The Signal
He, Wang, Song and Gu Set a CSS-Code Decoupling Criterion
A new paper identifies inter-layer gluing as an obstacle to separating higher-dimensional, translation-invariant topological CSS codes into toric-code stacks. Its authors give a criterion for decoupling and an algorithm to construct the cir
Quantum.world Team · Written with AI · fact-checked ·
Translation symmetry can look settled in a higher-dimensional quantum code while a hidden connection between its excitation layers still prevents it from being separated into simpler pieces. In a paper posted October 29, Mingxin He, Yifei Wang, Hao Song and Yingfei Gu identify that inter-layer “gluing” as the remaining obstruction—and show how coarse-graining can remove it.
The result gives researchers a test for when a translation-invariant topological CSS code can be transformed into a stack of higher-form toric codes using a finite-depth local circuit. The paper’s authors—Mingxin He, Yifei Wang, Hao Song and Yingfei Gu—write:
“Finally, we give an explicit algorithm that constructs the decoupling as a circuit of SUM (qudit CNOT) and SWAP gates.”
Why trivial translations are not enough
CSS codes organize their stabilizer checks into X and Z types. In the topological codes studied here, excitations can be grouped into layers according to their dimension. A decoupling maps the code, through a local CSS-preserving Clifford circuit and possibly with added ancillas, to a stack of toric codes while retaining the translations of a chosen superlattice.
In two dimensions, the paper says, decoupling is possible once translations in that superlattice act trivially on every anyon type. Higher dimensions add another condition: the layers can remain glued together even after translations act trivially on each one. The authors point to a three-dimensional example in which that gluing prevents decoupling that preserves the original translations. Their criterion makes the distinction explicit: both the translation action and the inter-layer gluing must permit a toric-code stack.
The coarse-graining bound depends on the layers
For translation-invariant p-ary CSS codes, where p is prime and p = 2 corresponds to qubits, the paper proves a general bound. Once translations act trivially on every layer, a code with m nonzero layers can have its gluing removed by further coarse-graining by a factor of pm−1 in every direction. Each additional factor of p detaches one more layer from those below it.
That bound guarantees a sufficient refinement, not necessarily the smallest one. In the paper’s three-dimensional example, two layers are glued, and increasing the unit cell by a factor of p along just one direction suffices; the authors show that this is the smallest possible refinement for that example. They also say their general method can use less coarse-graining when it computes the gluing classes directly.
The circuit is constructive, but not yet cheap
The authors turn the criterion into two algorithms. One finds a superlattice where the code’s excitation profile matches that of a toric-code stack, or reports that no such superlattice works; the other constructs the decoupling circuit from SUM and SWAP gates. Both algorithms are implemented in an open-source Python package, which uses exact arithmetic over the finite field.
The complexity result carries an important qualification. For fixed spatial dimension, the circuit-construction step has a polynomial bound in the size of its coarse-grained input, and the same bound applies to circuit depth and ancillas per unit cell. But the paper calls the polynomial far from practical: its proof uses worst-case degree bounds, and the result does not bound the size of the required coarse-graining or the cost of finding it. The theorem supplies a systematic route, not evidence that every eligible code can be transformed efficiently in practice.
A test for logical capacity on a torus
The criterion also applies when the code is placed on a finite torus. The paper shows that the torus reaches the maximum logical-qudit count allowed by the code’s excitation layers exactly when the code can be decoupled on that torus’s period lattice. A nontrivial translation action or surviving gluing lowers the count.
That connection gives the algebraic test a direct finite-system consequence: it identifies when a code’s periodic geometry preserves the full logical capacity associated with its layers. The authors’ rule therefore addresses both how a code can be reorganized into toric-code components and when a finite torus realizes the maximum number of logical qudits that organization permits.