Why Quantum Error Correction Is the Most Expensive Part of a Quantum Computer
The mainstream route to useful quantum computing is the surface code: bundling many noisy physical qubits into a few reliable logical qubits. The industry-standard figure of merit: a single high-quality logical qubit can require over a thousand physical qubits. Redundancy is a tax.
Quantum LDPC codes (QLDPC) are seen as a tax cut: clever code constructions keep the check matrices sparse, dramatically improving encoding rates. But the field has been stuck for years on one problem: these codes only proved they can store quantum information, not that they can compute with it. When it came to logical gates (the lattice-surgery toolkit), the qubit savings evaporated and gate depths grew with code distance d.
Work published May 20 in Nature Communications (reported by Quantum Zeitgeist on August 25) sawed through that bottleneck: researchers at Canada's Photonic Inc., led by Malcolm et al., proposed the SHYPS codes, claimed to be the first QLDPC code family with efficient logical gates. Chief Quantum Officer Stephanie Simmons put it bluntly: "efficient QLDPC logic is no longer a theoretical promise — it is a demonstrated result."
One caveat up front: all results are circuit-level numerical simulations, not hardware demonstrations. But the simulated accounting is worth reading line by line.
What Is the SHYPS Code?
The name decomposes as Subsystem HYpergraph Product Simplex — a subsystem hypergraph product construction using classical simplex codes as raw material. The parameter family is written SHYPS(r), and for r ≥ 3 gives codes with parameters [(2^r − 1)², r², 2^(r−1)]: n physical qubits encoding k logical qubits at distance d. The paper's worked examples use two members: [49, 9, 4] and [225, 16, 8]. Forty-nine physical qubits carrying 9 logical qubits is an encoding rate near 1/5 — extravagant next to the surface code (which starts at k=1 and needs modules to scale).
The real weapon is its symmetry. The code family's automorphism group is on the order of |GL_r(2)|², translating to 2^O(k) fault-tolerant logical generators per block, each with physical circuit depth of just 1. In other words: the code's massive permutation symmetry can be used directly as logic gates, acting in a single layer of circuit depth.
Where Exactly Does "Computing" Get Cheap?
The standard way to perform logical gates is lattice surgery: patching two code blocks together and repeatedly measuring syndromes, with depth scaling linearly in code distance, O(md). SHYPS logical gates instead use code automorphisms and ZX duality, with a five-segment decomposition (D_Z–D_X–D_Z–D_X–D_Z) assembling the full Clifford group at depth independent of code distance.
How big is the difference? The paper's concrete numbers: at d=20, a CNOT takes roughly 40 logical cycles via lattice surgery versus roughly 4 for SHYPS. For an arbitrary worst-case Clifford acting on b·k logical qubits, SHYPS needs 4bk(1+o(1)) logical cycles — compared to 2bk + O(log²(bk)) for bare, unencoded qubits. The gap is compressed to constant factors, which is nearly unheard of in the world of error-correcting codes.
The Simulation Results
Two head-to-head comparisons, same error model:
| SHYPS | Surface code baseline | Result | Qubit savings | |---|---|---|---| | [49, 9, 4] | [[81, 9, 3]] | Outperforms | 2x | | [225, 16, 8] | [[784, 16, 7]] | Matches | 3.5x |
A tougher test: a random 18-logical-qubit Clifford compiled on two [49,9,4] code blocks into 63 depth-1 generators, run as a depth-126 full circuit (operators plus inverses), suppressed logical error rates at nearly the same level as pure storage. Translation: computing is as reliable as storing.
The decoder wasn't cut corners either: a small-window sliding BP+LSD decoder in (2,1) and (3,1) configurations, performing close to single-shot behavior — while the surface code baseline used the optimal MWPM decoder. Winning against the opponent's full-strength lineup is a respectable win.
The Cold Water — Colder Than Expected
The other column of the ledger:
- Lower pseudo-threshold. The [225,16,8] code's pseudo-threshold is around 0.35%, versus roughly 0.8% for the d=7 surface code baseline. SHYPS's error-rate curve only beats the surface code's slope once physical error rates drop below ~0.05%. Translation: it needs cleaner physical qubits to cash in its efficiency advantage.
- Connectivity hungry. The paper's efficiency arguments assume high-connectivity architectures. Photonic's own Entanglement First roadmap (silicon spin qubits with telecom-band photonic interconnects) targets exactly this — but that hardware line has to deliver first.
- T gates are billed separately. The full Clifford group is cheap; non-Clifford operations still go through the old magic-state-injection accounting.
- Simulation is still simulation. Distance-8 code blocks of a few hundred physical qubits have not yet faced real-machine wiring, crosstalk, or decoding latency.
- Paper: Computing Efficiently in QLDPC Codes, Malcolm et al., Nature Communications 17, 7286 (2026-05-20) https://www.nature.com/articles/s41467-026-73061-9
- Quantum Zeitgeist coverage: https://quantumzeitgeist.com/photonic-code-cuts-qubit-needs-quantum/
- Open-source code: https://github.com/PhotonicInc/ComputingEfficientlyInQLDPCCodes
Why the Buzz This Week
The paper came out in May; the heat arrived in late August when Quantum Zeitgeist re-covered it and Simmons began arguing that "architectures and timelines will be affected." An earlier report from the same outlet cited an extrapolated 20x qubit-saving figure, and the company has consistently steered toward "accelerated commercialization timelines."
This post's take: the most valuable part of this work is not the qubit savings per se, but that it shifts code selection from storage-metric-driven to compute-metric-driven. Previous code comparisons raced on thresholds and encoding rates; SHYPS pulls "logical clock speed" into the first-tier metrics. Other code teams (IBM's bivariate bicycle codes, Google's surface code improvements, Amazon's LDPC variants) will likely have to catch up — no gate-depth comparison, no seat at the practicality table.
References: