Quantum computing · Theory and numerical analysis

Persistent-mediator architecture for universal fault-tolerant quantum computation with color codes

Zhi-Cheng He; Jingbo Wang; Dong E. Liu

arXiv:2611.16691v1 ↗PDF ↗Submitted 12 November 2026

Question
Two-dimensional color codes have transversal Clifford gates and three-dimensional tetrahedral color codes have a transversal \(T\) gate, but no code has both. Code switching converts the data register into the 3D code, and teleportation schemes prepare a 3D encoded state and read it out destructively for every \(T\) gate. Can one 3D block stay encoded and protected while it supplies successive logical \(T\) gates to a separate 2D register?
Result
A logical cycle does this deterministically: starting from \(\ket\psi_S\ket+_K\), apply a logical CZ, measure \(\bar Z_S\bar Z_K\), apply transversal \(T\) on the 3D block, measure \(\bar Z_S\bar X_K\), apply a Clifford correction chosen by the two outcomes, and apply CZ again. The output is \(T\ket\psi_S\ket+_K\) in every branch. Physically, CZ is a pairwise layer between the 2D patch and a matched boundary face of the tetrahedral code, \(\bar Z\bar Z\) uses established string-like lattice surgery, and \(\bar X_S\bar X_K\) uses a new membrane-like surgery with ancillas on faces and interior edges. The merged code keeps one logical qubit and distance \(d\). Simulations at \(d=3,5,7\) of a Clifford surrogate, with \(T\) replaced by \(S\), give a phenomenological crossing near 0.7% and a circuit-level crossing near \(1.5\times10^{-4}\), close to 3D memory, and errors that grow linearly over up to 20 reuses.

The paper in detail

Fault tolerance needs a universal gate set, and the Eastin–Knill theorem forbids a single code from implementing one transversally. Color codes split the work cleanly: the 2D triangular code implements every Clifford gate transversally, and the 3D tetrahedral code, the 15-qubit Reed–Muller code at distance 3, implements \(T\). The question is how to connect them. Code switching moves the data into 3D and back; teleportation moves the gate by preparing a 3D resource and consuming it.

The proposal keeps both blocks alive. The logical cycle is gate teleportation in which the destructive readout of the mediator is replaced by a joint measurement. Conjugating by CZ maps \(\bar X_K\) to \(\bar Z_S\bar X_K\), so measuring \(\bar Z_S\bar X_K\) is the mediator's \(X\) readout expressed in the entangled frame. After the Clifford correction \(C_{ab}=(S^aZ^b)\otimes Z^b\) and a second CZ, the mediator is back in \(\ket+_K\), still encoded, and ready for the next call.

Two pieces of geometry make this fault-tolerant. A pairwise CZ between the 2D patch and an isomorphic boundary face of the tetrahedron implements logical CZ, provided faces map to faces and every 3D cell meets the boundary in a valid 2D check. The joint parity \(\bar X_S\bar X_K\) is harder, because the 3D logical \(X\) is a membrane: string-like surgery does not reach it. The membrane surgery adds an ancilla per matched face and per selected interior edge and one new \(X\) check per vertex pair, whose product is \(X(M_S)X(M_K)\).

The authors track the full stabilizer group through the merge. The merged code has one logical qubit with \(\bar X_M=[\bar X_S]=[\bar X_K]\) and \(\bar Z_M=\bar Z_S\bar Z_K\); individual check outcomes reveal only the joint parity, and the spatial distance remains \(d\). Repeating each merged stage for \(d\) rounds protects against measurement errors, and flag circuits extract the weight-up-to-24 checks.

What remains is to simulate the real non-Clifford cycle, establish its circuit-level fault distance, and decide when keeping a 3D block alive beats preparing fresh resources in parallel. The authors propose the Steane–Reed–Muller pair as the smallest experimental test.

Transferable methods

  1. A teleportation cycle that returns its resource

    Teleporting a gate normally ends by measuring the resource. To keep the mediator, the cycle measures only joint parities, which never reveal the mediator's own logical state, and undoes the entangling gate at the end.

    Editorial derivation of the frame change: \(\mathrm{CZ}\,\bar X_K\,\mathrm{CZ}=\bar Z_S\bar X_K\) and \(\mathrm{CZ}\) commutes with \(\bar Z_S\bar Z_K\). Before the cycle the state is \(\ket\psi_S\ket+_K\), stabilised by \(\bar X_K\); after the first CZ it is stabilised by \(\bar Z_S\bar X_K\). The \(\bar Z_S\bar Z_K\) measurement and the transversal \(T_K\) then act as in the standard \(T\)-teleportation gadget, and the \(\bar Z_S\bar X_K\) measurement replaces the destructive \(X\) readout of \(K\).

    Editorial check: simulating the four-step circuit on two qubits for random \(\ket\psi\), each outcome pair \((a,b)\) occurs with probability \(1/4\), and after \(C_{ab}=(S^aZ^b)\otimes Z^b\) and the final CZ the state equals \(T\ket\psi\otimes\ket+\) with fidelity 1 in all four branches.

    How to reuse it

    Any gate teleportation whose resource is a stabilizer state can be made non-consuming: replace the destructive readout by a measurement of the same observable conjugated by the entangling gate, and undo the entangling gate after the Pauli-frame correction.

  2. Membrane-like lattice surgery

    A joint parity is measured in lattice surgery by new local checks whose product is the logical operator. When one logical operator is a membrane, the new checks must also commute with the modified \(Z\) checks across a two-dimensional interface; ancillas on faces and edges supply the needed even overlaps.

    For each matched vertex pair \((v,\pi_0(v))\), \[G_v=X_vX_{\pi_0(v)}\prod_{f\ni v}X_{a_f}\prod_{e\ni\pi_0(v)}X_{b_e}.\] A modified \(Z\) check \(S_f^ZZ_{a_f}\) overlaps \(G_v\) on \(v\) and on \(a_f\) exactly when \(v\in f\), and otherwise not at all, so the overlap is even. Color-code faces have even weight, so each \(a_f\) appears an even number of times in \(\prod_vG_v\), and each edge ancilla twice. Therefore \(\prod_vG_v=X(M_S)X(M_K)=\bar X_S\bar X_K\).

    Any product of a subset \(A\) of the new checks acts on the two data membranes as \(X\) on \(A\) and on its image, so it has the same commutation parity \((-1)^{\lvert A\rvert}\) with \(\bar Z_S\) and with \(\bar Z_K\). It therefore cannot represent \(\bar X_S\) or \(\bar X_K\) alone: the measurement reveals only the joint parity.

    Editorial check on the smallest pair, labelling the 15 Reed–Muller qubits by nonzero 4-bit strings and taking the face where the fourth bit is 0 as the Steane code: the pairwise CZ maps every Reed–Muller \(X\) stabilizer to a Steane \(Z\) stabilizer times itself and every Steane \(X\) stabilizer to a Reed–Muller \(Z\) stabilizer, and maps \(\bar X_S\) to \(\bar X_S\) times a weight-7 logical \(\bar Z_K\). Table 1 is reproduced by \((3d^2+1)/4\), \((d^3+d)/2\), \(3(d-1)^2/2\) and \((4d^4+9d^3+3d)/2\) at \(d=3,5,7\).

    How to reuse it

    To measure a joint logical operator whose support in one code is a surface, add one ancilla per interface face and per interior edge, build one \(X\) check per matched vertex that includes the ancillas around it, and use even face weights to make the ancillas cancel in the product.

Source: Sec. II (Eqs. (1)–(2), Fig. 1c); Sec. III (Eqs. (3)–(5)); Sec. IV (Eq. (6)); Appendix A (Eqs. (8)–(15)); Appendix C.1 (Eqs. (19)–(29)); Table 1. Based on the linked paper version. Derivations marked as editorial, and suggested uses in other problems, are editorial guidance rather than results of the paper.

Context & limitations

Why it matters. It makes the non-Clifford module a persistent, error-corrected part of the processor rather than a consumable, and supplies the missing protected interface: a non-destructive, distance-preserving measurement of a joint parity whose 3D logical operator is a membrane rather than a string.

Limitations. The simulations never include a noisy \(T\) gate: the surrogate replaces it by \(S\) to stay in the stabilizer formalism, so fault propagation through the actual non-Clifford layer and the circuit-level fault distance of the true cycle are not established. Evidence covers \(d\le7\) only. The circuit-level threshold near \(10^{-4}\) is set by weight-24 checks with up to 25 flags. The cycle costs \(O(d^4)\) qubit-rounds, the same scaling as preparation-based schemes, and the 3D block occupies hardware between calls; no break-even analysis is given. The paper has no AI-use statement.

Description based on the full arXiv HTML of v1, read 2026-11-15: Sections I–VII, Appendices A–B and Appendix C.1; the remaining parts of Appendix C (interface-only logicals, distance proofs, flag schedules) were not read. The logical cycle was simulated on two qubits for random inputs and all four outcome pairs; the transversal CZ between the Steane code and the 15-qubit Reed–Muller code was checked on stabilizers and logicals; and the resource formulas of Table 1 were recomputed.