---
title: Clifford-Deformed Codes
url: https://www.emergentmind.com/topics/clifford-deformed-codes
type: topic
---

# Clifford-Deformed Codes

Clifford-deformed codes are code constructions in which Clifford structure alters stabilizers, logical operators, or decoding while preserving a controlled error-correcting framework. In the most common stabilizer-code usage, a tensor product of single-qubit Clifford unitaries \(U=\bigotimes_{i=1}^n U_i\) maps a Pauli stabilizer group \(\mathsf S\subset \mathsf P_n\) to \(\mathsf S' = U \mathsf S U^\dagger\), so the deformed code remains a Pauli stabilizer code with the same parameters \([[n,k,d]]\) but with locally rotated Pauli axes [2605.15348]. This mechanism has been used to tailor surface, compass, tile, and bivariate-bicycle-derived qLDPC codes to dephasing-biased noise [2201.07802]. In parallel, related literature uses closely allied Clifford language for code conversion and doubling that strengthen transversal gate sets [2408.12752], for code deformation that realizes logical Clifford gates by moving and braiding defects [1006.5260], and for representation-theoretic Clifford codes formulated through projective error models and operator algebra quantum error correction [2606.02531].

## 1. Terminological scope

The phrase “Clifford-deformed code” does not denote a single universally fixed construction. The sources instead organize several distinct, though related, uses of Clifford structure.

| Setting | Mechanism | Representative source |
|---|---|---|
| Pauli stabilizer deformation | Apply \(U=\bigotimes_i U_i\) and conjugate \(\mathsf S\) to \(U\mathsf S U^\dagger\) | [2605.15348] |
| Gate-oriented code conversion | Use doubling, folding, or related transformations to enlarge transversal gate support | [2408.12752] |
| Geometric code deformation | Change measured checks over time so that twist motion or braiding implements logical Clifford gates | [1006.5260] |
| Representation-theoretic Clifford codes | Define codes from induced or projective representations, sometimes beyond stabilizer form | [2606.02531] |
| Circuit-to-code constructions | Derive outcome and spacetime codes from Clifford circuits; the source does not use “Clifford-deformed code” as its own term | [2304.05943] |

A recurrent misconception is to identify all of these with surface-code-style code deformation. One source explicitly states that its “Clifford-deformed” theme is **not** “code deformation” in the surface-code/lattice-surgery sense, but rather code conversion and doubling between classical and quantum code families [2408.12752]. This suggests that the term is best read as a family resemblance across constructions in which Clifford conjugation, Clifford symmetry, or Clifford representation theory is the operative mechanism.

## 2. Local Clifford conjugation as a code transformation

For Pauli stabilizer codes, the basic construction is elementary: starting from \(\mathsf S\subset \mathsf P_n\), apply \(U=\bigotimes_{i=1}^n U_i\) with each \(U_i\) Clifford, and set \(\mathsf S' = U \mathsf S U^\dagger\). Because each \(U_i\) maps Pauli operators to Pauli operators, \(\mathsf S'\) is again a Pauli stabilizer group with unchanged \([[n,k,d]]\), while the local Pauli frame is permuted qubit by qubit [2605.15348]. Under biased dephasing, the point of the transformation is not to alter the topology “in a gross sense,” but to align stabilizers and logicals more favorably against dominant \(Z\)-type noise.

For surface codes, the relevant local Clifford choices under the bias model \(p_X=p_Y\) are \(I\), \(H\), and \(H\sqrt{Z}H \equiv YZ\). One may either deform the stabilizers directly or keep the CSS stabilizers fixed and transform the local noise model in the Heisenberg picture [2201.07802]. The same paper introduces a bias-dependent effective distance \(d'\) and half-distance \(t'\),
\[
d'=\mathcal N^{-1}\log\!\big(p_{\rm log}(1-p)^{-n}\big), \qquad
t'=\mathcal N^{-1}\log\!\big(p_{\rm cor}(1-p)^{-n}\big),
\]
with
\[
\mathcal N=\log\!\left(p\,\eta(1+\eta)^{-1}(1-p)^{-1}\right),
\]
and notes that for unbiased depolarizing noise \((\eta=0.5)\) these reduce to the usual code parameters \(d'=d\) and \(t'=t=\left\lceil \frac{d-1}{2}\right\rceil\) [2201.07802].

For elongated compass codes, Clifford deformation preserves the support of stabilizers, their weights, and the code dimension, but can change the Pauli content and thereby make the code non-CSS. The paper defines two tailored deformations. In the \(XZZX\square\) deformation, Hadamards are applied to the top-right and bottom-left qubits of each weight-4 \(X\)-stabilizer. In the \(ZXXZ\square\) deformation, Hadamards are applied to the top-left and bottom-right qubits of each weight-4 \(X\)-stabilizer [2412.03808]. Both are designed to preserve many weight-2 \(X\)-stabilizers while introducing XZZX-like symmetry.

The Romanesco construction uses a still more structured version of the same principle. Starting from a self-dual bivariate bicycle code built from two reflected cellular-automaton codes, Hadamards are applied to all gray qubits and the CSS parity-check matrix
\[
\left( \begin{array}{c c \mid c c}
H_X^b & H_X^g & 0 & 0 \\
0 & 0 & H_Z^b & H_Z^g
\end{array} \right)
\]
is transformed into
\[
\left( \begin{array}{c c \mid c c}
H_X^b & 0 & 0 & H_Z^g \\
0 & H_X^g & H_Z^b & 0
\end{array} \right).
\]
The result is a non-CSS code in which each stabilizer is obtained by multiplying an all-\(X\) stabilizer from one classical code with an all-\(Z\) stabilizer from the other, or conversely [2506.00130].

## 3. Bias-tailored families and lattice architectures

Clifford-deformed surface codes, or CDSCs, are obtained by applying single-qubit Clifford operators to the ordinary surface code on the square lattice. On the \(3\times 3\) lattice there are \(3^9=19683\) such codes, since each qubit independently receives one of \(I\), \(H\), or \(YZ\), and the paper reports that under strong bias their logical error rates can differ by orders of magnitude even though all have standard distance \(d=3\) [2201.07802]. In the thermodynamic study of random CDSCs, the parameters \((XZ,YZ)\) interpolate between the CSS surface code \((0,0)\), the XZZX code \((0.5,0)\), and the XY code \((0,1)\). The same work identifies a region in the \((XZ,YZ)\) plane with a \(50\%\) threshold at infinite bias and constructs a translation-invariant code in the nearby rational family \(\left(\frac{2}{9},\frac{4}{9}\right)\), which outperforms XZZX and XY in the tested regime [2201.07802].

Elongated compass codes begin as CSS gauge-fixings of the square-lattice quantum compass model, parameterized by an elongation \(\ell\). As \(\ell\) grows, weight-2 \(X\)-stabilizers proliferate while \(Z\)-stabilizers become weight \(2\ell\), producing an asymmetric code whose optimal bias \(\eta_\ell^*\) balances \(Z\)- and \(X\)-error performance [2412.03808]. After Clifford deformation, the paper reports threshold tables in the format CSS \(|\) \(XZZX\square\) \(|\) \(ZXXZ\square\). Representative entries include, at \(\eta=10\), \(\ell=3\): \(14.6 \,|\, 18.0 \,|\, 27.3\); at \(\eta=25\), \(\ell=3\): \(14.1 \,|\, 21.5 \,|\, 33.6\); and at \(\eta=100\), \(\ell=4\): \(16.8 \,|\, 24.9 \,|\, 40.0\). The paper further states that the \(ZXXZ\square\)-deformed compass codes surpass the XZZX surface code threshold at moderate bias, specifically between \(\eta=10\) and \(\eta=25\), and remain better up to \(\eta=100\) [2412.03808].

Clifford-deformed tile codes extend the same bias-tailoring program to planar, geometrically local qLDPC codes with bounded-weight checks, typically weight 6 in the bulk. The open-boundary family has fixed \(k=8\), while the periodic family used for phase-diagram studies has \(k=6\) [2605.15348]. Random deformations independently apply \(H\), \(HSH\), or \(I\) with probabilities \(\Pi_{XZ}\), \(\Pi_{YZ}\), and \(1-\Pi_{XZ}-\Pi_{YZ}\). At infinite bias, the paper finds an extended region with threshold \(50\%\), approximately
\[
0.5 < \Pi_{XZ}+\Pi_{YZ} < 1,\qquad \Pi_{XZ}\le 0.5.
\]
It also gives three translationally invariant examples—linear deformation, XY deformation, and a realization associated with \((\Pi_{XZ},\Pi_{YZ})=(0.25,0.5)\)—for which a \(50\%\) threshold is reported at infinite bias for the sizes studied [2605.15348].

Romanesco codes move the same idea into qLDPC codes derived from fractal or cellular-automaton input codes. They are defined on a bipartite honeycomb or hexagonal lattice, are non-CSS, and have stabilizer generators that are each half \(X\) and half \(Z\). In the main families discussed, bulk stabilizer weight is typically 8, while boundaries can reduce to 6 or 4 [2506.00130]. The paper reports toric examples such as \([[288,12,12]]\) with \(d_c=54\), and planar examples such as \([[244,4,12]]\) with \(d_c=40\), and argues that under strong bias the effective distance approaches the classical distance of the input codes [2506.00130].

## 4. Threshold theory, effective distance, and decoding

In the infinite-bias limit \(p_X=p_Y=0\), \(p_Z>0\), decoding failure is governed by pure-\(Z\) logical operators
\[
\mathcal L_Z = \big(\mathsf P_n^Z\cap \mathcal C(\mathsf S)\big)\setminus \mathsf S_Z.
\]
For a pure-\(Z\) logical \(L\), a \(Z\)-error \(E\) causes ambiguity if
\[
|E\cap L|\ge \frac{|L|}{2}.
\]
A general sufficient condition proved for zero-rate LDPC codes is that if every nontrivial pure-\(Z\) logical operator satisfies \(|L|\ge K n^\alpha\) and the total number of pure-\(Z\) logical operators is subexponential, \(|\mathcal L_Z|\le \exp(o(n^\alpha))\), then for any \(p_Z<\tfrac12\),
\[
P_{\mathrm{fail}(n)\le \exp(-\Omega(n^\alpha)),
\]
so the family has a \(50\%\) infinite-bias threshold [2605.15348]. The same paper reformulates this in terms of a basis of logical operators and overlap-load conditions: few, long, weakly overlapping pure-\(Z\) logicals are the favorable regime.

Surface-code work emphasizes a complementary, bias-weighted notion of distance. The effective distance \(d'\) and half-distance \(t'\) are defined through the most likely logical and uncorrectable events, and explain why two Clifford-deformed codes with the same ordinary distance \(d\) can differ by orders of magnitude in logical error rate under biased noise [2201.07802]. Romanesco codes introduce a related modified weight for a logical with \(s\) \(X\)-type operators and total Hamming weight \(L(s)\),
\[
w = \log(p_X^s p_Z^{L(s)-s})/\log(p_Z)=L(s)+s\delta,\qquad
\delta = -\frac{\log(\eta)}{\log(p_Z)},\qquad
\eta = \frac{p_Z}{p_X},
\]
and define \(d'\) as the minimum of this modified weight over logical operators. In the strong-bias limit, \(d' \to d_c\) [2506.00130].

For compass-code deformations, decoding can be reduced to the undeformed CSS decoding graph with an inhomogeneous noise model. The paper writes
\[
H'^{\intercal} = O_U H^{\intercal}, \qquad H'\vec e'=\vec s = H\vec e, \qquad \vec e = O_U \vec e',
\]
and therefore the effective local error probabilities become qubit dependent:
\[
p_{x,q} =
\begin{dcases}
p_x & U_q = I\\
p_z & U_q = H
\end{dcases}
\qquad
p_{z,q} =
\begin{dcases}
p_z & U_q = I\\
p_x & U_q = H.
\end{dcases}
\]
Minimum-weight perfect matching then uses inhomogeneous edge weights
\[
w=\log\!\left(\frac{1-p}{p}\right)
\]
on the original CSS graph [2412.03808].

Circuit-level decoding adds another layer. For Clifford-deformed elongated compass codes, correlated MWPM improves thresholds for all tested biases relative to standard MWPM under circuit-level noise [2605.27598]. For tile codes, the relevant figure of merit is the residual bias after a full syndrome-extraction cycle. The paper estimates effective single-qubit marginals \((\bar p_X,\bar p_Y,\bar p_Z)\) and defines
\[
p_{\mathrm{eff}=\bar p_X+\bar p_Y+\bar p_Z,\qquad
\eta_{\mathrm{eff}=\frac{\bar p_Z}{\bar p_X+\bar p_Y}.
\]
This links hardware-specific microscopic circuits to phenomenological threshold curves [2605.15348].

## 5. Logical Clifford gates by deformation, folding, and doubling

A distinct branch of the literature uses code deformation to implement logical Clifford operations directly. In planar topological subsystem color codes, code deformation means changing which check operators are measured so that twists are moved, braided, created, or annihilated while logical operators are dragged with the geometry [1006.5260]. The paper introduces colored Majorana operators \(k_i\) attached to twists, with braid rules such as
\[
k_j \rightarrow k_{j+1},
\]
and a color-dependent transformation for \(k_{j+1}\), and derives from this the full Clifford group on encoded qubits. The same work emphasizes that TSCCs retain 2-local gauge measurements in a 2D setting while achieving all Clifford gates by deformation [1006.5260].

Folded surface codes realize a related objective through mirror-dual geometry rather than moving twists. By folding square or diamond-derived surface codes, the paper constructs two families with transversal logical \(H\) and \(S\) gates for arbitrary qudit dimension \(d\) [1603.02286]. The square family uses \(2D^2 - 2D + 1\) data qudits for distance \(D\), while two cone variants for the diamond family use either \(2D^2 - 2D + 1\) or \(2D^2 - 1\) data qudits [1603.02286]. The construction is presented as a logical identity that changes physical layout but preserves code space.

Short-length transversal-gate constructions use a different mechanism again. One paper constructs smallest known doubly even CSS codes, all of which admit a transversal implementation of the Clifford group, then applies a doubling procedure to obtain smallest known weak triply even codes for the same distances and one encoded qubit [2408.12752]. The comparison is explicit: DE codes provide transversal Clifford support; TE* codes provide transversal logical \(T\) with no Clifford corrections; triorthogonal codes implement logical \(T\) via transversal \(T\) plus extra physical Clifford corrections \(S\) and \(CZ\). The paper lists QR-based DE examples \([[7,1,3]]\), \([[23,1,7]]\), \([[47,1,11]]\), up to \([[199,1,31]]\), and QR-based TE* examples from \([[15,1,3]]\) through \([[3239,1,31]]\) [2408.12752].

Hypergraph product codes supply an LDPC analogue of fault-tolerant Clifford gates by deformation. The construction uses generalized punctures and wormholes inside a single code block. Punctures are created by removing connected subgraphs and measuring interior qubits in \(X\) or \(Z\) basis, while wormholes pair smooth and rough punctures through boundary measurements such as
\[
X(c,c')\otimes Z(c',c)
\]
and
\[
X(u,u')\otimes Z(u',u).
\]
These hybrid stabilizers mix \(X\) and \(Z\) at the logical level and enable single-qubit Clifford gates; together with state injection, the paper proposes universal computation within a single HGP block [1909.07424].

## 6. Representation-theoretic, anyonic, and nonstandard generalizations

The term “Clifford code” also has a precise representation-theoretic meaning that is not identical to stabilizer conjugation by local Clifford gates. In the Ising-anyon or Majorana setting, commuting families of even Clifford operators are classified by \(q\)-isotropic subspaces of \(\mathbb F_2^{2n}\), where
\[
q(x,y)=x\cdot y+\mathrm{wt}(x)\,\mathrm{wt}(y)\pmod 2.
\]
If \(q(x,y)=0\), the associated even Clifford operators commute, and \(q\)-isotropic subspaces are in one-to-one correspondence with punctured all-even self-orthogonal codes in one higher dimension [2503.08736]. The resulting code space is denoted \([[n,k]]_{\mathrm{Cl}}\), and detectable even Clifford errors are governed by the \(q\)-orthogonal complement [2503.08736].

Projective representation theory broadens the notion further. A projective error model is a finite group \(G\) with a projectively faithful irreducible projective representation \(\pi\), and a Clifford subspace code is defined by a normal subgroup \(L\trianglelefteq G\) and a projective representation \(\gamma:L\to U(C)\) such that \(Ind_L^G\gamma\simeq \pi\) [2606.02531]. Hybridization is obtained by taking orthogonal translates \(\pi(t)C\) over a coset transversal, and subsystem structure is introduced through a factorization \(C\simeq A\otimes B\), leading to hybrid Clifford subsystem codes \(C(G,\pi,L,G,T_0)\) and an associated operator algebra
\[
\mathcal A(G,\pi,L,G,T_0)=\bigoplus_{t\in T_0}\pi(t)\bigl(\mathcal L(A)\otimes I_B\bigr)\pi(t)^{-1}
\]
[2606.02531].

A closely related paper reformulates stabilizer codes, Clifford codes, and weak stabilizer codes uniformly inside projective error models [2506.01843]. For a Clifford code \(W=V^{Cl}(G,\pi,I,\rho)\), it gives
\[
L_{(G,\pi)}(W)=I,\qquad
S_{(G,\pi)}(W)=\ker(q\circ\rho),\qquad
D_{(G,\pi)}(W)=(G\setminus I)\cup \ker(q\circ\rho).
\]
The same source proves that every stabilizer code is a Clifford code, while also giving infinite families of non-stabilizer Clifford codes and non-Clifford weak stabilizer codes [2506.01843]. A common misconception is therefore that “Clifford code” is synonymous with “stabilizer code”; in this representation-theoretic sense, it is not.

Outside qubit Pauli and Majorana settings, Clifford deformation also organizes homological rotor and oscillator codes. For planar rotors, the \(n\)-rotor Clifford group is
\[
\mathrm{U}(1)^{n(n+1)/2} \rtimes \mathrm{GL}_n(\mathbb Z),
\]
and homological rotor CSS codes are classified up to CSS-preserving rotor Clifford transformations by the Smith normal form of the \(X\)-check matrix \(H_X\) [2311.07679]. Reversing the correspondence maps rotor Clifford operations into oscillator number-phase settings and yields multimode bosonic codes protecting against occupation-number changes and dephasing [2311.07679].

A still broader, non-quantum use appears in distributed space-time coding. There, extended Clifford algebras \(\mathbb A_n^L\), with Clifford generators \(\gamma_k\) and commuting involutions \(\delta_i\), are used to construct full-diversity \(4\)-group ML decodable distributed space-time codes for any power-of-two number of relays. The resulting codes permit lattice decoding on a lattice of four times lesser dimension than in the general case [0704.2505]. This is not a quantum error-correcting construction, but it shows that Clifford deformation language also appears in coding theory more broadly.

Source: https://www.emergentmind.com/topics/clifford-deformed-codes