Optimal Error Thresholds in 2D Toric Codes: Lattice Geometry and Duality Splitting

Beyond MWPM: why the optimal threshold matters

Earlier work established the toric code's mapping to the two-dimensional random-bond Ising model and used it to estimate thresholds under minimum-weight perfect matching, a suboptimal but fast decoder [2]. That line of work located the zero-temperature threshold at approximately 0.1031 on the square lattice and showed that the failure probability obeys finite-size scaling with correlation-length exponent near 1.46 [2]. The new paper argues that near-optimal decoders such as tensor-network decoders have made the maximum-likelihood code capacity the relevant benchmark, because it provides a universal baseline against which approximate decoders can be judged [1]. The authors therefore compute the full finite-temperature phase boundary of the RBIM and extract its intersection with the Nishimori line, which corresponds to the optimal threshold rather than the MWPM threshold [1].

The distinction is quantitative, not semantic. On the square lattice the new data-collapse analysis gives an optimal threshold of 0.10924 ± 0.00027, compared with the zero-temperature MWPM value of about 0.1034 from the same phase-boundary extrapolation [1]. That gap of roughly 0.006 in error probability is the headroom that better decoding can recover, and it sets the scale for interpreting the geometry-dependent results that follow.

Duality-driven splitting around the self-dual square lattice

The central result is that dual-lattice pairs split around the self-dual square-lattice threshold in opposite directions. The triangular lattice, with average coordination number six, reaches 0.1633 ± 0.0003, while its dual honeycomb lattice, with coordination three, drops to 0.0654 ± 0.0007 [1]. The dice and kagome lattices both have average coordination four, yet the dice lattice is more robust at 0.1175 ± 0.0006 versus 0.1000 ± 0.0005 for kagome [1]. The authors verify the generalized duality relation H₂(p_c) + H₂(p*_c) ≈ 1 across all pairs, with the sum staying within about one percent of unity [1].

The interpretation is that average coordination number sets the primary scale, but the detailed arrangement of vertices and plaquettes shifts the precise value. High-connectivity vertices penalize error propagation, while large plaquettes accommodate more error configurations without frustration [1]. A Taylor expansion around the self-dual point shows that the leading correction to the dual threshold is positive and quadratic, so any asymmetry in connectivity systematically increases the sum of primal and dual thresholds above twice the square-lattice value [1].

How the new thresholds compare with earlier estimates

The new thresholds are slightly lower than previous estimations across lattices, with offsets of about +0.054 and −0.044 relative to the square-lattice benchmark for triangular and honeycomb respectively [1]. The authors attribute the discrepancy to methodological differences and finite-size effects, noting that the earlier de Queiroz values are systematically higher [1]. The square-lattice benchmark itself agrees well with high-precision literature values near 0.10919 and the analytic conjecture of 0.109187, which supports the reliability of the replica-exchange approach [1].

Independent decoder work provides a useful cross-check on the square-lattice optimal threshold. A stabilizer-aware neural decoder reports a threshold of 10.99% under independent noise for the toric code, essentially reaching the maximum-likelihood bound estimated between 10.9% and 11.0% [5]. That agreement between a machine-learning decoder and the statistical-mechanics calculation strengthens the case that the Nishimori-line multicritical point is the correct target for optimal decoding [1][5].

What the threshold means physically

The threshold is not merely a decoder performance metric; it coincides with a separability transition in the decohered state. For the two-dimensional toric code under bit-flip errors, the decohered density matrix can be written as a convex sum of short-range entangled states for p > p_c, where p_c is related to the paramagnetic-ferromagnetic transition in the two-dimensional random-bond Ising model along the Nishimori line [3]. This connects the new geometry-dependent thresholds to a broader question about when topological order survives local decoherence [3].

The random-bond Ising model itself has a rich analytical structure that supports the new calculations. The Nishimori line exhibits enhanced supersymmetry, with transfer matrices having osp(2n+1|2n) symmetry compared with osp(2n|2n) elsewhere in the phase diagram [6]. That exact structure is what makes the multicritical point a well-defined target and gives confidence that the numerical thresholds are not artifacts of finite-size scaling [1][6].

Where the conclusion stops

The results are specific to two-dimensional toric codes under bit-flip or phase-flip errors, which map cleanly onto the random-bond Ising model [1]. The authors explicitly identify three-dimensional toric codes and other topological stabilizer families such as color codes and fracton architectures as future directions [1]. Hyperbolic lattices present a further boundary: on self-dual hyperbolic surfaces the dual random-bond Ising model is not equivalent to the original, unlike in the Euclidean case, so the duality-splitting framework may not transfer directly [7].

Decoder comparisons also come with caveats. The maximum-likelihood threshold is an upper bound on what any decoder can achieve, but practical decoders operate under circuit-level noise with measurement errors and correlated faults that the code-capacity model omits [1]. A distributed lattice-surgery study using minimum-weight perfect matching reports thresholds around 0.86% under a phenomenological noise model with noisy entangled pairs, far below the code-capacity values discussed here [4]. That gap illustrates how much of the theoretical threshold is consumed by realistic noise and decoding constraints, and it reinforces the authors' framing of maximum-likelihood capacity as a baseline rather than a hardware prediction [1][4].

About These Sources

This research page is built on 7 studies (6 peer-reviewed, 1 preprint) — published from 2000 to 2026, 4 from 2024 or later, collectively cited 214 times — selected as the most relevant from 13 studies that passed quality screening, drawn from 64 papers retrieved from a database of over 500 million.

Sources used in this answer