arXiv:2610.30087
Injection is all you need: parallel universal logic in high-rate concatenated self-dual codes
Joseph Sullivan, David Amaro, Michael A. Perlin
Abstract
We propose a quantum architecture that achieves addressable and parallel universal fault-tolerant logical in high-rate codes. The codes are obtained from the flexible concatenation of small-distance high-rate self-dual CSS codes. Addressable and parallel universal logic on the resulting $K$ logical qubits is injected by fault-tolerantly preparing and then consuming diverse logical $K$-qubit CSS stabilizer states and logical $\lvert H \rangle^{\otimes K}$ global magic states. Flag-at-origin is used to prepare the stabilizer resource states while a concatenated cultivation protocol prepares the global magic states. These resource states are consumed by Steane-QEC or Knill-QEC teleportation gadgets, which simultaneously perform quantum error correction. Any quantum algorithm with $T$-depth $D_T$ can be executed by consuming $12D_T$ or fewer resource states. We benchmark the fault-tolerant preparation of the resource states for several two-level concatenations of the $[[k+4,k,2]]$ H codes and the Hamming $[[15,7,3]]$ code. In particular, we find that the concatenated cultivation offers at least one order of magnitude spacetime reduction compared to cultivation in the surface code.
AI Summary
High-rate codes encode multiple logical qubits per block, but addressable gates can restrict computational parallelism. This paper develops an injection-based architecture for concatenated self-dual CSS codes with transversal global logical $H$ and $S$ gates, shifting addressability into resource-state preparation. A Bruhat decomposition reduces any logical Clifford circuit to at most nine injections using CSS stabilizer resource states, prepared through flag-at-origin rather than distillation. Global $|H\rangle^{\otimes K}$ magic states enable addressable parallel $T$ layers, giving a stated bound of $12D_T$ resource states for a circuit of $T$-depth $D_T$. Steane- and Knill-style injection gadgets also extract error syndromes. A concatenated cultivation protocol cross-checks Hadamard parities across encoding levels, suppressing undetectable logical errors to $O(p^{2^L})$ at concatenation level $L$; decoding can reduce the effective suppression. Benchmarks cover two-level codes encoding 4, 8 and 14 logical qubits. Magic-state results use Clifford proxy circuits, not direct non-Clifford simulations. At physical error rate $p=5\times10^{-4}$, reported block error rates range from $1.14\times10^{-10}$ to $2.21\times10^{-8}$, with discard rates of 41–87%. The first two examples have roughly 12–15-fold lower estimated spacetime cost than the compared surface-code cultivation protocol. The benchmarks characterize individual resource states, not complete computations. Limitations include decoder-heralded rejection that aborts an ongoing computation, runtime-dependent Clifford corrections after $T$ injection, high rejection for some Clifford resources, and lookup-table decoders that do not scale. End-to-end compilation, scheduling and resource accounting remain open.
Error found: The stated injection accounting gives at most 12D_T+9 resource states for a general Clifford+T circuit, unless an additional boundary-Clifford convention or absorption argument is supplied. The third cultivation benchmark establishes only a greater-than-1.89 cost advantage, not the greater-than-ten advantage asserted for every benchmark. The numerical magic-state performance comparison relies on Clifford proxy simulations without an established quantitative transfer to the actual non-Clifford cultivation circuits. The concatenation [[6,2,2]]∘[[15,7,3]] has parameters [[90,14,6]], inconsistent with its repeated labeling as [[96,14,6]].
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
- Comments
- 21 pages, 13 figues