Quantum computing · Theory and numerics (AI-assisted)
The limits of magic-state cultivation
arXiv:2611.24852v1 ↗PDF ↗Submitted 18 November 2026
- Question
- Magic-state cultivation prepares a \(T\) state in a small colour code and checks it repeatedly with physical \(T\) gates, keeping it only if every check passes. Fault tolerance is normally analysed with Pauli algebra, which fails at non-Clifford gates, so cultivation circuits have been designed and benchmarked with every \(T\) gate replaced by an \(S\) gate (the S proxy). Simulations with real \(T\) gates later found that published circuits tolerate fewer faults than their proxies. Which faults are harmful when the \(T\) gates are real, how much distance do the circuits in use actually keep, and how can a check be certified when it is too large to simulate?
- Result
- For any circuit that prepares a magic state with Clifford gates, non-Clifford gates and postselection, one stabilizer code on the \(T\) gates decides harmfulness. A set of faults is harmless if and only if it fires a detector or that code still projects the perturbed \(T\) inputs onto the target or onto zero. This is the condition that governs distillation, so a circuit of fault distance \(d\) needs at least as many \(T\) gates as a distillation protocol of order \(d\) needs inputs (at least \(\max(2d,d+7)\)). For one check on an encoded state, the S proxy can only overstate the fault distance, by any amount. The published colour-code cultivation circuits of distance 3 and 5 have fault distance 2 and 3, against 3 and 4 for their proxies. The circuit run on a superconducting processor has 2 against 3, GHZ-fold checks on codes of distance 3, 5 and 7 have 2, 3 and 4, and an \(\mathbb{RP}^2\) surface-code scheme loses distance in the same way. Checks with a verified cat state keep the full distance, including nearest-neighbour ring and ladder layouts certified up to distance nine.
The paper in detail
A cultivation check measures the logical operator \(TXT^\dagger\), of which the encoded \(T\) state is an eigenstate, by applying a layer of \(T\) and \(T^\dagger\) gates to the data qubits, measuring the logical \(X\) and undoing the layer. A failed check discards the state. The fault distance is the least number of faults that can make the circuit output a wrong state without being caught, and it sets the leading power \(p^d\) of the logical error. Pauli algebra decides it for Clifford circuits through a detector error model, but a Pauli error passing a \(T\) gate becomes a superposition of Pauli errors. The S proxy, standard since cultivation was introduced, keeps the circuit Clifford at the price of changing exactly what happens to \(X\) errors at the \(T\) gates.
The first result makes real \(T\) gates tractable. Each \(T\) gate is replaced by a gadget that consumes an ideal \(T\) state, leaving a noisy Clifford skeleton acting on \(t\) perfect \(T\) inputs. Every fault can be pushed to the end of the skeleton, where it either fires a detector or acts as a Pauli error on the \(T\) inputs. Both maps are linear and computable with a stabilizer simulator, and the skeleton projects the inputs onto one stabilizer code. A fault set is harmless exactly when that code maps the perturbed inputs to the target or to zero, the condition that decides distillation. Because \(X\ket T\) is a superposition of \(\ket T\) and \(Z\ket T\), an error's \(X\) part splits into \(2^{|a|}\) \(Z\) patterns, and adding their amplitudes as Gauss sums gives an exact polynomial-time test for each fault set.
For checks on an encoded state the theory gives a sharp picture. If no fault spreads errors to more data qubits than its own count, layers of \(T\) gates are harmless, as in ordinary fault tolerance. If faults leave \(X\) on part of a minimum-weight logical support just before the last \(T\) layer, the layer turns it into a coherent family of errors, and the decoder keeps the bare \(X\) error with probability \(\prod\cos^2\theta_q\). Later faults then complete a logical operator. \(S\) gates (\(\cos\theta=0\)) are the only diagonal gates that remove this branch, which is why the proxy misses it. A GHZ fold, the tree of CNOTs used to collect the data parity, spreads one fault onto a branch of data qubits, so \((\delta-1)/2\) hooks plus one data fault complete a logical operator on a code of distance \(\delta\): about half the distance.
The computer-assisted certificates confirm the loss on the published circuits. At physical error rate \(10^{-3}\) under the SI1000 superconducting noise model, real \(T\) gates fail about three times as often as the proxy at distance three, 5.8 times at distance five and 6.1 times on the hardware circuit. The fix is Shor-style checks with a verified cat state. On a square grid the cat is grown along a path from its middle, so that a fault spreads \(X\) only towards one end and the parity of the two ends catches it. A ring at distance three (17 qubits) and ladders at distances five, seven and nine (46, 91 and 151 qubits) keep fault distance equal to the code distance. Open problems are a noisy growth stage, cheaper stabilizer rounds on the grid and coherent noise.
Transferable methods
One stabilizer code on the T inputs decides every fault
Non-Clifford gates break the Pauli bookkeeping of fault tolerance. Moving them to the inputs restores it: every \(T\) gate is implemented by consuming an ideal \(T\) state, so the rest of the circuit is Clifford, and every fault can be pushed through Clifford gates to a Pauli error on the \(T\) inputs or a flipped detector. The remaining question, whether a Pauli error on the inputs changes the projected output, is the same question that decides distillation.
A \(T\) gate equals a Clifford circuit acting on an ancilla in \(\ket T\): \[\sqrt2\,\bra0_a\,\mathrm{CX}_{q\to a}\ket\psi_q\ket T_a=T\ket\psi_q.\] With every \(T\) replaced, the postselected skeleton acts on \(\ket T^{\otimes t}\) as a projection \(\Pi_C\) onto a stabilizer code \(C\) followed by decoding. A fault set \(F\) maps linearly to a detector pattern \(D(F)\) and to a Pauli operator \(E(F)\) on the \(t\) inputs. It is harmless if and only if \(D(F)\ne0\) or \(\Pi_CE(F)\ket T^{\otimes t}\in\mathbb C\,\Pi_C\ket T^{\otimes t}\).
\(Z\) errors map \(\ket T\) to an orthogonal state and behave as in distillation, while \(X\) errors do not: \[X\ket T=\tfrac1{\sqrt2}\bigl(\ket T+iZ\ket T\bigr).\] An error whose \(X\) part touches \(|a|\) inputs is therefore a superposition of \(2^{|a|}\) \(Z\) patterns, each of which the code accepts, flips or rejects deterministically in cultivation. The amplitudes are Gauss sums \(\sum_zi^{q(z)}\) of a quadratic form modulo 4, giving the acceptance \(2^{-|a|}(|g_0|^2+|g_1|^2)\) and the harmful weight \(2^{-|a|}|g_1|^2\). These are evaluated variable by variable in about a millisecond per configuration at distance seven.
Because \(Z\) faults on the \(T\) gates never fire a detector and act on \(C\) like noise on distillation inputs, a cultivation circuit of fault distance \(d\) contains a distillation protocol of order \(d\). Its \(T\)-count is therefore at least the least input count of such a protocol: \(\max(2d,d+7)\), and \(3d-4\) for odd \(d\). Fault distance two needs exactly nine \(T\) gates, against the \((3\delta^2+1)/2\) of a single check on a colour code of distance \(\delta\). Editorial check: the identity for \(X\ket T\) and the conjugation rule \(VXV^\dagger=(\cos\theta\,I-i\sin\theta\,Z)X\) for \(V=\mathrm{diag}(1,e^{i\theta})\) were verified numerically.
How to reuse it
Gate teleportation turns any circuit with a few non-Clifford gates into a Clifford circuit on ideal resource states, so detector error models, linear fault propagation and stabilizer simulation apply again. The same reduction should certify fault tolerance of other non-Clifford gadgets, such as \(CCZ\) factories, non-Clifford rotations by injection, or code switching, wherever their harmful faults are decided by a code on the resource inputs.
Coherent completion and why the S proxy overstates distance
Pauli faults that leave \(X\) on part of a logical operator are normally harmless because the code detects the partial error. A diagonal non-Clifford layer changes this: it turns the partial \(X\) error into a superposition that includes both a detectable member and the bare \(X\) error. The decoder projects onto the bare branch, and a few later faults complete it to a logical error. An \(S\) layer has no bare branch, so a design checked with \(S\) gates cannot see this failure mode.
A diagonal gate \(V=\mathrm{diag}(1,e^{i\theta})\) keeps the \(X\) part of an error and adds \(Z\) only where it meets an \(X\) error, \(VXV^\dagger=(\cos\theta\,I-i\sin\theta\,Z)X\). For \(X^u\) before the last layer, the layer produces \[\prod_{q\in u}(\cos\theta_qI-i\sin\theta_qZ_q)\,X^u.\] If \(u\) is a proper subset of a minimum-weight logical support \(\ell\) containing no \(Z\) stabilizer, every member with a \(Z\) part on \(u\) is detected. Only the bare \(X^u\) survives, with amplitude \(\prod\cos\theta_q\). Faults putting \(X\) on \(\ell\setminus u\) after the layer then complete the logical \(X\), and the decoder accepts with probability \(\prod_{q\in u}\cos^2\theta_q\) and outputs the target with logical \(X\) applied.
For \(T\) gates \(\cos^2(\pi/4)=\tfrac12\); for \(S\) gates \(\cos(\pi/2)=0\), so the proxy rejects the configuration with certainty. A GHZ fold spreads one \(X\) fault on its tree into a hook on a branch of data qubits. If \((\delta-1)/2\) hooks each cover two qubits of one minimum-weight support, they cover \(\delta-1\) of its \(\delta\) qubits, and one more fault completes it, so the fault distance is at most \((\delta+1)/2\). For a single check on an ideally encoded state, every late configuration harmful with \(S\) gates is also harmful with \(T\) gates, so the proxy is an upper bound on the real fault distance. Spreading of \(X\) cancels from the proxy's failure condition, so the proxy can overstate the distance by any amount.
Editorial check on the seven-qubit colour code: with the logical support \(\{2,4,5\}\), \(X\) on qubits 2 and 4 before a layer of \(T\) gates and \(X\) on qubit 5 after it gives acceptance 0.25 and output \(\bar X\ket{\bar T}\), whose fidelity with \(\ket{\bar T}\) is 0.5. Replacing the layer by \(S\) gates gives acceptance 0. This is the mechanism of the paper's distance-three example, where one hook and one data fault form one of 48 harmful pairs in the published circuit. Its exact leading coefficient \(32p^2/75\) agrees with the independent exact series of Wan and Zapirain.
How to reuse it
Wherever a Clifford stand-in is used to design or simulate a circuit with diagonal non-Clifford gates (cultivation, magic-state injection, small-angle rotation gadgets, non-Clifford measurements in code switching), the stand-in hides the bare-error branch of every \(X\) error that crosses those gates. Checking that no few faults leave \(X\) on more qubits of a minimum-weight logical support than their number, before the last non-Clifford layer, is a quick design rule for any such gadget.
Context & limitations
Why it matters. Cultivation is among the cheapest known sources of \(T\) states and has been run on hardware, so its real fault distance sets the factory size needed for a target error rate. The result shows that today's GHZ-fold checks keep only about half their code's distance with real \(T\) gates: reaching fault distance \(d\) needs a code of distance about \(2d-1\), roughly four times the qubits, and projections of the hardware run to lower noise along the proxy's \(p^3\) slope are optimistic by a factor growing as \(1/p\). The theory turns certification back into Pauli algebra and supplies design rules and full-distance checks. It extends the authors' exact analysis of distillation, covered here on 23 October (arXiv:2610.31115).
Limitations. Fault distances are for stochastic Pauli noise with an ideal final decoder that discards every state with a detected error; coherent errors and leakage are not covered, and whether a noisy growth stage keeps the checks' distance is open. The certificates at distance nine rest on an integer program whose infeasibility is established by a floating-point solver rather than by exhaustive search (a second solver and exact searches at distance seven agree). The full-distance checks fail less often than the published ones on the same ideal input, but the authors' complete nearest-neighbour protocol at distance three fails more often than the published circuit (\(4.7\times10^{-6}\) against \(8.0\times10^{-7}\)) because of the stabilizer measurements it needs. The paper does not analyse the Steane-code cultivation run on IBM hardware covered here on 19 November (arXiv:2611.22035). The authors state that the work was carried out with the assistance of Claude Opus 5.5. Code and circuits are available on request.