Willow Quantum Processor

• Willow Quantum Processor is a square-lattice-like planar architecture with a coordination number of 4, influencing tensor-network simulations.
• Its denser connectivity generates strong loop correlations that significantly affect belief propagation accuracy and contraction performance.
• The architecture imposes higher simulation complexity and memory scaling—O(N×χ⁴)—compared to sparse, tree-like layouts such as heavy-hex.

Searching arXiv for the cited paper and closely related references on Willow processor geometry and tensor-network simulation. The Willow Quantum Processor is treated in recent tensor-network simulation work as a concrete instance of a 2D planar processor topology whose connectivity is close to a square lattice rather than a sparse, tree-like graph (Rudolph et al., 15 Jul 2025). In that framework, “geometry” denotes the connectivity graph on which qubits reside and on which two-qubit gates are applied. Willow is therefore not defined merely as a hardware layout, but as a square-lattice-like planar architecture with coordination number $z=4$, whose denser connectivity generates short loops that materially affect tensor-network ansätze, contraction accuracy, sampling quality, and ultimately classical simulability (Rudolph et al., 15 Jul 2025).

1. Geometric definition and processor-graph interpretation

In the cited simulation framework, the Willow processor geometry is instantiated as a planar network in which vertices represent qubits or site tensors and edges represent both physical couplings and virtual bonds in the tensor network (Rudolph et al., 15 Jul 2025). The paper states explicitly that “we have $z = 3$ for heavy-hex geometries and $z = 4$ for the Willow processor geometry,” making the Willow layout a higher-coordination architecture than IBM’s heavy-hex (Rudolph et al., 15 Jul 2025).

The paper further characterizes Willow as a square-lattice processor and a rotated-square-lattice-like planar architecture. This places it in a class of denser, more grid-like processor graphs with smaller primitive loops than heavy-hex. For the tensor-network ansatz, the structure is said to “reflect that of the underlying processor,” and in the Willow case each tensor has at most four virtual legs, consistent with $z=4$ (Rudolph et al., 15 Jul 2025).

This graph-theoretic interpretation is central to the article’s technical claims. The processor geometry is not treated as an implementation detail external to simulation; rather, it determines the amount of loop structure in the corresponding tensor network. A plausible implication is that the processor graph and the tensor-network contraction problem should be understood as tightly coupled objects, especially for planar superconducting architectures.

2. Role in circuit construction and Trotterized dynamics

The Willow geometry appears in the paper’s study of a domain-wall quench in a two-dimensional discrete-time Heisenberg model defined directly on the device graph (Rudolph et al., 15 Jul 2025). The circuit is built through a Trotter-Suzuki decomposition,

$U = \prod_{l=1}^L U_l,$

with each Trotter slice decomposed into non-overlapping two-qubit substeps over edge sets $E_k$. For bipartite lattices the paper sets $K=z$, so Willow uses $K=4$ while heavy-hex uses $K=3$ (Rudolph et al., 15 Jul 2025).

That difference has an immediate operational meaning: Willow requires more layers of non-overlapping two-qubit substeps per Trotter slice than heavy-hex. Because its connectivity is denser, the geometry also creates more opportunities for entanglement loops during time evolution. The paper’s comparison therefore links circuit compilation depth at the substep level to graph coordination number, rather than to qubit count alone.

The broader comparison in the paper distinguishes two families of layouts. Heavy-hex is described as sparse and “tree-like,” with larger primitive loops and lower coordination. Willow, by contrast, is grouped with rotated square lattices: denser, more grid-like, and characterized by smaller loops. This topological contrast underlies the paper’s conclusion that geometry is a primary determinant of classical simulability (Rudolph et al., 15 Jul 2025).

3. Tensor-network representation and geometry-dependent cost

The many-body wavefunction is represented by a 2D tensor-network ansatz with one low-rank tensor per qubit and virtual bonds carrying entanglement (Rudolph et al., 15 Jul 2025). In this formalism, the memory scaling is

$\mathcal{O}(N_{\rm qubits}\chi^{z}),$

where $z = 3$0 is the number of qubits, $z = 3$1 is the maximum bond dimension, and $z = 3$2 is the coordination number. Since Willow has $z = 3$3, its memory scaling becomes

$z = 3$4

whereas heavy-hex with $z = 3$5 scales as $z = 3$6 in the same parameterization (Rudolph et al., 15 Jul 2025).

Gate application is handled by a BP-based simple-update-like procedure. One-qubit gates are applied exactly. Two-qubit gates are applied and then truncated back to bond dimension $z = 3$7 via SVD. The approximate per-gate truncation error is defined as

$z = 3$8

with per-gate fidelity $z = 3$9, and an approximate final fidelity

$z = 4$0

These definitions apply generally, but their practical severity depends on the underlying graph. For Willow, the higher coordination number and denser loop structure make the truncation problem more demanding (Rudolph et al., 15 Jul 2025).

The same dependence appears in sampling complexity. For square-lattice-like geometry such as Willow with $z = 4$1, and for $z = 4$2, the cost of generating $z = 4$3 samples is given as

$z = 4$4

when partitioning by columns or rows. For heavy-hex with $z = 4$5, the scaling improves to

$z = 4$6

This is one of the paper’s clearest quantitative statements that Willow is harder to simulate classically because the extra coordination number directly increases the scaling exponents (Rudolph et al., 15 Jul 2025).

4. Belief propagation, boundary MPS, and the effect of loops

To improve truncation quality, the paper computes message tensors using belief propagation (BP) (Rudolph et al., 15 Jul 2025). BP is exact when there are no loops. That criterion is decisive for the Willow geometry, because the paper repeatedly emphasizes that Willow contains many short loops, so BP alone can fail badly.

To go beyond BP, the work employs a generalized boundary Matrix Product State contraction algorithm. The method groups the planar tensor network into partitions $z = 4$7, $z = 4$8, chosen so that the partitions form a line. The network is then contracted sequentially using an MPS of maximum bond dimension $z = 4$9, with MPS-MPO contractions fitted along the partitions; in the limit $z=4$0, the contraction becomes exact (Rudolph et al., 15 Jul 2025). The paper emphasizes that this generalized boundary-MPS construction applies to any planar tensor network, not only to rectangular square-lattice PEPS, and that it is used for both norm contractions $z=4$1 and amplitude contractions $z=4$2.

The Willow topology is particularly challenging for this workflow. The paper states that at the relevant system size, “the bond dimension of the Willow tensor network is notably larger than that typically considered in literature for square-lattice tensor networks, and extracting accurate information from it, with current methods, beyond the belief propagation approximation is very challenging” (Rudolph et al., 15 Jul 2025). By contrast, the same is “not true for the heavy-hex topology” because lower connectivity makes the boundary-MPS scheme scale more favorably in $z=4$3 and because the larger loops mean only minimal corrections to BP are needed (Rudolph et al., 15 Jul 2025).

This contrast is central to the article’s interpretation of Willow. The challenge is not only that contractions are more expensive in the abstract, but that the geometry specifically degrades the effectiveness of approximations that become accurate on weakly loopy or tree-like graphs.

5. Loop-correlation diagnostics and Willow-specific behavior

A geometry-sensitive diagnostic in the paper is an approximate BP-error proxy based on primitive loops (Rudolph et al., 15 Jul 2025). For each loop $z=4$4, the paper defines

$z=4$5

where the $z=4$6 are transfer-matrix eigenvalues. These are then averaged over primitive loops to obtain $z=4$7, which is used as a first-order approximation to BP error and as a proxy for loop-correlation strength (Rudolph et al., 15 Jul 2025).

Willow exhibits rapid loop buildup under this metric. The paper states that “The Willow topology is notably different, even at shallower circuit depths,” and that “It becomes clear that the Willow topology generates significantly stronger loop correlations than the heavy-hex topology, even at a third of the circuit depth” (Rudolph et al., 15 Jul 2025). The loop-correlation plot is summarized by the statement: “We observe a drastically larger BP error (many orders of magnitude) for the Willow square-lattice versus the heavy-hexagonal lattice” (Rudolph et al., 15 Jul 2025).

The paper further notes a heuristic expectation that larger heavy-hex loops might suppress correlations roughly as

$z=4$8

but reports that the observed difference is even stronger and may indicate some kind of “loop interference” effect (Rudolph et al., 15 Jul 2025). Since the paper presents this as a suggestion, a cautious reading is that Willow’s short-loop structure does more than merely increase loop count; it may qualitatively amplify the way correlations interact under contraction.

6. Local observables, sampling quality, and misconceptions about hardness

The most important Willow-specific conclusion is that loop correlations build up rapidly on this geometry and significantly affect even local observables (Rudolph et al., 15 Jul 2025). In the Heisenberg quench study, the paper reports that at 7 layers, local $z=4$9 expectation values on Willow can be converged, but require $U = \prod_{l=1}^L U_l,$0 on the order of the state bond dimension $U = \prod_{l=1}^L U_l,$1. At 15 layers, it becomes “very difficult to converge the single-site expectation value with either direct computation or sample-based computation” (Rudolph et al., 15 Jul 2025).

This stands in direct contrast with heavy-hex. The paper states that on heavy-hex, local $U = \prod_{l=1}^L U_l,$2 expectation values are almost insensitive to $U = \prod_{l=1}^L U_l,$3, whereas on Willow they depend strongly on the boundary-MPS dimension (Rudolph et al., 15 Jul 2025). In other words, even local one-site observables become sensitive to contraction accuracy on Willow at comparatively shallow depths.

The same distinction appears in the paper’s treatment of sampling. It defines the sampled distribution $U = \prod_{l=1}^L U_l,$4 and the exact tensor-network distribution

$U = \prod_{l=1}^L U_l,$5

and measures sample quality using the Kullback-Leibler divergence

$U = \prod_{l=1}^L U_l,$6

It also computes the unbiased estimator

$U = \prod_{l=1}^L U_l,$7

For the hardest Willow case, the paper notes that samples at $U = \prod_{l=1}^L U_l,$8 can have probabilities “off by dozens of orders of magnitude,” while larger $U = \prod_{l=1}^L U_l,$9 improves them markedly (Rudolph et al., 15 Jul 2025).

A common misconception addressed implicitly by these results is that the difficulty of recovering local observables and the difficulty of recovering the full distribution must track each other closely. The paper states the opposite: local observables can be recovered before the full distribution is fully accurate (Rudolph et al., 15 Jul 2025). Another misconception is that depth and qubit count alone determine hardness. The Willow results show that geometry, specifically coordination number and primitive-loop structure, can dominate simulability even before very large depth is reached.

Feature	Willow	Heavy-hex
Coordination number	$E_k$0	$E_k$1
Graph character	Square-lattice-like, planar, denser	Sparse, tree-like, low-loop
BP and boundary-MPS behavior	Strong loop correlations; BP alone can fail badly	BP already remarkably accurate; minimal corrections often sufficient

7. Significance for classical simulability and quantum-advantage assessment

The paper’s broader message is that processor geometry is a primary determinant of classical simulability for near-term quantum circuits (Rudolph et al., 15 Jul 2025). Heavy-hex’s lower coordination and larger loops make BP and low-$E_k$2 boundary MPS remarkably effective, even at relatively deep circuits. Willow’s denser square-lattice structure, by contrast, generates short loops that rapidly frustrate BP and require more expensive contraction.

Several broader implications are stated directly. First, circuit hardness is not only about depth and qubit count, but also about geometry. Second, local observables may remain easy longer than full sampling distributions, so claims of advantage should distinguish between tasks. Third, geometry can either suppress or amplify loop correlations, which are a major source of tensor-network difficulty. Fourth, a processor like Willow may reach regimes where classical tensor-network methods become much more costly at lower depth than heavy-hex devices of similar size (Rudolph et al., 15 Jul 2025).

Within that perspective, Willow functions as a case study in how a square-lattice-like superconducting layout can stress state-of-the-art classical simulators through its loop structure rather than through size alone. The conclusion drawn in the paper is that geometry can make a deep circuit classically easy or hard even before one reaches very large depth, and that any assessment of quantum advantage on superconducting processors must therefore account carefully for the underlying layout (Rudolph et al., 15 Jul 2025).