Heavy-Hex Graph Ising Models

• Heavy-Hex Graph Ising Models are spin systems defined on sparse, planar, and bipartite heavy-hex graphs that integrate hardware constraints with theoretical formulations.
• They support various Hamiltonian families—including quadratic, cubic, and transverse-field models—facilitating studies from dynamical confinement to optimization via QAOA.
• Their unique structure enables efficient routing and compilation on superconducting processors while offering insights into classical-versus-quantum scaling and entanglement dynamics.

Heavy-hex graph Ising models are Ising-type spin systems whose interaction structure is either the heavy-hex graph itself or a lattice that is compiled onto heavy-hex hardware. In the literature, the topic therefore spans several closely related settings: classical quadratic and geometrically local cubic spin glasses defined directly on heavy-hex graphs, transverse-field and kicked Ising dynamics on the heavy-hex lattice, and line-graph Ising circuits that are deterministically routed onto heavy-hex processors. Across these settings, the heavy-hex graph functions simultaneously as a graph-theoretic object, a native superconducting-qubit coupling map, and a constrained platform for studying compilation overhead, dynamical phenomena, and classical-versus-quantum scaling (Kattemölle et al., 2023, Pelofske et al., 2024, Tindall et al., 2024).

1. Graph-theoretic structure of the heavy-hex lattice

The heavy-hex graph is a sparse, planar, bipartite, geometrically local graph with bounded degree. In the construction used for large-scale random Ising studies, every generated heavy-hex graph is bipartite with vertices partitioned as $V = V_2 \sqcup V_3$ such that $E \subset V_2 \times V_3$; vertices in $V_2$ have maximum degree $2$, vertices in $V_3$ have maximum degree $3$, and the number of edges scales linearly with the number of vertices $n$, producing $O(n)$ edges (Pelofske et al., 2024). The transverse-field literature describes the same structure as a decorated hexagonal lattice: starting from a regular hexagonal lattice, every edge is decorated with one additional site, yielding two sublattices, $A$ of coordination number $2$ and $E \subset V_2 \times V_3$0 of coordination number $E \subset V_2 \times V_3$1, with no $E \subset V_2 \times V_3$2 or $E \subset V_2 \times V_3$3 nearest-neighbor bonds (Tindall et al., 2024).

In the routing literature, heavy-hex appears as the heavy graph $E \subset V_2 \times V_3$4 of a hexagonal base lattice $E \subset V_2 \times V_3$5. One places a node on every edge of $E \subset V_2 \times V_3$6 and connects that new node to the two original vertices incident to the edge. The new nodes are the “heavy nodes,” while the original vertices serve as “light” mediator nodes. This representation is central for compiling line-graph interaction patterns onto heavy-hex hardware because algorithm qubits can be assigned to heavy nodes and routed through unique mediators on the light sublattice (Kattemölle et al., 2023).

The architecture is also a hardware design choice. IBM’s superconducting processors use heavy-hex connectivity to mitigate frequency-collision and crosstalk errors while maintaining a manufacturable sparse topology. For the 127-qubit heavy-hex processor considered in magnetization-dynamics studies, the hardware graph has 144 two-qubit couplers and is 3-edge-colorable, a combinatorial property that enables parallel scheduling of all nearest-neighbor ZZ interactions in depth $E \subset V_2 \times V_3$7 (Pelofske et al., 2023).

2. Hamiltonian families defined on heavy-hex graphs

Several distinct Hamiltonian classes appear under the heading of heavy-hex graph Ising models. They differ in whether the model is classical or quantum, whether interactions are quadratic or higher order, and whether the graph is used directly or as a hardware target.

Family	Hamiltonian	Structural rule
Quadratic classical Ising	$E \subset V_2 \times V_3$8	Pairwise terms only on heavy-hex edges
Quadratic + local cubic Ising	$E \subset V_2 \times V_3$9	Cubic terms centered on degree-2 vertices
Heavy-hex TFIM	$V_2$0	Uniform nearest-neighbor ferromagnetic couplings on heavy-hex
Kicked heavy-hex TFIM	$V_2$1	Stroboscopic IBM circuit model on heavy-hex
XX+Z transverse-field model	$V_2$2	Native heavy-hex edge set used directly

For random combinatorial optimization instances, the coefficients $V_2$3, $V_2$4, and, when present, $V_2$5 are independently drawn from $V_2$6 with probability $V_2$7 each. The cubic extension is geometrically local by construction: each 3-local term involves a degree-2 vertex and its two neighbors, so every cubic monomial occupies a local heavy-hex motif rather than an arbitrary hyperedge (Pelofske et al., 2024).

For heavy-hex-compatible QAOA studies, the cost Hamiltonian is written as

$V_2$8

with $V_2$9 and $2$0 consisting of triples centered on degree-2 vertices. This model class is “heavy-hex compatible” because all terms are supported on native couplers or native two-edge motifs of the hardware graph (Pelofske et al., 2023).

A broader equilibrium formulation also exists. For ferromagnetic Ising models with general external field on a finite graph $2$1, the Hamiltonian is

$2$2

with $2$3. Because the random-cluster and Edwards–Sokal formalism applies on any finite graph and the heavy-hex graph is planar, bounded-degree, and amenable, this formulation specializes directly to finite heavy-hex patches (Cioletti et al., 2015).

3. Native compilation, mediated routing, and hardware realization

When the Ising model is already defined on the heavy-hex graph, heavy-hex processors admit direct spin-to-qubit mapping with no routing SWAPs. This is the regime targeted in heavy-hex QAOA and short-time evolution studies. A 3-edge-coloring of the heavy-hex coupling map partitions couplers into three disjoint matchings, so all pairwise phase-separation gates can be scheduled in parallel within each color class. For a quadratic term $2$4, the compilation uses a parity-compute $2$5-uncompute pattern,

$2$6

For a cubic term on a degree-2 center $2$7 with neighbors $2$8, the same parity chains are reused: $2$9 In the layered schedule, cubic terms therefore add only one additional $V_3$0 per triple and do not increase the CNOT depth, which remains $V_3$1 per QAOA layer independent of instance size (Pelofske et al., 2023). The same strict two-qubit depth $V_3$2 per layer is retained in later high-round parameter-transfer experiments on 127-, 133-, and 156-qubit heavy-hex devices, where a $V_3$3-qubit circuit at $V_3$4 contains $V_3$5 CZ gates (Pelofske et al., 16 Sep 2025).

A different use of heavy-hex arises when the algorithmic interaction graph is not heavy-hex but a line graph $V_3$6. In that setting, the hardware graph is $V_3$7, and every logical two-qubit ZZ interaction on $V_3$8 is mediated through a unique heavy-hex light node. If

$V_3$9

then line-graph routing replaces it by

$3$0

where $3$1 is the unique mediator with $3$2 and $3$3 hardware edges. The routing algorithm first tests whether the coupling graph is a line graph via $3$4, builds $3$5, replaces every ZZ gate by the mediated form, removes superfluous SWAPs, and optionally embeds $3$6 into the labeled device graph using VF2 (Kattemölle et al., 2023).

The formal guarantee is explicit: every circuit with coupling graph $3$7 can be performed on hardware with coupling graph $3$8 with

$3$9

and overall classical routing cost

$n$0

with $n$1 the total gate count. For the kagome-to-heavy-hex benchmark, the routed circuit has depth $n$2, SWAP count $n$3, active qubits $n$4, and routing time $n$5 s; compared to SABRE at its highest optimization level, line-graph routing achieved $n$6 smaller depth while also using fewer SWAPs (Kattemölle et al., 2023).

A common misconception is that heavy-hex itself is a line graph. The routing construction states the opposite: heavy-hex is the hardware target $n$7 when the algorithmic coupling graph is the line graph $n$8, as in kagome $n$9 (Kattemölle et al., 2023).

4. Dynamical phenomena on heavy-hex Ising lattices

The heavy-hex transverse-field Ising model exhibits a dynamical regime in which the lattice geometry itself induces confinement. For

$O(n)$0

defined on the decorated hexagonal heavy-hex lattice, a quench from the broken-symmetry state $O(n)$1 at small $O(n)$2 produces persistent oscillations, suppressed correlation spreading, and saturation of the entanglement entropy. The mechanism is geometric: flipping a single $O(n)$3 site creates two domain walls and costs $O(n)$4, whereas flipping a single $O(n)$5 site together with any subset of its three neighboring $O(n)$6 sites creates three domain walls and costs $O(n)$7. The effective confining potential is summarized heuristically as

$O(n)$8

and the low-energy zero-momentum sector is captured by a $O(n)$9 effective Hamiltonian whose eigenvalues define five hadronic quasiparticle masses. For quenches from $A$0, the regime $A$1 shows long-lived oscillations and bounded entanglement, whereas $A$2 shows linear growth of entanglement entropy density, spreading correlations, and relaxation consistent with ETH (Tindall et al., 2024).

A closely related heavy-hex dynamics problem was implemented as a kicked ferromagnetic TFIM on IBM’s 127-qubit heavy-hex architecture. The Hamiltonian is parameterized as

$A$3

with first-order Trotterization

$A$4

Because the heavy-hex graph is 3-edge-colorable, the ZZ layer fits in ECR/CNOT depth $A$5. This model was simulated on D-Wave Pegasus quantum annealers by matching the ratio of transverse-field and Ising energy scales through an anneal pause $A$6 satisfying

$A$7

together with pause-time matching equations for the ZZ and $A$8 actions. Using reverse quantum annealing and h-gain state encoding, the study reported equivalent magnetization dynamics for $A$9, $2$0, and up to $2$1 time steps, consistent with exact classical $2$2-qubit heavy-hex Trotterized circuit dynamics for whole-lattice magnetization (Pelofske et al., 2023).

Heavy-hex graph Ising models have also been used for short-time quantum-evolution-based ground-state estimation in the XX+Z transverse-field model

$2$3

Here the heavy-hex edge set is used directly, so no minor embedding is required. The computation combines shallow Trotterized sampling with a classical subspace diagonalization step, and success is diagnosed by the information-theoretic ratio

$2$4

On ibm_torino, with median CZ error $2$5 and median readout assignment error $2$6, reliable ground-state energy approximations were reported up to $2$7 heavy-hex qubits when $2$8 is not too large and $2$9 (Stenger et al., 28 Apr 2026).

These dynamical studies collectively show that heavy-hex geometry is not merely a hardware constraint. It also organizes the excitation spectrum, limits or enhances entanglement growth depending on the quench protocol, and strongly affects which observables remain accessible on current devices.

5. Classical complexity, equilibrium structure, and exact graphical representations

For random quadratic and geometrically local cubic Ising models defined directly on 2D heavy-hex graphs, the dominant structural fact is sparsity. The graph has maximum degree $E \subset V_2 \times V_3$00, is planar and bipartite, and supports only local pairwise and local cubic interactions. In an empirical scaling study covering $E \subset V_2 \times V_3$01 random instances per size for $E \subset V_2 \times V_3$02 from $E \subset V_2 \times V_3$03 up to $E \subset V_2 \times V_3$04, simulated annealing with $E \subset V_2 \times V_3$05 or $E \subset V_2 \times V_3$06 sweeps showed evidence of exponential Time-to-Solution scaling on quadratic-only instances, whereas Gurobi solved the same sparse instances efficiently on average. For quadratic-only instances, the best fit for Gurobi IQP runtime was linear with $E \subset V_2 \times V_3$07, $E \subset V_2 \times V_3$08, and RMSE $E \subset V_2 \times V_3$09. With cubic terms, ILP lifting showed linear scaling with $E \subset V_2 \times V_3$10, $E \subset V_2 \times V_3$11, and RMSE $E \subset V_2 \times V_3$12, while IQP order reduction showed weakly quadratic scaling with $E \subset V_2 \times V_3$13, $E \subset V_2 \times V_3$14, $E \subset V_2 \times V_3$15, and RMSE $E \subset V_2 \times V_3$16. The central conclusion was that these sparse heavy-hex instances are not classically hard on average for strong exact solvers, even though a simple heuristic such as simulated annealing can display weak exponential scaling (Pelofske et al., 2024).

This empirical result is important because heavy-hex-native instances are often chosen precisely to incur little or no embedding overhead on current quantum hardware. A second common misconception is therefore that hardware-native automatically implies classically challenging. The large-instance Gurobi study shows the opposite for the random heavy-hex ensembles tested: sparse locality can make exact classical optimization efficient on average (Pelofske et al., 2024).

The equilibrium statistical mechanics of ferromagnetic heavy-hex Ising models can be analyzed exactly through random-cluster and Edwards–Sokal representations. For a finite heavy-hex patch $E \subset V_2 \times V_3$17, with $E \subset V_2 \times V_3$18, the random-cluster weight is

$E \subset V_2 \times V_3$19

where $E \subset V_2 \times V_3$20. The exact single-site magnetization is

$E \subset V_2 \times V_3$21

and the exact spin-spin correlation is

$E \subset V_2 \times V_3$22

Because heavy-hex graphs are planar, periodic, bounded-degree, and amenable, the associated GRC measures inherit strong FKG, uniqueness of the infinite connected component under free and max-wired limits, and quasilocality (Cioletti et al., 2015).

For classical sampling beyond cluster algorithms, the Hybrid Monte Carlo formalism offers a graph-agnostic continuous-variable representation of the Ising model. After a Hubbard–Stratonovich transformation with a stabilizing shift $E \subset V_2 \times V_3$23, HMC targets

$E \subset V_2 \times V_3$24

with leapfrog dynamics driven by

$E \subset V_2 \times V_3$25

In the reported benchmarks, acceptance near criticality was about $E \subset V_2 \times V_3$26–$E \subset V_2 \times V_3$27, and the integrated autocorrelation time of $E \subset V_2 \times V_3$28 scaled with dynamic exponent $E \subset V_2 \times V_3$29 in $E \subset V_2 \times V_3$30 and $E \subset V_2 \times V_3$31. Since the method is explicitly formulated for arbitrary lattices and couplings, it carries over directly to heavy-hex graphs by building the symmetric coupling matrix from the heavy-hex edge set (Ostmeyer et al., 2019).

6. QAOA, parameter transfer, and the current research frontier

Heavy-hex-compatible higher-order Ising spin glasses have become a standard benchmark for whole-chip QAOA because the graph structure allows constant-depth compilation. In noiseless simulations on random $E \subset V_2 \times V_3$32 linear, quadratic, and cubic heavy-hex instances, a single set of angles trained on one 16-qubit instance transferred effectively to ensembles at 16, 27, and 127 qubits for $E \subset V_2 \times V_3$33 through $E \subset V_2 \times V_3$34. For 127 qubits, the average noiseless expectations across the ensemble were reported as roughly $E \subset V_2 \times V_3$35, $E \subset V_2 \times V_3$36, $E \subset V_2 \times V_3$37, $E \subset V_2 \times V_3$38, and $E \subset V_2 \times V_3$39 for $E \subset V_2 \times V_3$40. The $E \subset V_2 \times V_3$41 mean energy landscapes were nearly identical across 27-, 127-, and 414-qubit devices, with the best angles concentrating near $E \subset V_2 \times V_3$42–$E \subset V_2 \times V_3$43 and $E \subset V_2 \times V_3$44–$E \subset V_2 \times V_3$45 modulo periodicity. On hardware, the best 27-qubit processors improved mean energies up to $E \subset V_2 \times V_3$46, the best 127-qubit processors up to $E \subset V_2 \times V_3$47, and noisy means remained about a factor of two worse in magnitude than the noiseless expectations (Pelofske et al., 2023).

Later work pushed the same heavy-hex-compatible program to much larger QAOA depths. Angles trained classically on 16- and 27-qubit donor instances were transferred to 127-, 133-, and 156-qubit targets, with classical validation by statevector, PEPS, MPS, and LOWESA. The central result is qualified rather than monotone: transferred schedules do not in general improve strictly with $E \subset V_2 \times V_3$48, and temporary performance decreases occur, but the overall trend is toward improved expectation value with increasing depth, in many cases converging close to the true ground-state energy of the $E \subset V_2 \times V_3$49-qubit instances. On hardware, continuous improvement of solution quality was observed up to $E \subset V_2 \times V_3$50 on ibm_fez, $E \subset V_2 \times V_3$51 on ibm_torino, and $E \subset V_2 \times V_3$52 on ibm_pittsburgh, with no dynamical decoupling or error mitigation (Pelofske et al., 16 Sep 2025).

Several boundaries of the field are now clear. Heavy-hex-native compilation is exceptionally favorable: it avoids routing, fixes two-qubit depth at $E \subset V_2 \times V_3$53 per QAOA layer, and accommodates geometrically local cubic terms without extra entangling depth (Pelofske et al., 2023, Pelofske et al., 16 Sep 2025). At the same time, the random heavy-hex optimization problems most compatible with current hardware are structurally sparse, and sparse heavy-hex instances can be easy for strong classical exact solvers on average (Pelofske et al., 2024). In dynamical settings, heavy-hex geometry can either suppress entanglement through geometric confinement or support long-time analog emulation through native embedding on annealers, depending on the Hamiltonian and observable (Tindall et al., 2024, Pelofske et al., 2023).

Open directions reported in the literature include denser and/or nonlocal heavy-hex-compatible benchmarks, alternative classical heuristics such as cluster updates and parallel tempering, measurement of structural quantities such as treewidth and cutwidth, dynamic endpoint selection in line-graph routing to enhance SWAP cancellation, angle extrapolation beyond the trained QAOA depth, and lightweight online refinement of transferred schedules (Pelofske et al., 2024, Kattemölle et al., 2023, Pelofske et al., 16 Sep 2025). Taken together, these directions suggest that heavy-hex graph Ising models are best understood not as a single model, but as a tightly connected family of graph-native statistical-mechanical and quantum-circuit problems whose defining feature is the persistent interaction between lattice geometry, hardware constraints, and algorithm design.