On the Expressive Power of the Transverse-Field Ising Model for Graph Learning

Abstract: We study the quantum evolution induced by graph-indexed Ising Hamiltonians as a source of structural signal for graph learning. Graph automorphisms preserve symmetries of the Hamiltonian, and these symmetries constrain the quantum evolution in a way that turns time-dependent local measurements into informative probes of graph structure. Leveraging this idea, we introduce QDAGer, a quantum-inspired graph-pair Transformer that injects quantum-dynamical features from time series of node occupations and connected two-point correlators directly into the attention mechanism. We apply QDAGer to learning Graph Edit Distance (GED), an NP-hard similarity measure, using either a direct permutation-invariant embedding discrepancy or an alignment-based surrogate loss. Experiments on multiple GED benchmarks under different edit cost settings show that the proposed dynamical features provide a stronger inductive bias than classical structural alternatives under the same training protocol. In addition, we report ablations where the dynamical signal is replaced by standard random-walk and heat-kernel features while keeping the architecture fixed, highlighting that the gain comes from the injected dynamics rather than model capacity alone.

• The paper establishes that graph-indexed transverse-field Ising dynamics generically distinguish automorphism orbits and provides a probabilistic measurement bound for identifying non-isomorphic graph pairs.
• QDAGer injects time-resolved node occupations and connected correlators into a swap-consistent graph-pair Transformer, achieving leading uniform-cost GED results on AIDS, Yeast, Mutag, MolHIV, and MolPCBA.
• Controlled ablations show quantum dynamical features outperform heat-kernel and random-walk signals, while state-vector emulation, finite-shot noise, hardware constraints, and scalability remain important open challenges.

Overview

This paper investigates whether quantum dynamics generated by graph-indexed transverse-field Ising Hamiltonians can serve as a structural signal for graph learning, and operationalizes this idea in a graph-pair Transformer for Graph Edit Distance (GED) prediction (2608.17750). The central observation is that graph automorphisms are symmetries of the Hamiltonian, so time-dependent local measurements—node occupations $\langle \hat n_u(t)\rangle$ and connected two-point correlators—are automatically permutation-consistent and sensitive to automorphism orbits rather than arbitrary node labels. The authors build on prior work using neutral-atom dynamics as static kernels or positional encodings (e.g., the quantum evolution kernel and quantum positional encodings), but differ by injecting time-resolved dynamical features directly into the attention mechanism of a trainable architecture.

Symmetry structure of graph-indexed Ising dynamics

Each vertex is mapped to a qubit, and the Hamiltonian couples occupations along edges:

$$\widehat H_{\mathcal G} = \sum_{(u,v)\in\mathcal E}\widehat n_u\widehat n_v + \Omega\sum_u \widehat\sigma_u^x - \delta\sum_u \widehat n_u,$$

an occupation-number parametrization of a transverse-field Ising model, also motivated by the Rydberg van der Waals Hamiltonian realizable on neutral-atom hardware. Because every $g \in \mathrm{Aut}(\mathcal G)$ commutes with $\widehat H_{\mathcal G}$, the trivial symmetry sector $\mathcal H_0$ (states invariant under all automorphisms) is dynamically invariant, and computational-basis amplitudes are constant on automorphism orbits.

The main theoretical result states that within $\mathcal H_0$, occupation observables associated with distinct subset orbits yield generically distinct expectation-value trajectories: for Haar-almost every initial state, the signal $f_{S,S'}(t)$ is a nonzero trigonometric polynomial whose zeros form a Lebesgue-null set. The proof requires only finite dimensionality and automorphism invariance of the Hamiltonian, so it applies to any graph-indexed Hamiltonian. Combined with Frucht's theorem on automorphisms of disconnected unions, this yields an orbit-crossing criterion: two connected graphs are isomorphic if and only if some orbit of $\mathrm{Aut}(\mathcal G_1 \cup \mathcal G_2)$ crosses both components. Since the propagator factorizes over disjoint components, vertex-level occupation dynamics can be compared independently across graphs, giving a measurement-complexity bound of $O(\log(N/\alpha)\,\mathrm{gap}(t_0)^{-2})$ shots to distinguish non-isomorphic pairs at confidence $1-\alpha$.

Two caveats deserve emphasis. First, the guarantee is generic over Haar-random initial states; the practical pipeline uses the deterministic initialization $$\widehat H_{\mathcal G} = \sum_{(u,v)\in\mathcal E}\widehat n_u\widehat n_v + \Omega\sum_u \widehat\sigma_u^x - \delta\sum_u \widehat n_u,$$0, for which separation is not guaranteed by the theorem—the paper shows that at $$\widehat H_{\mathcal G} = \sum_{(u,v)\in\mathcal E}\widehat n_u\widehat n_v + \Omega\sum_u \widehat\sigma_u^x - \delta\sum_u \widehat n_u,$$1 the state is stationary and the compressed signal vanishes identically, so the nonzero transverse field is essential to generate a nontrivial Krylov subspace. Second, the comparison against $$\widehat H_{\mathcal G} = \sum_{(u,v)\in\mathcal E}\widehat n_u\widehat n_v + \Omega\sum_u \widehat\sigma_u^x - \delta\sum_u \widehat n_u,$$2-WL is carefully qualified: the dynamical criterion is probabilistic and its cost depends on instance-specific gaps, so no uniform computational advantage over WL is claimed.

Empirical separability on hard graphs

The theory is validated on strongly regular graph families that defeat low-dimensional WL tests. For the Rook's $$\widehat H_{\mathcal G} = \sum_{(u,v)\in\mathcal E}\widehat n_u\widehat n_v + \Omega\sum_u \widehat\sigma_u^x - \delta\sum_u \widehat n_u,$$3 versus Shrikhande pair—non-isomorphic srg$$\widehat H_{\mathcal G} = \sum_{(u,v)\in\mathcal E}\widehat n_u\widehat n_v + \Omega\sum_u \widehat\sigma_u^x - \delta\sum_u \widehat n_u,$$4 graphs requiring 4-WL, whose exact cost scales as $$\widehat H_{\mathcal G} = \sum_{(u,v)\in\mathcal E}\widehat n_u\widehat n_v + \Omega\sum_u \widehat\sigma_u^x - \delta\sum_u \widehat n_u,$$5—the one-body occupation trajectories under $$\widehat H_{\mathcal G} = \sum_{(u,v)\in\mathcal E}\widehat n_u\widehat n_v + \Omega\sum_u \widehat\sigma_u^x - \delta\sum_u \widehat n_u,$$6, $$\widehat H_{\mathcal G} = \sum_{(u,v)\in\mathcal E}\widehat n_u\widehat n_v + \Omega\sum_u \widehat\sigma_u^x - \delta\sum_u \widehat n_u,$$7 separate visibly, and vertex-transitivity reduces the comparison to a single cross-graph trajectory difference certifiable at 95% confidence with a modest shot budget. Extending beyond vertex-transitive instances, permutation-invariant RMS distances between node-occupation traces (computed exactly via linear assignment) become strictly positive for every non-isomorphic pair in srg$$\widehat H_{\mathcal G} = \sum_{(u,v)\in\mathcal E}\widehat n_u\widehat n_v + \Omega\sum_u \widehat\sigma_u^x - \delta\sum_u \widehat n_u,$$8 (10 graphs) and srg$$\widehat H_{\mathcal G} = \sum_{(u,v)\in\mathcal E}\widehat n_u\widehat n_v + \Omega\sum_u \widehat\sigma_u^x - \delta\sum_u \widehat n_u,$$9 (15 graphs), demonstrating that one-point dynamics alone suffice for these families from the fixed $g \in \mathrm{Aut}(\mathcal G)$0 initialization.

The QDAGer architecture

QDAGer (Quantum Dynamics-based Attention Graph transformer) is a GPS-style graph-pair Transformer. Preprocessing emulates the state-vector evolution of each graph and records node occupation series $g \in \mathrm{Aut}(\mathcal G)$1 and connected correlator series $g \in \mathrm{Aut}(\mathcal G)$2. Attention logits are built from swap-consistent symmetric/antisymmetric combinations $g \in \mathrm{Aut}(\mathcal G)$3 (and analogously for pairs), softmaxed jointly over nodes and node pairs. Layer updates couple the two levels: pair embeddings are refreshed through gated interactions of node-derived queries/keys with pair features, and node embeddings aggregate attention-weighted values plus updated pair information. The construction is swap-equivariant at the layer level.

Two training heads are considered: a blindfold loss, regressing a permutation-invariant pooled discrepancy between embedding sets onto the true GED, and a neural set divergence surrogate following Jain et al., where a Sinkhorn-based PermNet produces doubly-stochastic transport plans over nodes and pairs, trained end-to-end with the XOR-DiffAlign surrogate. Swap invariance holds exactly for the surrogate only when a two-sided surrogate is used and Sinkhorn has converged to genuinely doubly-stochastic plans; the blindfold read-out is not swap-invariant regardless of edit costs, and enforcing it via absolute values empirically degrades performance. An expressiveness theorem establishes that, for Haar-generic initial states and almost every sampling-time choice, there exists a parameter setting whose embeddings separate all non-isomorphic pairs in any finite dataset—proved constructively by exhibiting parameters that preserve the input dynamical features exactly (identity FFN, zeroed value/pair projections).

Experimental results

Evaluation covers seven datasets (AIDS, Yeast, Mutagenicity, MolHIV, MolPCBA, Code2, Linux) under uniform ($g \in \mathrm{Aut}(\mathcal G)$4) and non-uniform ($g \in \mathrm{Aut}(\mathcal G)$5) edit costs, with ground-truth GED computed exactly via the F2 solver in GEDLIB. Key findings:

Setting	Result
Surrogate loss, data-rich datasets	QDAGer achieves best MSE on AIDS (0.472), Yeast (0.699), Mutag (0.453), MolHIV (0.578), MolPCBA (0.628) under uniform costs
Blindfold loss	QDAGer (dir) improves over strongest baselines (e.g., EGSC, ERIC) on most large datasets
Ablations	Replacing quantum features with heat-kernel or random-walk features degrades performance in most blindfold cases and all surrogate cases

The ablation design—holding architecture, alignment module, and loss fixed while swapping only the input signal—is the paper's strongest evidence that the gain derives from the dynamical inductive bias rather than model capacity. Notably, QDAGer underperforms GraphEDX on the small Code2 (3,629 pairs) and Linux (1,774 pairs) datasets, which the authors attribute to optimization variance under data scarcity, a plausible but not fully isolated explanation since hyperparameters were tuned per dataset. All surrogate experiments use a single fixed surrogate without selection or tuning, keeping comparisons conservative.

Limitations and open questions

The paper is explicit about several constraints. State-vector emulation limits feasible graphs to roughly 25 nodes and makes feature generation expensive, though the cost is amortized over the $g \in \mathrm{Aut}(\mathcal G)$6 pairs constructed from $g \in \mathrm{Aut}(\mathcal G)$7 emulated graphs. All results use noiseless expectation values; robustness under finite-shot estimation—and hence the relevance of the shot-complexity bound in practice—is untested. Hardware deployment would require geometric embeddings of input graphs into neutral-atom arrays, which many benchmark graphs do not admit, and residual long-range van der Waals interactions would perturb the idealized edge-indexed coupling. The expressiveness theorem assumes Haar-generic initial states, whereas experiments use $g \in \mathrm{Aut}(\mathcal G)$8, leaving open which graph families and Hamiltonian parameter regimes guarantee separation from this specific initialization. Finally, the method is purely structure-oriented and was not evaluated on labeled graphs.

Conclusion

The paper establishes graph-indexed Ising dynamics as a principled, symmetry-aware source of structural features for graph comparison, proves generic orbit-level distinguishability and architecture-level expressiveness guarantees, and demonstrates empirically—via controlled ablations—that these dynamical features outperform classical random-walk and heat-kernel alternatives for GED prediction on data-rich benchmarks. The main unresolved issues are scalability of feature generation beyond state-vector emulation, finite-shot robustness, and hardware realizability, all of which bear directly on whether the demonstrated inductive bias can be exploited at scale.