weprovideclosed−formexpressionsforthesemixers,togetherwithexplicitquantumcircuitimplementations.Toempiricallyevaluateourapproach,wecompareQAOAvariantsemployingthestandardmixer-invariant mixer provides the first example of a QAOA protocol whose dynamics (up to final measurement) are confined entirely within a nontrivial irreducible representation of a symmetry group of the objective function. Although our work does not investigate the benefits of exploiting such subspaces as computational resources, we think that the very realization of a variational algorithm whose evolution is restricted to a nontrivial symmetry-adapted subspace is of fundamental conceptual interest. We provide closed-form expressions for these mixers, together with explicit quantum circuit implementations. To empirically evaluate our approach, we compare QAOA variants employing the standard mixer−invariantmixerprovidesthefirstexampleofaQAOAprotocolwhosedynamics(uptofinalmeasurement)areconfinedentirelywithinanontrivialirreduciblerepresentationofasymmetrygroupoftheobjectivefunction.Althoughourworkdoesnotinvestigatethebenefitsofexploitingsuchsubspacesascomputationalresources,wethinkthattheveryrealizationofavariationalalgorithmwhoseevolutionisrestrictedtoanontrivialsymmetry−adaptedsubspaceisoffundamentalconceptualinterest.Weprovideclosed−formexpressionsforthesemixers,togetherwithexplicitquantumcircuitimplementations.Toempiricallyevaluateourapproach,wecompareQAOAvariantsemployingthestandardmixerB = \sum X_{i}withthoseusingourproposedHamiltonianswith those using our proposed HamiltonianswiththoseusingourproposedHamiltoniansH_{M}andandandH_\chi$on edge coloring and graph partitioning problems. Across multiple graph instances, our symmetry-adapted mixers consistently yield objective values closer to the optimum, demonstrating statistically significant improvements over classical baselines.