Cy2Mixer: Cycle-Based Mixing in Graph Algorithms

• Cycle to Mixer (Cy2Mixer) is a framework that exploits cycle structures to enhance mixing efficiency and expressivity across graph neural networks, Markov processes, and quantum settings.
• It employs a triple-block design with temporal convolution, spatial message passing, and dedicated cycle message passing to capture rich topological features.
• Empirical benchmarks demonstrate improved forecasting accuracy and faster mixing times, validating the integration of cycle-based techniques in diverse algorithmic contexts.

Cycle to Mixer (Cy2Mixer) encompasses several independent yet thematically connected concepts that exploit cycle-based structures for enhanced mixing, expressiveness, or optimization in graph learning, Markov chains, and quantum-inspired algorithms. The term, as implemented in recent literature, designates both algorithmic primitives for spatio-temporal graph neural networks (GNNs) and mixing operators for classical/quantum walks on cycles. The following sections detail both the spatio-temporal neural network architecture and mixer designs for cycles in Markov and quantum walk settings, based strictly on published results.

1. Spatio-Temporal Graph Neural Architectures: The Cy2Mixer Model

Cy2Mixer (Lee et al., 2024) constitutes a spatio-temporal GNN architecture built for forecasting on real-world networks, exemplified by large-scale traffic data. The design operates by capturing dependencies through three parallel “gated” multilayer perceptron (MLP) blocks: a temporal block, a standard message-passing block, and a cycle message-passing block. Each block processes information relevant to a specific axis: time, local spatial adjacency, and higher-order cyclic topological structure, respectively.

At the core of each Cy2Mixer encoder layer (indexed by $\ell$), the hidden state tensor $H^{(\ell)}\in\mathbb{R}^{T\times N\times 2d_h}$ is linearly mixed and divided along the channel axis. The outputs of the three parallel blocks are gated in the element-wise form $\tilde{Z} = Z_1 \odot \varphi(Z_2)$, where $\varphi$ is specific to the block: a small convolution for the temporal block, classical message-passing for spatial, and message-passing over clique-augmented adjacency for cycles. Ultimately, the outputs are concatenated and projected for the next layer.

2. Formal Structure and Block Operations

2.1 Temporal Block

The temporal component applies a $3\times 3$ spatio-temporal convolution on $Z_2$ along the $(T, N)$ grid, yielding $$\varphi_{\text{temporal}}(Z_2) = \mathrm{Conv}_{3\times 3}(Z_2)$$. The output is gated with $Z_1$ as $$\tilde{Z}^{\text{temp}} = Z_1 \odot \varphi_{\text{temporal}}(Z_2)$$, in analogy with gated MLPs (gMLPs).

2.2 Standard Message-Passing Block

A classical message-passing neural network (MPNN) routine operates on the spatial adjacency $H^{(\ell)}\in\mathbb{R}^{T\times N\times 2d_h}$0 of the underlying graph $H^{(\ell)}\in\mathbb{R}^{T\times N\times 2d_h}$1, aggregating node messages from neighbors: $H^{(\ell)}\in\mathbb{R}^{T\times N\times 2d_h}$2 followed by a combination function for $H^{(\ell)}\in\mathbb{R}^{T\times N\times 2d_h}$3.

2.3 Cycle Message-Passing Block

A highlight of Cy2Mixer is the direct encoding of 1-dimensional topological cycles (homology classes) of $H^{(\ell)}\in\mathbb{R}^{T\times N\times 2d_h}$4 into a clique-adjacency matrix $H^{(\ell)}\in\mathbb{R}^{T\times N\times 2d_h}$5. This is constructed by:

• Computing a cycle basis $H^{(\ell)}\in\mathbb{R}^{T\times N\times 2d_h}$6 (e.g., via Paton’s algorithm)
• For each $H^{(\ell)}\in\mathbb{R}^{T\times N\times 2d_h}$7 with indicator $H^{(\ell)}\in\mathbb{R}^{T\times N\times 2d_h}$8, defining $H^{(\ell)}\in\mathbb{R}^{T\times N\times 2d_h}$9

MPNN is then executed over $\tilde{Z} = Z_1 \odot \varphi(Z_2)$0, and the output is gated as in the spatial case.

3. Mathematical Rationale: Topological Expressivity

The justification for the cycle-centric block leverages homotopy theory: for a spatio-temporal graph $\tilde{Z} = Z_1 \odot \varphi(Z_2)$1 (with $\tilde{Z} = Z_1 \odot \varphi(Z_2)$2 a time interval), the natural projection $\tilde{Z} = Z_1 \odot \varphi(Z_2)$3 induces isomorphisms on the first homology group: $\tilde{Z} = Z_1 \odot \varphi(Z_2)$4. Therefore, the cycle basis of $\tilde{Z} = Z_1 \odot \varphi(Z_2)$5 suffices to capture non-trivial cycles in spacetime, and models relying purely on edge-local connectivity (i.e., standard MPNNs with $\tilde{Z} = Z_1 \odot \varphi(Z_2)$6) are insufficient to capture this topological information. Augmenting with $\tilde{Z} = Z_1 \odot \varphi(Z_2)$7 ensures network capacity to distinguish between graphs based on cycle-level structure, not just edge adjacencies.

Ablation results confirm the empirical necessity of all three parallel blocks; removing any of temporal, spatial, or cycle components degrades accuracy on standard benchmarks (Lee et al., 2024).

4. Algorithmic Complexity and Implementation

Per-layer computational costs are as follows:

• Temporal convolution: $\tilde{Z} = Z_1 \odot \varphi(Z_2)$8 for $\tilde{Z} = Z_1 \odot \varphi(Z_2)$9
• MPNN over $\varphi$0 or $\varphi$1: each $\varphi$2
• Linear mixing: $\varphi$3
• Total space: $\varphi$4

Adjuvant "tiny self-attention" modules may be fused within blocks to support additional mixing, and both $\varphi$5 and $\varphi$6 are precomputed (cycle basis via $\varphi$7 algorithms).

5. Empirical Performance

Cy2Mixer achieves superior or state-of-the-art results on major spatio-temporal benchmarks such as PEMS04, PEMS08, and others. For example, on PEMS04 (307 nodes, 340 edges, 45 cycles), Cy2Mixer attains MAE 18.14, RMSE 30.02, MAPE 11.93% across a 12-step forecast, outperforming earlier models including DCRNN, STGCN, and STAEFormer. The inclusion of the cycle block specifically yields measurable gains over variants lacking cycle awareness (MAE drops from 18.81 without the cycle block to 18.14 with $\varphi$8). Training uses Adam, batch size 16, dropout 0.1–0.4, 50 epochs, early stopping on MAE (Lee et al., 2024).

6. Cycle-Based Mixers in Markov Chains on Graphs

In the context of Markov chains, “Cycle to Mixer” refers to constructions that manipulate random walks on cycles for enhanced mixing properties. Consider a basic $\varphi$9-cycle $3\times 3$0 perturbed by $3\times 3$1 random shortcut edges (matching $3\times 3$2 random hubs), then imposing a non-reversible transition scheme: with probability $3\times 3$3 move forward, $3\times 3$4 move backward, and $3\times 3$5 jump across a shortcut if present. This Markov kernel $3\times 3$6 admits the unique uniform stationary distribution but is not reversible for $3\times 3$7.

A main result is that adding only $3\times 3$8 extra random edges, the total-variation mixing time reduces from $3\times 3$9 (pure cycle) to

$Z_2$0

for fixed $Z_2$1 as $Z_2$2 (Feng et al., 2024). The mechanism combines deterministic drift along cycle arcs and rapid coverage via shortcut usage, with non-reversibility being essential for speedup. This stands in contrast to classical cycle-mixing without shortcuts where diffusion dominates and mixing is strictly quadratic in $Z_2$3.

Algorithmic simulation involves sampling shortcuts, generating a transition array, and stepping accordingly. The construction stands as a prototypical instance of a “Cy2Mixer” Markov process.

7. Cycle-to-Mixer in Quantum and QAOA Settings

Quantum approximate optimization algorithms (QAOA) employ mixers alternating with cost Hamiltonians. For cycle-structured problems such as the Hamiltonian cycle, “Cy2Mixer” refers to using Hamiltonians or superoperators derived from continuous-time walks on $Z_2$4 (Godsil, 2011). Specifically:

• The mixing operator is $Z_2$5, the adjacency of the $Z_2$6-cycle. The corresponding unitary $Z_2$7 acts as a full-circulant mixer, efficiently delocalizing quantum amplitude across the cycle.
• The average mixing matrix $Z_2$8 (Schur product), is doubly stochastic, positive semidefinite, and rational, contracting all states to uniform. For odd $Z_2$9, $(T, N)$0; for even $(T, N)$1, an additional permutation structure appears.
• In QAOA simulations for the Hamiltonian cycle (triangle/square), standard $(T, N)$2-mixers outperform $(T, N)$3-mixers, and non-trivial noise can sometimes enhance solution probability (Ye, 2023, Godsil, 2011).

The “cycle-to-mixer” paradigm in quantum settings thus refers to leveraging cycle symmetry for efficient, unbiased mixing within QAOA or quantum walk algorithms.

---

In summary, Cycle to Mixer (Cy2Mixer) formalizes the exploitation of cycles for enhanced expressive power or mixing speed in modern spatio-temporal GNNs (Lee et al., 2024), mixing-accelerated random walks on cycle-augmented graphs (Feng et al., 2024), and as quantum mixer operators in variational algorithms (Godsil, 2011). The central insight is that cycle-based topological invariants or symmetries can be harnessed in both classical and quantum settings to yield provable advantages in expressivity, discriminability, and convergence rates.