Fault-Tolerant Renormalization Group Decoder for Abelian Topological Codes

I. INTRODUCTION

• Topological codes are geometrically constrained quantum error-correcting codes whose local checks can support fault-tolerant thresholds, transversal Clifford gates, and efficient decoding.The paper focuses on decoding as one component of their potential use in quantum memories and computers.Source: §I · p. 1
• The paper adapts its 2D renormalization-group decoder to faulty syndrome measurements, treated for topological codes by adding a time dimension.It also describes the approach in physical terms applicable to codes supporting Abelian anyons, not only stabilizer codes.Source: §I · p. 1

II. HEURISTIC PHYSICAL DESCRIPTION

• The heuristic description models information loss as a particle-antiparticle pair moving around a nontrivial cycle and fusing with its partner.Excitations carry conserved Abelian charges with deterministic fusion rules.Source: A. Decoding problem · p. 2

A. Decoding problem

• Decoding infers the most likely homology of particle world-lines from two consecutive location snapshots, equivalently their net charge flow across gauge lines.The gauge lines represent the torus’s two nontrivial cycles.Source: A. Decoding problem · p. 2

B. RG algorithm

• Each cell’s boundary currents are assigned probabilities conditioned on its observed charge configuration, using the local noise model.Current values represent possible charge flows across the cell’s north, east, south, and west walls.Source: B. RG algorithm · pp. 2–3
• Boundary-current values are vectors of elementary-charge components, with each component defined modulo the order of its charge.For the toric code, the elementary charges e and m each have order 2.Source: B. RG algorithm · p. 2
• The algorithm repeatedly merges four neighboring cells and sums compatible bulk processes to obtain boundary-current distributions for larger cells.At the final scale, currents across walls corresponding to the torus’s nontrivial cycles determine the decoding.Source: B. RG algorithm · p. 3
• Charge conservation and equality of opposite currents on shared walls create global correlations that make the exact joint distribution inefficient to compute.The cell constraint equates total boundary current with the observed charge.Source: B. RG algorithm · p. 3
• The algorithm approximates shared-wall consistency by requiring neighboring cell distributions to have matching marginals, enforced heuristically with belief propagation.Cells store only north- and west-wall variables because the other wall variables are redundant under the shared-boundary constraint.Source: B. RG algorithm · p. 3
• Leaving one region’s charge unmeasured makes the unit-cell total undetermined, relaxing constraints on current variables until the next RG iteration.The modification is a bookkeeping change rather than a physical measurement change.Source: B. RG algorithm · p. 3

C. Fault-tolerant decoding

• Repeated measurements leave changes ambiguous between intervening physical errors and faulty measurements; the space-time model represents them as cube charges.Source: C. Fault-tolerant decoding · p. 3
• Space-like qubit errors create excitations in overlapping cubes, whereas faulty measurements create excitations in cubes separated by the measurement.Source: Figure 3 · p. 4
• The fault-tolerant decoding problem differs from perfect-measurement decoding only by lattice dimension, so the RG algorithm applies directly.Source: C. Fault-tolerant decoding · p. 4

III. FORMAL DESCRIPTION FOR KITAEV’S TORIC CODE

• The formal KTC model uses an ℓ×ℓ periodic square lattice with two qubits per site and site and plaquette stabilizer generators.Source: §III · p. 4, A. Model · p. 4

A. Model

• Figure 4 assigns edge qubits to sites and depicts site X and plaquette Z stabilizers as blue and green strings.Source: Figure 4 · p. 4
• Pauli errors flip stabilizer eigenvalues into electric-charge or magnetic-flux excitations, while local memoryless Pauli channels specify their dynamics.Source: A. Model · p. 4

B. Decoding problem

• A correction succeeds only when it is homologically equivalent to the error; otherwise recovery can enact a logical operation.Source: B. Decoding problem · p. 5
• For independent uniform noise, decoding reduces to lowest weight, and the Perfect Matching Algorithm performs this task in O(ℓ^6) complexity.Source: B. Decoding problem · p. 5
• Optimal decoding sums error probabilities over stabilizer and logical equivalence classes, but is intractable because |S| grows exponentially with system size.Source: B. Decoding problem · p. 5
• The RG visualization groups 2×2 cells; renormalized generator eigenvalues encode region charge, while the noise model encodes net boundary charge flow.Source: Figure 5 · p. 5

C. RG decoding algorithm

• The 2 × 2 unit cell contains 12 qubits, includes all qubits participating in measured excitations, and leaves out one magnetic and one electric operator.Shared qubits between neighbouring cells correlate their current variables.Source: C. RG decoding algorithm · p. 5
• The basis separates six stabilizer excitation measurements, canonical pure errors, logical operators tracking boundary currents, and E operators changing the omitted charges.The logical operators track magnetic current through the north and west walls and electric current through the east and south walls.Source: C. RG decoding algorithm · pp. 5–6
• Figure 6 labels neighbouring cells’ qubits, including shared qubits that have the same physical identity in both cells.For example, qubit 6 in L is qubit 9 in R, and qubit 10 in L is qubit 1 in R.Source: Figure 6 · p. 6
• Each RG iteration sums probabilities of equivalent processes consistent with observed excitations to produce a distribution over boundary currents.For example, distinct Pauli errors can contribute to the same magnetic current when they decompose into the same logical operator and pure error.Source: C. RG decoding algorithm · p. 6

D. Belief propagation

• The per-qubit marginal is computed from the unit-cell Pauli-group distribution, accounting for correlations introduced by RG.With an independent base prior, the marginal reduces to the qubit’s noise model.Source: D. Belief propagation · p. 6
• Shared qubits can receive different locally inferred conditional distributions, so the decoder relaxes global consistency to agreement among their marginal distributions.A shared qubit between magnetic fluxes illustrates how its global error probability can be sharper than either cell’s local inference.Source: D. Belief propagation · p. 6
• Belief propagation enforces marginal agreement by exchanging one message per shared qubit between neighbouring cells.Outgoing messages initially equal the cell’s marginal error probabilities; a few rounds, such as three, can precede an RG step.Source: D. Belief propagation · pp. 6–7

E. Fault-tolerant decoding

• Space-like errors affect qubits between measurements, while time-like errors invert measurement outcomes; consecutive syndrome differences combine these errors.The 3D cubic lattice represents spatial and temporal errors as a history, with bits on faces and three orientations H, V, and T.Source: E. Fault-tolerant decoding · p. 7, Figure 7 · p. 7
• The 3D decoding approximation seeks the most likely equivalent error history consistent with the syndrome; perfect matching instead finds the most probable consistent history.The stated 3D setup assumes bit-flip errors, while the same method also applies to Z errors and can consider both error types simultaneously, including Y errors.Source: E. Fault-tolerant decoding · p. 7
• Periodic spatial boundaries create time-like nontrivial loops that are not true memory corruptions; this distinction does not arise with open boundaries.The authors identify these periodic boundaries as an unphysical simplification for presentation and numerical simulations.Source: E. Fault-tolerant decoding · p. 7
• A 2 × 2 × 2 cell has 33 qubits and seven check constraints, requiring a sum over 26-bit configurations, so the decoder uses smaller cells.The cost of summing over bulk processes grows exponentially with the number of qubits in a cell.Source: E. Fault-tolerant decoding · p. 7
• The 2 × 1 × 1 cell renormalizes one lattice dimension at a time, with rotations used to process the other two dimensions.In this cell, qubits lie on faces and qubits 1, 2, 6, and 7 are shared between cells and participate in belief propagation.Source: E. Fault-tolerant decoding · p. 7, Figure 8 · p. 8

IV. NUMERICAL RESULTS

• 1.9(4)% for the 2 × 2 × 1 cell and 1.8(2)% for the 2 × 1 × 1 cell, versus 2.9% for perfect matching under the same error model.The threshold estimates use independent bit-flip noise on qubits and measurements; the larger-cell estimate has finite-size effects for ℓ=8.Source: §IV · p. 8, Figure 10 · p. 8, Figure 11 · p. 9
• The 2 × 2 × 1-cell crossing point has limited support because decoding was practical only for ℓ=16 and ℓ=64.That cell is compatible only with lattice sizes that are powers of four, and the crossing of those two curves gives little confidence in the threshold estimate.Source: §IV · p. 8
• The RG scheme has complexity proportional to the lattice space-time volume and can be parallelized to logarithmic time, but its unit-cell-dependent constant can be prohibitive.The parallel version was not implemented.Source: §IV · p. 8
• Hybrid decoding with a 2 × 2 × 1 cell until the last step and a 2 × 1 × 1 cell at the last step yields a crossing consistent with all four curves.The authors interpret this agreement as support for combining unit-cell shapes for different lattice shapes deep in the RG flow.Source: §IV · p. 9
• No significant threshold anisotropy was observed across the three directions, but better statistics are needed to quantify possible subtler effects.The analysis considered marginal error rates by direction; possible differences in scaling exponents remain unresolved.Source: §IV · p. 9

V. CONCLUSION AND OUTLOOK

• The algorithm targets fault-tolerant decoding of topological quantum error-correcting codes supporting Abelian anyonic excitations.Source: §V · p. 9
• The extension of prior work permits error correction in the presence of faulty measurements.Source: §V · p. 9
• The numerical benchmark finds a fault-tolerant error threshold of nearly 2%, in the same ballpark as other leading techniques.Source: §V · p. 9

A. Relation to other work

• The color-code RG decoder of Sarvepalli and Raussendorf splits overlapping stabilizer generators into cell-local parts, using auxiliary binary variables instead of shared qubits.Their decoder achieves 7.8%, compared with 8.7% for mapping the code to multiple Kitaev toric codes and decoding them with this RG algorithm [13].Source: A. Relation to other work · p. 9
• Bravyi and Haah’s hard-decision RG decoder proves a finite threshold, while this paper’s soft-decision approach is supported only by numerical evidence.The hard-decision decoder achieves 6.7%; the passage also notes prior results that soft decoders can achieve higher thresholds and greater noise suppression below threshold.Source: A. Relation to other work · p. 9
• Wootton and Loss’s Monte Carlo decoder achieves 18.5%, versus 16.4% for this method under the same noise model, but its runtime scales morally exponentially with lattice size.Source: A. Relation to other work · p. 9
• Fowler, Whiteside, and Hollenberg’s parallelized perfect-matching implementation runs in constant average time without performance loss, but its runtime is not predetermined.The authors caution that on very large lattices, the probability that at least one cluster node takes at least any finite time approaches one.Source: A. Relation to other work · pp. 9–10

B. Extensions

• Rounding current-variable probabilities to binary distributions gives a conceptual hard-decision degradation of this algorithm, and intermediate trinary or truncated-probability schemes are also possible.The binary choice retains only the maximizing probability; a trinary option can use a flag to signal a potential error, as in Knill’s scheme.Source: B. Extensions · p. 10
• Larger unit cells improve error correction but make exact summation of bulk processes exponential in cell volume.Approximate summation using Monte Carlo or tensor networks, or a transfer-matrix approach, could reduce the cost; these techniques are unnecessary for the small cells considered.Source: B. Extensions · p. 10
• The algorithm’s description extends to 2D subsystem codes with local stabilizer generators, including topological subsystem color codes, but excludes Bacon-Shor codes.Gauge charges can traverse a non-trivial cycle without changing the ground state, so their current need not be monitored; an extra sum over these harmless processes yields a reported 1.95% threshold for the topological subsystem color code.Source: B. Extensions · p. 10

• A probability distribution over the n-qubit Pauli group can be represented, relative to a generating set, as a distribution over 2n binary variables.Source: Appendix A · p. 10
• Standard Bayesian calculus defines marginal and conditional probabilities over those variables, including marginalizing over x4 through x2n and conditioning x1 and x2 on x3.Source: Appendix A · pp. 10–11
• These probabilities depend on the chosen generating basis, can be converted to a different basis, and can be relabeled using the corresponding generators.Source: Appendix A · p. 11
• Subgroups of the Pauli group provide coarse-grained variables whose joint, marginal, and conditional probabilities are defined in the same natural way.Source: Appendix A · p. 11