REVIEW 4 major objections 4 minor 1 cited by
Quantum error correction at ultra-low overhead
T0 review · 4 major / 4 minor · reviewed 2026-08-15 · deepseek-v4-flash
Pith's one-line read Cornucopia codes claim to break the rate–threshold–hardware trade-off: rate above 1/2, pseudo-threshold above 0.4%, and constant-depth syndrome extraction.
desk verdict A serious qLDPC co-design paper whose headline numbers are extrapolations that need a closer look. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing construction is a block-convolutional CSS code on the grid $\mathbb{Z}_3\times\mathbb{Z}_q$: twelve data blocks and six check blocks are coupled by twelve coordinate permutations $A_i,B_i$ that are pure cyclic shifts in the column direction, with only four of them performing row swaps or row-cyclic shifts. This restricted permutation class is what makes nonlocal checks executable as rigid atom translations, keeps the CNOT schedule to twelve layers independent of code length, and enables exact distance certification via row-pair enumeration. The algebraic identity carrying the logical-routing claim is the module decomposition $L_{X/Z}\simeq 18\cdot\mathbb{F}_2[C_q]\oplus 4\cdot\mathbf{1}$, which says that a global column shift acts on each logical space as eighteen independent length-$q$ cyclic registers together with four fixed modes.
What would settle it
Run a direct Monte Carlo simulation of one full syndrome cycle of the $[[2844,1426,18]]$ code at $p=0.001$ including the circuit-level noise model together with the complete routing sequence using the paper's own trap-transfer, acceleration, and idle-error parameters; if the observed logical error rate per logical qubit per cycle is not near $2.6\times10^{-16}$, or even below $p$, the extrapolation is falsified.
Extended reading notes
Core claim
The central claim is that the Cornucopia family of CSS quantum LDPC codes simultaneously achieves an encoding rate $r=k/n>1/2$, a circuit-level pseudo-threshold above $0.4\%$ under the standard depolarizing noise model, and a syndrome extraction cycle of exactly twelve parallel CNOT layers whose required atom rearrangements are only collective cyclic shifts and a few row swaps. The largest instance, $[[2844,1426,18]]$, uses 2,844 data qubits and 1,422 check qubits to encode 1,426 logical qubits, and Monte Carlo simulations with a relay-BP/BP-OSD decoder extrapolate to $p_L=2.6\times10^{-16}$ per logical qubit per cycle at $p=10^{-3}$ and $1.9\times10^{-31}$ at $p=10^{-4}$. The paper also proves equality of $X$- and $Z$-distances, certifies all listed distances by exhaustive symmetry-reduced search, and identifies a logical-routing automorphism in which one physical column shift acts as eighteen simultaneous cyclic logical registers plus four invariant modes.
Load-bearing premise
The headline logical error rates are extrapolations from Monte Carlo data at physical error rates 0.15%–0.25% down to 0.01%–0.1% under a noise model that ignores errors from the 10–16 ms of atomic motion per syndrome cycle; if the fitted curve bends upward or the motion introduces comparable errors, the quoted rates are not established.
Editorial extensions
If this is right
- The $[[2844,1426,18]]$ block is projected to reach $p_L=2.6\times10^{-16}$ per logical qubit per cycle at $p=0.1\%$, using 4,266 physical qubits total, i.e. roughly three physical qubits per logical qubit.
- Because a full syndrome cycle uses only twelve entangling layers and an estimated 10–16 ms of atom routing, the protocol fits within coherence times already demonstrated in neutral-atom experiments.
- If the pseudo-threshold holds, Cornucopia codes offer an order-of-magnitude reduction in physical-qubit overhead compared with bivariate bicycle codes for the same logical workload at comparable logical error rates.
- The column-shift automorphism gives a native logical-routing primitive: a single global physical shift moves logical information across eighteen cyclic registers, without deep SWAP networks or code deformation.
Reading between the lines
- The paper does not construct universal fault-tolerant logical gates: lattice surgery for Cornucopia codes is left open, so the ultra-low-overhead memory does not yet translate into a full fault-tolerant computation stack.
- Because each block contains hundreds or thousands of logical qubits, the block-level failure probability grows with $k$; readers comparing thresholds should separate per-logical-qubit error rates from the probability that all logical qubits in one block survive together.
- The $O(\sqrt{q})$ growth of ancilla routing time means the constant-depth claim is about entangling layers, not wall-clock time; at very large block sizes the cycle time could approach coherence limits, which is a testable scaling question the paper's estimates can be pushed to answer.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper introduces Cornucopia codes, a family of high-rate CSS qLDPC codes with a block-convolutional structure tailored to reconfigurable neutral-atom arrays. The codes have rate above 1/2, weight-12 checks, and a depth-12 syndrome extraction schedule implemented by rigid atom motions. The authors certify exact code distances, prove structural lemmas about the syndrome schedule and the logical action of a column-shift automorphism, and present Monte Carlo logical-memory simulations under a circuit-level depolarizing model for instances up to [[2844,1426,18]]. The headline claims are an encoding rate exceeding 1/2, a pseudo-threshold exceeding 0.4%, and extrapolated logical error rates of 2.6e-16 per logical qubit per cycle at p=1e-3 and 1.9e-31 at p=1e-4, which would give a more than tenfold physical-qubit overhead reduction relative to bivariate bicycle codes.
Significance. If the quantitative claims are supported, this would be a notable step toward demonstrating that highly efficient qLDPC codes with rate above 1/2 can be compatible with the connectivity constraints of neutral-atom hardware. The explicit construction, certified distances, constant-depth syndrome extraction schedule, and atom-routing estimates are valuable and largely self-contained. The paper also contributes a clear co-design principle and a logical-routing primitive. The main significance, however, rests on the pseudo-threshold and extrapolated error-rate figures, which are the load-bearing claims that the paper must defend; those figures currently rest on a fitting ansatz that is extrapolated well beyond the simulated range and appears inconsistent with the reported fit coefficients.
major comments (4)
- [Methods, 'Numerical simulation details'; Tab. S1] The claim that all simulated Cornucopia codes have a pseudo-threshold exceeding 0.4% is inconsistent with the paper's own fitting coefficients. For [[2844,1426,18]], substituting c0=10.36, c1=18370, c2=-2448603 into p_L = p^9 exp(c0+c1 p+c2 p^2) and solving p_L = p gives a crossing at approximately p=0.0038, below 0.004. The main text and Fig. 2a should be corrected to report the pseudo-threshold as determined by the simulated data or by the fit, and the statement 'exceeding 0.4%' must be revised.
- [Methods, 'Numerical simulation details'; Tab. 1; Tab. S1] The extrapolated logical error rates at p=1e-3 and p=1e-4 rely on the ansatz p_L = p^{d/2} exp(c0+c1 p+c2 p^2), whose exponent d/2 is not established for these codes. Tab. 1 reports only upper bounds on the circuit-level distance, and for [[1044,526,12]] this bound is 11 < 12, demonstrating that the effective sub-threshold exponent can be smaller than d/2. If the circuit-level distance of [[2844,1426,18]] were 16 or 14, the extrapolated p_L at p=1e-3 would increase by roughly 10^3 or 10^6, respectively. The quoted rates 2.6e-16 and 1.9e-31 are therefore model-dependent extrapolations, not established predictions; the manuscript should either obtain direct estimates at lower p, use the certified circuit-level distance upper bounds to give conservative bounds, or clearly label these numbers as upper-bound-free extrapolations with explicit sensitivity.
- [Methods, 'Numerical simulation details'; Tab. S2] The truncated-cycle simulations use N_c = 8 for the d=18 code and N_c = 6 for d<=12, well below the conventional N_c = d. The quantity p_L is derived from P_fail over N_c cycles as a per-cycle rate, which assumes that the logical failure probability accumulates independently per cycle. With N_c much smaller than the code distance, the per-cycle rate may not have converged to the steady-state value, and the fit could be biased. Please provide data for at least two or three N_c values per code to demonstrate convergence, or justify the truncated-cycle model quantitatively.
- [Methods, 'Implementation in neutral-atom arrays'; 'Circuit-level noise simulation and decoding'] The logical-error simulations neglect idle errors during the 10.4-16.2 ms atom-routing time per syndrome cycle, arguing only that this time is much smaller than the ~10 s coherence time. However, at p=1e-3 the per-cycle idle dephasing error is approximately 1.6e-3, comparable to the physical error rate per operation and to the break-even value. The statement that routing time is 'orders of magnitude shorter than coherence times' is true but does not make the idle error negligible relative to the depolarizing error budget. Please provide an explicit estimate of the idle-error contribution per cycle or include a decoherence term in the noise model for at least one representative code.
minor comments (4)
- [Methods, 'Numerical simulation details'; Tab. S1] The fitting procedure uses unweighted least squares in log space over data points with very different statistical weights (e.g., 5 failures out of 2e6 samples versus 26 failures out of 1000 samples for the largest code). Please report the fit residuals or use a weighted fit, and state the uncertainty of the extracted coefficients.
- [Main text, Fig. 2] The extrapolated lines in Fig. 2 are presented with the same visual weight as the simulated marker points; consider using a different line style or shaded region to mark the extrapolation region beyond the lowest simulated p.
- [Main text, Tab. 1] The table lists p_L values at p=1e-3 and p=1e-4 without a footnote that these are extrapolations from fits, not direct simulation results; such a footnote would prevent misinterpretation.
- [Abstract and Introduction] The phrase 'ultra-high encoding rate exceeding 1/2' is accurate for the presented instances but 'ultra-high' is subjective; consider using a more neutral phrasing such as 'encoding rate above 1/2' or 'rate greater than 1/2'.
Circularity Check
No significant circularity: the code construction, exact distance certification, and Monte Carlo logical performance are independent; the low-p logical rates are explicitly extrapolations from a fitted ansatz, not fitted inputs relabeled as predictions.
full rationale
The paper's central derivation chain is self-contained. Cornucopia codes are defined by explicit parity-check matrices in Methods Eqs. (1)-(6), with the CSS orthogonality condition verified algebraically; the distances in Table 1 are certified by exhaustive symmetry-reduced search, not assumed or imported from prior work. The headline logical error rates at p=10^-3 and p=10^-4 are not raw simulation outputs but are values of the fitting ansatz p_L = p^{d/2} exp(c0 + c1 p + c2 p^2) (Methods, 'Numerical simulation details'; Tab. S1), whose coefficients are least-squares fits to independent Monte Carlo data. For the largest [[2844,1426,18]] code the fitted range is p in [1.75e-3, 2.5e-3] (Tab. S2), so the reported p=1e-3 and p=1e-4 figures lie outside the fitting interval and are not used to determine the fit; they are extrapolations, and the paper explicitly labels them 'extrapolated' in the abstract and Table 1. The pseudo-threshold claim is likewise an extrapolation of the fitted curves to the break-even point, which is a robustness concern rather than a circular derivation; the same holds for the stated neglect of idling errors during atom routing. The load-bearing citations for the code construction (Refs. [47,48]) and for the fitting ansatz (Refs. [31,57]) are external to the present authors, so there is no self-citation chain on which the central claim depends. The only author-overlapping citation (Ref. [22]) is an experimental demonstration and is not load-bearing. The paper also benchmarks Cornucopia codes directly against bivariate bicycle and surface codes using the same noise model and decoder, providing independent empirical grounding. No definitional equivalence, fitted-parameter-as-prediction, or imported-uniqueness step could be identified from the text.
Assumptions & free parameters
free parameters (3)
- column-shift parameters s for A_i and B_i =
Per instance, e.g. [[2844,1426,18]]: A: 6,49,55,18,40,7; B: 24,41,78,53,68,21 (Extended Tab. 1)
- logical error rate fit coefficients c0,c1,c2 =
e.g. for [[2844,1426,18]]: 10.36, 18370.0, -2448603 (Tab. S1)
- truncated-cycle counts N_c =
6 for d<=12, 7 for d=14, 8 for d>=16
assumptions (6)
- domain assumption Uniform circuit-level depolarizing noise with independent faults on state prep, CNOT, measurement, reset, and readout
- ad hoc to paper The fit p_L = p^{d/2} exp(c0 + c1 p + c2 p^2) remains valid down to p = 1e-4
- domain assumption The exact distance certification exhaustively enumerates all logical operators of weight below d
- domain assumption The 12-layer CNOT schedule plus AOD rigid motions measures stabilizers without introducing additional faults
- domain assumption Relay-BP with BP-OSD fallback gives logical error rates close enough to true code performance for threshold comparisons
- standard math CSS stabilizer formalism and the constraint H_X H_Z^T = 0 define a valid code
Cite this review
Pith. "Pith review of Quantum error correction at ultra-low overhead." pith.science (2026). https://pith.science/paper/TV644WZD
@misc{pith2026260802773,
author = {Pith},
title = {Pith review of: Quantum error correction at ultra-low overhead},
year = {2026},
howpublished = {\url{https://pith.science/paper/TV644WZD}},
note = {Machine review of arXiv:2608.02773}
}
abstract
Suppressing errors is the central challenge for useful large-scale quantum computing. While quantum error correction promises a viable solution to this challenge, existing codes typically suffer from trade-offs among encoding efficiency, error threshold, and hardware feasibility. Here, we introduce Cornucopia codes, a family of practical, hardware-efficient quantum low-density parity-check codes that achieve an ultra-high encoding rate exceeding $1/2$ while maintaining a pseudo-threshold exceeding $0.4\%$ under the standard circuit-level noise model. Inspired by recent affine-permutation-based code constructions and the long-range connectivity available in reconfigurable neutral-atom arrays, we adopt a structured code geometry in which the code layout, atom rearrangement, and syndrome-extraction schedule are co-designed. This structure enables nonlocal syndrome measurements through simple, parallel atom rearrangements. A complete syndrome extraction cycle measures all $X$- and $Z$-type checks in parallel with $12$ entangling layers, independent of the code size. The resulting threshold is comparable to those of the surface code and bivariate bicycle codes. In particular, a single code block $[[2844,1426,18]]$ encodes $1{,}426$ distance-$18$ logical qubits, achieving an extrapolated logical error rate of $2.6\times10^{-16}$ ($1.9\times10^{-31}$) per logical qubit per cycle, assuming the physical error rate of $0.1\%$ ($0.01\%$). By comparison, a bivariate bicycle code implementation would require more than $68{,}000$ physical qubits to encode the same number of logical qubits at a comparable logical error rate. These results bring demonstrations of ultra-low-overhead quantum error correction within the reach of near-term quantum processors.
Figures
Forward citations
Cited by 1 Pith paper
-
Quantum Tanner Codes at Moderate Blocklength
A search over non-abelian groups yields new moderate-blocklength quantum Tanner code instances whose randomized distance bounds exceed 20, with decoder pseudo-thresholds comparable to shorter codes.
Reviewed August 15, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.