CAHR: Hypergraph Reconstruction in QEC

• CAHR is a topology-discovery method that reconstructs detector error models by leveraging exact correlation identities and a global, top-down pruning strategy.
• It formulates the inverse problem of hypergraph topology discovery from syndrome data using a three-stage algorithm that first identifies candidate hyperedges before pruning spurious ones.
• The method has demonstrated exact topology recovery in rotated surface codes and dense 2D color codes under realistic shot counts, emphasizing structural accuracy over precise parameter estimation.

Searching arXiv for the specified CAHR paper and closely related hypergraph reconstruction work. Correlation-Analysis-based Hypergraph Reconstruction (CAHR) is a topology-discovery method for reconstructing a detector error model (DEM) directly from syndrome data in fault-tolerant quantum error correction. In this setting, realistic circuit-level noise generates correlated multi-detector faults rather than only single-qubit stochastic noise, and those faults are represented as a hypergraph whose vertices are detectors and whose hyperedges are independent physical mechanisms that flip all detectors in the hyperedge simultaneously. CAHR addresses the discrete part of the inverse problem first—inferring which hyperedges exist, without introducing spurious ones—by combining exact correlation identities with a global, top-down pruning strategy. It was introduced for reconstruction of circuit-level noise in surface codes and color codes, with benchmark results showing exact topology recovery in the reported settings for a $d=5$ rotated surface code and a dense $d=5$, $r=4$ 2D color code (Ye et al., 15 Jun 2026).

1. Problem formulation and object of reconstruction

CAHR is defined around the reconstruction of a detector error model from experimental syndrome statistics. The underlying premise is that syndrome-extraction circuits in surface codes, color codes, and related QEC schemes generate correlated faults that are naturally represented as higher-order mechanisms, not merely as independent pairwise or single-detector events. In the DEM formalism used by CAHR, the noise model is a hypergraph $G=(V,E)$, where $V$ is the detector set and each hyperedge $e\in E$ is an independent physical mechanism acting on all detectors in $e$ simultaneously (Ye et al., 15 Jun 2026).

The reconstruction target is therefore structural before it is parametric. The immediate goal is to infer the hypergraph topology, namely which hyperedges exist. A subsequent goal is to estimate the corresponding probabilities $p_e$. The paper emphasizes that these two tasks are statistically different: topology discovery is combinatorial and structural, whereas probability calibration is a continuous parameter-estimation problem. This distinction is central to CAHR’s design, because dense correlated-noise settings make exact continuous inversion fragile even when the discrete structure is recoverable.

This formulation differs from reconstruction problems in which one observes an ordinary graph and infers a latent hypergraph behind it, or from time-series-based hypergraph inference. Here the observation is syndrome data from fault-tolerant quantum circuits, and the hidden object is a DEM hypergraph directly tied to physical fault mechanisms. A plausible implication is that CAHR belongs to a more specialized class of inverse problems in which algebraic consistency of observed correlations is as important as sparsity or statistical fit.

2. Exact correlation structure underlying CAHR

CAHR is built on an exact set of correlation equations. The syndrome bit $s_i\in\{0,1\}$ is mapped to a spin-like variable

$\sigma_i = 1 - 2s_i.$

For any detector subset $d=5$0, the multi-detector moment is

$d=5$1

The core identity states that

$d=5$2

so the moment factors over precisely those physical mechanisms whose overlap with $d=5$3 has odd parity (Ye et al., 15 Jun 2026).

The tensor-network derivation in the appendix writes the same relation as a contraction identity,

$d=5$4

with $d=5$5 and $d=5$6 as the projection states used to extract correlations from the tensor network. The significance of this formula is that the DEM is not inferred through an approximate phenomenological fit; rather, the syndrome moments obey an exact multiplicative structure under the model assumptions.

The paper also defines a recursive helper quantity

$d=5$7

and derives

$d=5$8

This makes explicit why lower-order inference depends on already-resolved higher-order “parent” mechanisms. The exact probability of a lower-order hyperedge can be resolved algebraically once all parent hyperedges $d=5$9 are known. That dependency is the mathematical reason that CAHR reconstructs from high degree downward rather than through low-order-first growth.

3. Algorithmic workflow and concurrent pruning

CAHR reconstructs topology in three stages. First, it computes pair correlations and uses them as a necessary geometric filter. The paper defines an apparent pair probability $r=4$0 from pair moments and notes that, in the small-noise regime,

$r=4$1

A real higher-order hyperedge must therefore project to a clique of positive pairwise correlations. Positive pairwise links are used only as a permissive candidate generator, and an initial conservative threshold $r=4$2 removes pairs with $r=4$3 (Ye et al., 15 Jun 2026).

Second, CAHR forms fully connected cliques up to a user-chosen maximum order $r=4$4 determined by the circuit or code structure. In the reported benchmarks, $r=4$5 for the rotated surface code and $r=4$6 for the 2D color code. This yields a candidate hypergraph $r=4$7 containing all sub-hyperedges of those cliques.

Third, the algorithm performs global correlation inversion on this candidate set in descending order of degree, pruning a hyperedge immediately if its inferred probability falls below the global threshold $r=4$8. This is the algorithm’s central device, termed “concurrent pruning.” The rationale is structural: false high-order candidates are removed before their spurious contributions can propagate downward through the recursion and contaminate lower-order inference.

The paper contrasts this with standard greedy, order-by-order inference. Greedy methods fit edges layer-by-layer, so unresolved higher-order mechanisms can leak into lower-order statistics, producing false positives and suppressing real mechanisms. CAHR reverses that logic: it begins with a permissive topology, then resolves and prunes from high degree to low degree. Once a spurious parent hyperedge is removed, its contribution is zeroed before the recursion continues. This makes the reconstruction globally consistent rather than locally residual-driven.

4. Benchmark behavior in surface-code and color-code settings

The reported benchmarks evaluate CAHR on two fault-tolerant code families. For the $r=4$9 rotated surface code with $G=(V,E)$0 and physical error rate $G=(V,E)$1, the method reconstructs the full hypergraph with zero false positives and zero false negatives at $G=(V,E)$2 shots. For the dense $G=(V,E)$3, $G=(V,E)$4 2D color code, which contains up to 8-body hyperedges, the method reaches zero false positives and zero false negatives at $G=(V,E)$5 shots (Ye et al., 15 Jun 2026).

Benchmark	Shots	Reported reconstruction
Rotated surface code	$G=(V,E)$6	5262 inferred edges, 3944 false positives, 361 false negatives
Rotated surface code	$G=(V,E)$7	2219 edges, 542 false positives, 2 false negatives
Rotated surface code	$G=(V,E)$8	zero false positives, zero false negatives
2D color code	$G=(V,E)$9	1106 edges, 422 false positives, 159 false negatives
2D color code	$V$0	860 edges, 31 false positives, 14 false negatives
2D color code	$V$1	843 edges, zero false positives, zero false negatives

These numbers illustrate two features of the method. First, CAHR is explicitly sample-limited: in the undersampled regime, the candidate graph can become badly polluted by spurious pair correlations, which then inflate the clique search. Second, once the shot count is sufficiently high, the global inversion-and-pruning strategy eliminates those artifacts and recovers the true topology exactly in the benchmark settings.

The paper also notes that clique search is only a bottleneck in the undersampled regime. For the $V$2 surface code, CPU time drops dramatically as shot count increases, reflecting the disappearance of ghost cliques once spurious pair correlations are suppressed. This suggests that the dominant computational difficulty is not intrinsic clique enumeration alone, but clique proliferation induced by finite-statistics artifacts.

5. Variance cascade, two-stage inference, and decoding relevance

A central conceptual result associated with CAHR is the “variance cascade.” Because lower-order probabilities are computed by subtracting off higher-order parent contributions, finite-sample uncertainty in high-degree estimates propagates into all descendants. Under a Gaussian approximation, the paper gives the approximate additive law

$V$3

The consequence is that absolute variance can accumulate linearly from high- to low-degree mechanisms, especially in dense hypergraphs (Ye et al., 15 Jun 2026).

In the reported figures, this appears as inflation of absolute error as degree decreases. The effect is much stronger in the dense color code than in the surface code because the color code has more high-degree parent hyperedges and a deeper subtraction hierarchy. The practical implication drawn in the paper is not that topology recovery fails, but that exact continuous probability estimation becomes unstable in dense codes. Lower-degree inferred probabilities can acquire large absolute errors or even become negative under insufficient statistics, which in turn increases false negatives when non-negativity is enforced.

This motivates the paper’s two-stage inference paradigm. Stage 1 is CAHR itself: extract the discrete topology, meaning which hyperedges exist. Stage 2 is continuous probability refinement on that fixed topology, using maximum likelihood, learning-based optimization, or another statistical estimator. The proposed separation is not merely procedural; it is tied directly to the variance-cascade limitation of exact continuous inversion.

The decoder experiments reinforce this distinction. Using BP-OSD decoding, the paper compares the “Inference DEM” from CAHR’s full finite-statistics reconstruction, a “Given DEM Topology” ablation in which the exact ideal topology is supplied but probabilities are still estimated from finite data, and the ideal DEM. The reported conclusion is that structural exactness matters more than high-precision parameter estimation. At low statistics, the Inference DEM performs poorly because false-positive short cycles corrupt belief propagation. By contrast, the Given DEM Topology remains close to the ideal logical error rate even with highly noisy parameter estimates, because it avoids the structural damage caused by wrong edges.

A common misunderstanding is therefore that CAHR is primarily a probability-estimation scheme. The paper argues the opposite emphasis in dense correlated-noise settings: topology discovery is the robust first step, while continuous calibration should be delegated to a second-stage estimator after the search space has been compressed to the correct hypergraph.

6. Assumptions, limitations, and relation to broader hypergraph-reconstruction research

The formulation underlying CAHR assumes Pauli noise. It does not claim quantitative benchmarks for explicit non-Pauli models such as leakage or coherent errors, although it suggests that some non-Pauli signatures may still be partially absorbed at the syndrome level. Extending the tensor-network formulation to those cases is identified as an open problem (Ye et al., 15 Jun 2026).

Within the broader reconstruction literature, CAHR occupies a distinct position. Bayesian hypergraph reconstruction from ordinary graph projections treats the observed graph as a projection of a latent hypergraph and uses a parsimonious posterior to prefer sparse higher-order explanations (Young et al., 2020). Reconstruction from noisy graph projections studies sharp thresholds and a detection–reconstruction gap, with a clique estimator recovering $V$4-uniform hyperedges only in a restricted signal-to-noise regime (Gong et al., 21 Jun 2025). Supervised remediation of hypergraph projection loss uses the $V$5-alignment statistic and the SHyRe pipeline to reconstruct hyperedges from projected graphs when same-domain training hypergraphs are available (Wang et al., 2024).

Other neighboring directions address different data modalities. Hypergraph reconstruction from dynamics uses Taylor-based Hypergraph Inference using SINDy (THIS) to infer higher-order effective interactions from time-series data and derivatives rather than from correlation identities on syndrome statistics (Delabays et al., 2024). Hypergraph reconstruction from noisy pairwise observations formulates a Bayesian model with 2-edges and 3-edges and uses a Metropolis-Hastings-within-Gibbs sampler to infer latent higher-order structure from Poisson pairwise counts (Lizotte et al., 2022). GraphCroc, although not a CAHR method, analyzes self-correlation versus cross-correlation decoders for structural reconstruction and identifies transfer ideas relevant to hypergraph incidence reconstruction, particularly the use of two embedding spaces and balanced sparse reconstruction objectives (Duan et al., 2024).

This comparison suggests that “correlation-analysis-based” reconstruction is not a single methodology across fields. In CAHR for quantum error correction, the term refers specifically to exact multi-detector correlation equations, global inversion, and top-down pruning for DEM topology recovery. In other literatures, related phrases may refer instead to Bayesian projection inversion, clique-based recovery from noisy projections, supervised distributional alignment, or dependency patterns in dynamical data. The shared theme is inference of latent higher-order structure from lower-order observables, but the identifiability assumptions, noise models, and algorithmic guarantees differ substantially.