Transversal gates and magic-state distillation

A physical unitary preserves a stabilizer code if conjugation maps its stabilizer group back to itself. Applying Hadamard on every qubit exchanges X-type and Z-type checks. For the seven-qubit CSS code with identical check matrices, this permutes the two check families, so the code is preserved. The all-seven logical X and Z representatives are also exchanged, identifying the resulting operation as a logical Hadamard with consistent logical conventions.

A bitwise controlled-NOT between corresponding qubits of compatible CSS blocks supplies another useful transversal operation. A fault can affect corresponding qubits in different blocks, but an ideal transversal layer does not fan one preexisting single-site error into several sites of the same block. This property requires the physical interactions really follow the transversal partition. Crosstalk, shared controls, or leakage can violate the simple independent-site fault picture and must be included when assessing the actual gadget.

Clifford gates conjugate every Pauli into another Pauli, making their action particularly compatible with stabilizer checks. A phase gate with angle pi over four is outside this group: conjugating X gives a superposition of X and Y rather than one Pauli. This is a concrete algebraic test, not a claim that the gate is intrinsically impossible or always less accurate on hardware. The challenge is implementing the desired logical operation while controlling fault propagation.

For a fixed nontrivial finite-dimensional code detecting arbitrary errors on a single physical subsystem, transversal logical unitary gates cannot form a universal gate set. The assumptions matter: the result does not prohibit universal fault-tolerant computation using measurements, ancillary states, code changes, or other constructions. Gate injection consumes a suitable non-Clifford ancilla through Clifford processing and measurement with feed-forward. Correct conditional corrections and a sufficiently clean resource state are part of that construction, not optional details.

Distillation combines several noisy resource states, checks suitable relations, rejects some outcomes, and conditionally outputs a cleaner state. For a standard fifteen-to-one protocol with independent, identically distributed errors diagonal in the appropriate magic-state basis and ideal Clifford processing, the leading small-error output scales as thirty-five times the cube of the input error. This is an asymptotic model, not an exact polynomial at arbitrary input error, nor a guarantee for coherent correlated resource faults.

Accepted outputs must remain below the protocol's improvement regime, and noisy Clifford gates introduce additional errors that can limit the benefit of another round. Rejection reduces yield and increases average input cost. Two fifteen-to-one levels nominally consume two hundred twenty-five raw inputs per final output before rejection and supporting-circuit overhead. Resource estimates therefore need acceptance rates, factories, storage, routing, and a task error budget. The clean scaling law isolates a mechanism; it does not establish an end-to-end hardware cost.